UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Assessing sub-daily exposure to wildfire smoke and its public health effects in British Columbia Yao, Jiayun 2019

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2019_november_yao_jiayun.pdf [ 3.82MB ]
Metadata
JSON: 24-1.0383432.json
JSON-LD: 24-1.0383432-ld.json
RDF/XML (Pretty): 24-1.0383432-rdf.xml
RDF/JSON: 24-1.0383432-rdf.json
Turtle: 24-1.0383432-turtle.txt
N-Triples: 24-1.0383432-rdf-ntriples.txt
Original Record: 24-1.0383432-source.json
Full Text
24-1.0383432-fulltext.txt
Citation
24-1.0383432.ris

Full Text

ASSESSING SUB-DAILY EXPOSURE TO WILDFIRE SMOKE AND ITS PUBLIC HEALTH EFFECTS IN BRITISH COLUMBIA by Jiayun Yao B.Sc., Fudan University, 2010 M.Sc., The University of British Columbia, 2013  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Population and Public Health)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  October 2019 © Jiayun Yao, 2019 ii  The following individuals certify that they have read, and recommend to the Faculty of Graduate and Postdoctoral Studies for acceptance, the dissertation entitled: Assessing Sub-Daily Exposure to Wildfire Smoke and its Public Health Effects in British Columbia  submitted by Jiayun Yao  in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Population and Public Health  Examining Committee: Sarah Henderson, Population and Public Health Co-supervisor Michael Brauer, Population and Public Health Co-supervisor  Kimberlyn McGrail, Population and Public Health Supervisory Committee Member Ian Mckendry, Geography University Examiner John Spinelli, Population and Public Health University Examiner  Additional Supervisory Committee Members: Sean Raffuse, Air Quality Research Center, University of California, Davis Supervisory Committee Member iii  Abstract Global climate change has created new public health issues, and evidence-based policies are needed for mitigating the health impacts. The increasing frequency and intensity of wildfires is one of the pressing concerns in Canada and globally. Epidemiological studies have found that daily average exposure to wildfire smoke is associated with a wide range of cardiopulmonary conditions. However, few studies have looked at the health effects of sub-daily exposures measured in hours, and little is known about the lag-response relationship at such temporal scales. Sub-daily impacts are highly relevant for public health response, especially for smoke episodes of limited duration. To address these knowledge gaps, this dissertation presents a machine learning approach to identify variables relevant to the vertical distribution of smoke in the atmosphere, which can improve the application of remote sensing data for population exposure assessment. Relevant variables included fire activity in the vicinity, geographic location of the smoke, and meteorological conditions. These variables were next combined with data from air quality monitors and ecological information, to develop an empirical model for estimating 1-hour average population exposure to fine particulate matter (PM2.5) during wildfire seasons from 2010 to 2015 in British Columbia, Canada, at a 5 km2 resolution. Compared with observations, model predictions had a correlation of 0.93, root mean squared error of 3.2 μg/m3, mean fractional bias of 15.1%, and mean fractional error of 44.7%. The model estimates were then linked to ambulance dispatches, paramedic assessments, and subsequent hospital admissions. Increased PM2.5 was associated with increased dispatches for respiratory and cardiovascular reasons within one hour following exposure, and for diabetic reasons within the iv  first 24-hour period. Each 10 μg/m3 increase in PM2.5 was associated with an increase in the cumulative odds over 48 hours of up to 10%, 20% and 10% for respiratory, cardiovascular, and diabetes calls, respectively. These results support further investigation into the health effects of sub-daily exposures and suggest that air quality standards and public health actions during wildfire smoke events should be based on the hourly time scale. Public health agencies and the general public should act promptly to reduce exposure when affected by wildfire smoke.   v  Lay Summary Inhaling smoke emitted by wildfires can affect human health. It can worsen conditions in people with chronic diseases, and it causes symptoms in people who are otherwise healthy. This is a growing concern for public health as wildfires will occur more frequently and affect more communities under the changing climate. This research aimed to improve our ability to monitor wildfire smoke exposures every hour at different locations in British Columbia during wildfire seasons, and to study whether more people call the ambulance during hours with more smoke. The findings show that ambulance calls for heart and lung conditions increased within one hour of exposure to smoke, while calls for diabetic conditions increased after 24 hours. These results suggest that the health effects from wildfire smoke can occur soon after exposure, and actions to reduce exposure should be taken promptly.         vi  Preface The work included in this dissertation was designed, carried out, and analyzed by myself, with guidance from co-supervisors Henderson and Brauer, and committee members Raffuse and McGrail. Detailed contributions are described below.  Publications and contributions For Chapter 1, I wrote and edited all of the content, and received feedback from Drs. Henderson, Brauer, McGrail, and Mr. Raffuse. Dr. Henderson also helped in the literature search for the evidence review. My contribution was at least 90%. A version of Chapter 2 has been published. Yao J, Raffuse SM, Brauer M, Williamson GJ, Bowman DM, Johnston FH, Henderson SB. Predicting the Minimum Height of Forest Fire Smoke Within the Atmosphere using Machine Learning and Data from the CALIPSO Satellite. Remote Sensing of Environment. 2018 Mar 1;206:98-106. doi:10.1016/j.rse.2017.12.027. I developed the research method; identified, collected, and processed the data, conducted the model building and evaluation; and wrote and edited the manuscript, representing a total contribution of at least 90% work. Mr. Sean Raffuse’s feedback represents a contribution of 2%; Drs. Brauer and Henderson provided guidance along the process of the research as well as suggesting revisions to drafts, representing a contribution of 3% each. The rest of the coauthors provided feedback on the drafts, representing a contribution of 2% in total.  vii  A version of Chapter 3 has been published. Yao J, Brauer M, Raffuse S, Henderson SB. Machine Learning Approach to Estimate Hourly Exposure to Fine Particulate Matter for Urban, Rural, and Remote Populations During Wildfire Seasons. Environmental Science & Technology. 2018 Oct 24;52(22):13239-49.doi: 10.1021/acs.est.8b01921. I developed the research method; identified, collected, and processed the data, conducted the model building and evaluation; and wrote and edited the manuscript, representing a total contribution of at least 90%. Mr. Sean Raffuse’s feedback represents a contribution of 2%; Drs. Brauer and Henderson provided guidance along the process of the research as well as suggesting revisions to drafts, representing a contribution of 4% each.  A version of Chapter 4 has been submitted for publication. Yao J, Brauer M, Wei J, McGrail KM, Johnston FH, Henderson SB. Association between Sub-Daily Exposure to Fine Particulate Matter and Ambulance Dispatches during Wildfire Seasons. I applied for data access to the various administrative datasets through Population Data BC, as well as direct contact with BC Emergency Health Services. I designed the study, prepared the data, conducted the analyses, interpreted the results, and wrote and edited the manuscript, representing a contribution of at least 90%. Dr. Johnston’s feedback represents a contribution of 0.5%. Ms. Julie Wei provided knowledge about the ambulance dataset, assisted the process of the data request, and reviewed the manuscript, representing a contribution of 1%. Dr. McGrail provided assistance in the process of data request, feedbacks on the study design, and reviewed the manuscript, representing a contribution of 1.5%. Drs. Henderson and Brauer provided guidance throughout viii  the process from the conception of the study to manuscript writing, representing a contribution of 4% each. For Chapter 5, I wrote and edited all of the content, and received feedback from Drs. Henderson, Brauer, McGrail, and Mr. Raffuse. My contribution was at least 95%. Ethics Approvals No ethics approvals were required or sought for Chapter 1, 2, 3 and 5. Initial and ongoing ethics approval for Chapter 4 (H15-02269) was obtained from the University of British Columbia’s Behavioural Research Ethics Board staring on September 14, 2016. ix  Table of Contents Abstract ......................................................................................................................................... iii Lay Summary .................................................................................................................................v Preface ........................................................................................................................................... vi Table of Contents ......................................................................................................................... ix List of Tables ................................................................................................................................xv List of Figures ............................................................................................................................. xvi List of Abbreviations ............................................................................................................... xviii Acknowledgements ......................................................................................................................xx Dedication .................................................................................................................................. xxii Chapter 1: Introduction and background ...................................................................................1 1.1 Background ..................................................................................................................... 1 1.1.1 Introduction ................................................................................................................. 1 1.1.2 Overview ..................................................................................................................... 4 1.1.3 Study area.................................................................................................................... 5 1.1.4 Study period ................................................................................................................ 5 1.1.5 Air quality in the study area ........................................................................................ 7 x  1.2 Rationale and objectives ................................................................................................. 9 1.2.1 Improving the utility of remote sensing data for smoke modeling (Chapter 2) ........ 11 1.2.2 Developing model to estimate 1-hour smoke exposure (Chapter 3) ......................... 11 1.2.3 Assessing the health effects of sub-daily exposure (Chapter 4) ............................... 12 1.3 Review of critical literature on wildfire smoke exposure ............................................. 12 1.3.1 Chemical and physical nature of smoke ................................................................... 12 1.3.2 Exposure assessment of smoke in epidemiology ...................................................... 14 1.3.3 Measuring the vertical distribution of smoke ........................................................... 15 1.3.4 Modeling vertical distribution of smoke ................................................................... 15 1.3.5 Modeling wildfire smoke exposure .......................................................................... 17 1.3.6 Machine learning and random forests ....................................................................... 17 1.4 Review of critical literature on health effects ............................................................... 19 1.4.1 Short-term exposure to wildfire smoke ..................................................................... 19 1.4.2 Sub-daily exposure to wildfire smoke ...................................................................... 21 1.4.3 Sub-daily exposure to PM2.5 ..................................................................................... 21 xi  1.5 Use of ambulance data for epidemiologic studies ........................................................ 23 1.6 Summary ....................................................................................................................... 25 Chapter 2: Predicting the minimum height of wildfire smoke within the atmosphere using machine learning and data from the CALIPSO satellite .........................................................32 2.1 Introduction ................................................................................................................... 32 2.2 Methods......................................................................................................................... 32 2.2.1 The CALIPSO data and the response variable ......................................................... 32 2.2.2 Other data and the potentially predictive variables................................................... 35 2.2.3 Statistical modelling.................................................................................................. 42 2.3 Results ........................................................................................................................... 44 2.4 Discussion ..................................................................................................................... 48 Chapter 3: A machine learning approach to estimate hourly exposure to fine particulate matter for urban, rural, and remote populations during wildfire seasons.............................56 3.1 Introduction ................................................................................................................... 56 3.2 Methods......................................................................................................................... 57 3.2.1 Data sources and variables ........................................................................................ 57 3.2.2 Model building and evaluation ................................................................................. 60 xii  3.2.2.1 Data reduction ................................................................................................... 60 3.2.2.2 Model training and out-of-bag evaluation ........................................................ 62 3.2.2.3 Leave-region-out cross-validation .................................................................... 64 3.2.2.4 Case studies ....................................................................................................... 64 3.2.3 Sensitivity analysis.................................................................................................... 66 3.2.3.1 Inclusion of PM lag 24-hour ............................................................................. 66 3.2.3.2 Inclusion of GOES AOD .................................................................................. 66 3.3 Results ........................................................................................................................... 67 3.3.1 Case study 1 .............................................................................................................. 71 3.3.2 Case study 2 .............................................................................................................. 72 3.3.3 Sensitivity analyses ................................................................................................... 72 3.4 Discussion ..................................................................................................................... 76 Chapter 4: Association between sub-daily exposure to fine particulate matter and ambulance dispatches during wildfire seasons..........................................................................81 4.1 Introduction ................................................................................................................... 81 4.2 Methods......................................................................................................................... 81 xiii  4.2.1 Health outcome data ................................................................................................. 81 4.2.2 Exposure assessment and assignment ....................................................................... 84 4.2.3 Statistical analysis ..................................................................................................... 86 4.3 Results ........................................................................................................................... 88 4.4 Discussion ..................................................................................................................... 97 Chapter 5: Conclusion ...............................................................................................................102 5.1 Research summary and contributions ......................................................................... 102 5.1.1 Additions to the health literature............................................................................. 103 5.1.2 Advancement in exposure assessment .................................................................... 103 5.1.3 Application of machine learning for public health ................................................. 105 5.1.4 Using symptomatic outcomes for epidemiologic studies ....................................... 106 5.1.5 Illustrating the lag-response relationship with distributed lag models ................... 107 5.2 Implications for public health policy and future work................................................ 108 5.2.1 Methodological limitations ..................................................................................... 108 5.2.2 Implications for public health policy ...................................................................... 110 5.2.3 Recommendations for future work ......................................................................... 113 xiv  Bibliography ...............................................................................................................................117 Appendices ..................................................................................................................................138 Appendix A Procedure for calculating summed fire radiative power (FRP) in each 5km grid cell, adjusting for the scan angle and bow-tie effects. ............................................................ 138 Appendix B Codes for ambulance dispatches and paramedic assessments. ........................... 139 B.1 Medical Priority Dispatch System (MPDS) codes.................................................. 139 B.2 Paramedic impression codes ................................................................................... 140  xv  List of Tables Table 1.1 Summary of case-crossover studies on myocardial infarction (MI) and out-of-hospital cardiac arrest (OHCA) .................................................................................................................. 26 Table 2.1 Description and summary statistics of variables for minimum height model .............. 41 Table 2.2 Descriptive analysis of the low altitude observations in the cluster of over-prediction 46 Table 2.3 Model performance for predictions of Minimum Height below thresholds.................. 47 Table 3.1 Data sources and variables for OSSEM-1h. ................................................................. 61 Table 3.2 Statistics used for evaluation of the OSSEM-1h model. .............................................. 63 Table 4.1 Definitions and number of cases for each health outcome measure. ............................ 85 Table 4.2 Matched Hospital Diagnosis for Ambulance Dispatch codes. ..................................... 87  xvi  List of Figures Figure 1.1 Annual area burned in British Columbia. ...................................................................... 6 Figure 1.2 Maps of wildfires in British Columbia. ......................................................................... 6 Figure 1.3 Fire activity in British Columbia by month. .................................................................. 7 Figure 1.4 Smoke impact in southwestern British Columbia in 2015. ........................................... 8 Figure 1.5 Time-series of fine particulate matter (PM2.5) concentrations during smoke event. ..... 9 Figure 2.1 Illustration of the CALIPSO data. ............................................................................... 35 Figure 2.2 Scatterplot comparing observations and predictions for minimum height model. ...... 47 Figure 2.3 Dotchart of variable importance of minimum height model. ...................................... 48 Figure 2.4 Partial dependence plots for minimum height model. ................................................. 52 Figure 2.5  Illustration of potential application of the minimum height model. ........................... 55 Figure 3.1 Areal extent of Biogeoclimatic Zones in British Columbia. ....................................... 58 Figure 3.2 Flowchart of data reduction for OSSEM-1h model training. ...................................... 62 Figure 3.3 Map of British Columbia, Canada. .............................................................................. 65 Figure 3.4 Variable importance plot for the OSSEM-1h model. .................................................. 68 xvii  Figure 3.5 OSSEM-1h model performance stratified by fire activity. ......................................... 69 Figure 3.6 Spatial distribution of evaluation statistics for OSSEM-1h model. ............................ 70 Figure 3.7 Illustration of OSSEM-1h performance in Case study 1. ............................................ 74 Figure 3.8 Illustration of OSSEM-1h performance in Case study 2. ............................................ 75 Figure 4.1 Flowchart of analytic data selection. ........................................................................... 83 Figure 4.2 Corresponding relationship between health outcome measures. ................................. 90 Figure 4.3 Results for PM2.5 exposure and all ambulance dispatch calls during wildfire season . 93 Figure 4.4 Results for PM2.5 exposure and respiratory health outcomes during wildfire season . 93 Figure 4.5 Results for PM2.5 exposure and circulatory health outcomes during wildfire season . 94 Figure 4.6 Results for PM2.5 exposure and diabetic outcomes during wildfire season ................. 95 Figure 4.7 Results of lag-response relationship in sensitivity analysis with cases of different severity. ......................................................................................................................................... 96 Figure 4.8 Results for subtypes of diabetic conditions. ................................................................ 99 Figure 5.1 Examples of non-linear exposure-response relationship between PM2.5 and ambulance dispatches. ................................................................................................................................... 113  xviii  List of Abbreviations AIC Akaike Information Criterion AOD Aerosol Optical Depth BC     British Columbia BCEHS British Columbia Emergency Health Service CALIO Cloud-Aerosol Lidar with Orthogonal Polarization CALIPSO Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation CO Carbon Monoxide COPD Chronic Obstructive Pulmonary Disease DLNM Distributed Lag Non-linear Model FIRMS Fire Information for Resource Management System FRP Fire Radiative Power FTP File Transfer Protocol GOES Geostationary Operational Environmental Satellite IDW Inverse Distance Weighted IAED International Academies of Emergency Dispatch ICD-10 International Classification of Diseases, 10th Revision IHD Ischemic Heart Disease IQR Interquartile Range MERRA Modern Era Retrospective-analysis for Research and Applications MFB Mean Fractional Bias xix  MFE Mean Fractional Error MI Myocardial Infarction MISR Multi-angle Imaging Spectro-Radiometer  MODIS Moderate Resolution Imaging Spectrometer MPDS Medical Priority Dispatch System NASA National Aeronautics and Space Administration NMOC Non-Methane Organic Compounds NOx Nitrogen Oxides O3 Ozone OHCA Out-of-Hospital Cardiac Arrest OSSEM Optimized Statistical Smoke Exposure Model PAHs Polycyclic Aromatic Hydrocarbons PBLH Planetary Boundary Layer Height PHN Personal Health Number PI Paramedic Impression PM Particulate Matter PM2.5 Particulate Matter with an aerodynamic diameter smaller than 2.5 micrometers PM10 Particulate Matter with an aerodynamic diameter smaller than 10 micrometers r Pearson’s Correlation Coefficient RMSE Root Mean Square Error  xx  Acknowledgements I owe special thanks to my co-supervisors Drs. Sarah Henderson and Michael Brauer, who have provided invaluable support and guidance throughout my PhD. They are not only advisors for my graduate studies, but more importantly the mentors for my career and life, showing me how to do research with innovation and integrity, take responsibility as an academia and practitioner in environmental health, and treat people with respect and genuineness. I am truly grateful to have them as my supervisors. I thank my committee member Dr. Kimberlyn McGrail for sharing her expertise in administrative health data and providing support in the process of my data request. I also thank committee member Mr. Sean Raffuse for providing his insights in smoke modeling and remote sensing data.  I am thankful to have labmates, classmates and friends who have provided technical and moral support throughout my PhD, including Raphael Arku, Anders Erickson, Hind Sbihi, Matt Shupler, Matt Wagstaff, Emily Rugel, Jessica Yu, and Weiran Yuchi, Annie Wang, Jennifer Guthrie, Nancy Laliberte, Amy Hall, Ther Aung, and Bojosi Gamontle. The Australian Research Council Linkage Projects and the BC Lung Association funded my PhD, and I thank them to make this work possible. I am also very fortunate to have the opportunity to collaborate with colleagues in Australia and borrow their expertise in wildfire smoke studies.  xxi  Special thanks are owed to my parents, who have supported me throughout my years of education, both morally and financially. The support from them and my mother-in-law was the only reason I could finish this dissertation while caring for a newborn baby. I also want to thank Seagle Yang Liu, a supportive and caring husband who has always been on my side, not to mention the free statistical consulting I have got over the years. Thank you for being in my life, and my journey has become so much more fun with you around! Finally, to baby Johanna – thank you for arriving at the final stage of mommy’s PhD journey. You will always be my inspiration to become a better version of myself! xxii  Dedication  To Johanna and Seagle. 1  Chapter 1: Introduction and background 1.1 Background 1.1.1 Introduction Globally, around 400 million hectares of the land surface is burned by fire each year (1). Although a recent study reported a 25% decline in global area burned over the last two decades driven by agricultural expansion and intensification in savannas and grasslands, the opposite was observed in regions with the most tree cover, such as parts of North America, Australia and Russia (2). Fire season length in North America increased by an average of 14 days during the period between 1979 and 2015 (3). Large wildfires (>405 ha) in the western US increased at a rate of seven fires per year, and the total fire area increased at a rate of 355 km2 per year between 1984 and 2011 (4). These trends can be partially attributed to the increasing temperature and drought severity caused by global climate changes that are expected to continue into the future (5-8). Wildfires can be caused by human activity or natural processes, such as lightening. They can also be suppressed or managed by humans to some degree, depending largely on the weather conditions, time between detection and management efforts, and location of the fire. The word wildfire can also be synonymous with other terms, such as vegetation fire, landscape fire, forest fire, bush fire, wildland fire, etc., depending on the context. To be consistent, I use the word wildfire to describe any fire on the landscape throughout the thesis, without intention to indicate the main cause(s).  2  Wildfires pose a direct threat to lives and property in the areas they affect. Smoke emissions from wildfires pose a different threat, which can affect much larger populations by degrading air quality at local, regional and global scales (9-12). They are an important source of ambient fine particulate matter (PM2.5) pollution in many regions (13-15), estimated to be responsible for 5-8% of the 3.3 million deaths attributable to outdoor ambient PM2.5 each year (16, 17). Exposure to wildfire smoke has also been linked to a wide range of cardiopulmonary health effects from increased symptoms to increased risk of mortality (18-21), and the burden for health care systems is increasing as well (22-25). Most studies on the health effects of wildfire smoke have examined daily average (24-hour) exposures, with few studies having examined the effects of sub-daily (1-hour) exposures. As a result, it remains unclear whether the cardiopulmonary events associated with wildfire smoke are due to short (i.e. hours) but extreme levels of exposure, or to cumulative exposures averaged over longer (i.e. days) periods. Because the levels of wildfire smoke in a community can change rapidly within a day, such evidence is needed to develop appropriate air quality standards and public health strategies for timely mitigation of the health risks.  To generate evidence on the short-term health impacts of wildfire smoke, the first challenge is to assess exposure with high spatial and temporal resolution. Most of the available studies on sub-daily exposure have used hourly measurements from air quality monitoring stations in the affected region (26). However, monitoring stations tend to be located in areas with high population density, excluding the rural and remote areas that are more frequently and severely impacted by wildfire smoke. In addition, assigning exposure from a central monitor assumes no 3  spatial variability over a large geographic area, where studies show that smoke levels can be heterogeneous over very small areas (27). Under these conditions, using a single monitor introduces non-differential exposure misclassification and likely biases the effect estimates towards the null. Another issue with monitoring stations is that the measurement instrumentation may fail when impacted by the heavy smoke or actual fire (28).  Remotely sensed fire and smoke products are promising alternatives to ground monitoring stations given their broad spatial coverage at relatively high spatial resolution. Such products have been used to directly or indirectly assess smoke exposure in many recent studies (29-33). However, most remote sensing products provide information integrated from the top to the bottom of the atmosphere, rather than information specifically at the surface where populations are exposed. Better information about the vertical distribution of smoke in the atmosphere has been identified as key to improving the utility of remote sensing products in air pollution exposure modeling and public health research (34, 35). However, routine measurements of the vertical distribution of air pollutants within the column are not yet feasible by surface monitoring or remote sensing.  The second challenge is to measure population health outcomes at the sub-daily scale. Because wildfire smoke events are usually sporadic, large populations and long time-series of data are needed to provide enough statistical power to detect their effects. This is most feasible using routinely collected administrative health data, such as hospital admissions, emergency room visits, medical billings, and pharmaceutical dispensations. However, most such databases are limited in their utility for studying sub-daily wildfire smoke exposures. First, they generally 4  provide the date of the health outcome, but not necessarily the exact time. Second, there may be gaps between the time of the initiation of the health outcome and the recorded time of the delivery of health care services, when available. For example, a physician visit appointment may be scheduled days after a patient first experiences symptoms. Third, some measures, such as medication dispensations, may not reflect actual disease exacerbations, but rather the preparation of patients for conditions perceived as potentially harmful. Finally, many databases record residential address, which may be different than the geographic location where the health outcome occurs. With this thesis, the aim was to add to the epidemiologic evidence on sub-daily wildfire smoke exposure by (1) developing a spatially and temporally resolved statistical model to estimate sub-daily exposure to PM2.5, and (2) assigning the modeled exposure to subjects in an administrative health database that records the exact time and location of health events, and then assessing the exposure-response relationship. 1.1.2 Overview This manuscript-based thesis has five chapters and two appendixes. Chapter 1 provides the background and critical review of relevant literature necessary for readers to understand the rationale of the thesis. Chapter 2 describes a machine learning approach to improve the use of remote sensing data in modeling human exposure to wildfire smoke. Chapter 3 uses data identified in Chapter 2 in combination with other relevant data to develop a model to estimate hourly population exposure to wildfire smoke across the province of British Columbia (BC). These exposure estimates are used in Chapter 4 to assess their associations with ambulance 5  dispatches, paramedic assessments, and hospital admissions. Chapter 5 concludes the thesis by discussing its contributions and implications.  1.1.3 Study area BC is the westernmost province of Canada, with a total land area of 925,186 km2, and 2017 population of 4.8 million people (36). With almost two-thirds of the land surface covered by forests, the province is prone to wildfires (37). As the global climate changes, the province has experienced rising temperatures and more drought in the summer months (38). At the same time, nearly 30% of trees were killed from 1990-2010 by the largest mountain pine beetle infestation in the recorded history of the province (39, 40). Fire-favouring weather, high fuel loads, and the compromised health of the forests have led to more extreme wildfire seasons in the past decade (Figure 1.1). The interior region of the province has been the most impacted by large wildfires ( Figure 1.2). 1.1.4 Study period The study period includes wildfire seasons from 2010 through 2015, during which data for both exposure and health outcomes were available with relatively high quality. This study period covers the 2010, 2014, and 2015 wildfire seasons, which were considered extreme compared with previous decades (Figure 1.1). The 2017 and 2018 seasons were even more extreme, but both occurred after the exposure assessment work and administrative health data application for this thesis was largely complete. Smoke events impacted a large proportion of the population in 6  the province in each of these years. Typical fire seasons in BC span from April to September, with the most intense activity in July (Figure 1.3).   Figure 1.1 Annual area burned in British Columbia.  Created with data from National Forestry Database and BC Wildfire Service.    Figure 1.2 Maps of wildfires in British Columbia. These maps show wildfires colour-coded by fire size in the first year of the study period (2010, left panel) and the last year of the study period (2015, right panel). Both were extreme wildfire seasons, compared with the previous decades. Created by BC Wildfire Service (https://www2.gov.bc.ca/gov/content/safety/wildfire-status/wildfire-statistics). 7   Figure 1.3 Fire activity in British Columbia by month.  Monthly area burned (red bar) and total number of fires reported (black line), averaged from 2010 to 2015.    1.1.5 Air quality in the study area Ambient air quality in BC is generally among the best in the world, even in large urban areas. One of the most important indicators of ambient air quality is the mass concentration of PM2.5. The annual average concentrations of PM2.5 ranged from 1.6 µg/m3 to 13 µg/m3 across BC in 2016, with most monitoring stations below the provincial objective of 8 µg/m3 and the World Health Organization guideline of 10 µg/m3 (41). However, episodes of extremely high PM2.5 concentrations can occur, especially during wildfire seasons. For example, multiple fires were burning north of the greater Vancouver area in early July of 2015. Under the meteorological conditions at the time, smoke was dispersed to the densely populated area of greater Vancouver and southern Vancouver Island (Figure 1.4). Daily average PM2.5 concentrations reached almost 8  100 µg/m3, which was almost unprecedented for these regions. The hourly PM2.5 concentrations were elevated to almost 200 µg/m3, with high temporal variability within each day (Figure 1.5).  Figure 1.4 Smoke impact in southwestern British Columbia in 2015.  Smoke dispersion over greater Vancouver and southern Vancouver Island from fires burning at multiple locations (red dots). Satellite image captured by Moderate Resolution Imaging Spectrometer (MODIS) aboard the Aqua Satellite on July 5th, 2015. Stars indicate the monitoring stations at Squamish (yellow) and Burnaby (blue).  9   Figure 1.5 Time-series of fine particulate matter (PM2.5) concentrations during smoke event. These figures show time-series of PM2.5 concentrations during a smoke event in 2015 at Burnaby (top) and Squamish (bottom). The black line indicates the 1-hour average observations and grey bars indicate the corresponding 24-hour average in the same calendar day. The red boxes show two days with the same 24-hour average, but different variability in the 1-hour averages.   1.2 Rationale and objectives Wildfire smoke exposure is unpredictable and episodic. It can last anywhere from a few hours to a few weeks in the Canadian context. So far, there is limited evidence about whether a few hours of smoke exposure can pose a significant public health risk, or whether the risk is higher during 10  peak hours of a more prolonged exposure. For example, the 24-hour average PM2.5 was just over 50 µg/m3 at Burnaby (blue star in Figure 1.4) on July 5th and at Squamish (yellow star in Figure 1.4) on July 9th, but these same 24-hour averages came from very different 1-hour average concentrations within the day (Figure 1.5, indicated by red boxes). The 1-hour average PM2.5 was below 25 µg/m3 for 15 out of the 24 hours on July 5th at Burnaby, while the elevated 24-hour average was mostly attributable to the few hours of concentrations over 100 µg/m3. On the other hand, the 1-hour PM2.5 was over 25 µg/m3 for 22 hours on July 9th at Squamish, with no concentrations over 100 µg/m3. In studies using 24-hour PM2.5 averages, these two situations would have been classified as the same exposure. However, it is largely unknown whether the different distributions of 1-hour exposures within a day make any difference to the observed population health effects. This is an important question to address, because it has implications for how public health authorities should respond to smoke events. The premise for my dissertation is that I can take advantage of recent advancements in machine learning to improve the utility of remote sensing data for modelling population exposure to wildfire smoke at the sub-daily scale. Using the modeled exposures and a unique administrative health database with resolved spatial and temporal information, we can improve understanding of the health effects of sub-daily exposure to wildfire smoke. This advancement in the epidemiologic knowledge will be valuable for improving emergency responses and reducing public health impacts from wildfire smoke in BC, as well as other regions with similar smoke impacts nationally and internationally.  11  1.2.1 Improving the utility of remote sensing data for smoke modeling (Chapter 2) The objective of Chapter 2 was to demonstrate a machine learning approach to estimate the vertical distribution of smoke in the atmosphere at 1-hour intervals for all of BC at a 5 km2 resolution. Some satellites, such as the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO), can produce snapshots of the vertical profile of the atmosphere. However, these satellite products have very limited spatial and temporal coverage, meaning that they intersect with wildfires and smoke relatively rarely. Instead of using these data directly, I developed a model to predict the vertical distribution of smoke observed by CALIPSO using other spatially and temporally varying data about fire activity in the vicinity, geographic location, and meteorological conditions.  1.2.2 Developing model to estimate 1-hour smoke exposure (Chapter 3) The objective of Chapter 3 was to develop a statistical model to generate 1-hour smoke exposure estimates for the population of BC at a 5 km2 resolution. In previous work, a statistical model was developed to estimate the 24-hour averaged exposure for the same population using remotely sensed fire and smoke, meteorological parameters, and ground-monitoring data (42). Here, the 24-hour model was expanded to generate 1-hour estimates by (1) including more temporally resolved data, (2) including variables predictive of the vertical distribution of smoke identified in Chapter 2, and (3) applying more advanced modeling techniques. The model was evaluated quantitively by comparing estimates with the observations at monitoring stations, and qualitatively by visualizations in case studies.   12  1.2.3 Assessing the health effects of sub-daily exposure (Chapter 4) The objective of Chapter 4 was to assess the acute health effects of sub-daily exposure to wildfire smoke. Hourly estimates of PM2.5 during wildfire seasons from 2010 to 2015 for all populated areas in BC were generated from the model developed in Chapter 3 and linked to ambulance dispatch data, which has information on the exact time and location where the health emergency occurred. A case-crossover study design was applied to assess whether the odds of dispatches for respiratory, cardiovascular, or diabetic complaints increased within hours of smoke exposures. The reasons for calls were assessed by three measures including the ambulance dispatch code, the subsequent paramedic assessment code (when available), and subsequent hospital admission diagnosis (when applicable), to examine the consistency between these somewhat different outcomes.   1.3 Review of critical literature on wildfire smoke exposure  1.3.1 Chemical and physical nature of smoke Wildfire smoke is a complex mixture of gases and particles resulting from the incomplete combustion of biomass. Some components known to be pertinent to human health include particulate matter (PM), carbon monoxide (CO), oxides of nitrogen oxides (NOx), and non-methane organic compounds (NMOC) such as polycyclic aromatic hydrocarbons (PAHs) and other volatile organic compounds (VOCs) (43). Many of these pollutants can interact with ultraviolet radiation to increase concentrations of ground-level ozone (O3), which has also been observed during smoke episodes (44, 45). The composition of the smoke mixture can be 13  influenced by various factors including the fire and fuel characteristics, weather, and the different stages and temperatures of combustion (43, 46-48). These factors can also influence the chemical composition and size distribution of PM emitted from wildfires. For example, PM emitted during the more efficient combustion of the flaming phase tends to be smaller in size (with a peak diameter < 1µm) and to have higher proportions of elemental carbon, PAHs, and heavy metals when compared with PM emitted during the less efficient combustion of the smoldering phase (49, 50). These differences in chemical composition and size may contribute to differences in the toxicity of the PM (50, 51). Although wildfire smoke PM can vary in size, the majority falls into the PM2.5 range, and the mass concentration is the most commonly used metric to represent the complex wildfire smoke mixture in epidemiologic studies. There are a few different reasons for this: (1) PM2.5 is widely measured for the purposes of ambient air quality assessment and regulation, and it is the most consistently elevated such pollutant during wildfire smoke events (18); (2) particles in PM2.5 range can penetrate deep into the gas-exchange region of the lung (52) and are more likely to be deposited in the lung and cause subsequent cytotoxic effects compared with larger particles (53); and (3) PM2.5 remains suspended in the atmosphere for long periods, and can be transported to regions distant from the fire, whereas other pollutants in the mixture volatilize and disperse more rapidly. For example, elevated PM2.5 concentrations have been observed in BC due to wildfire smoke originating within the province (31, 54), from the United States (55), and even across the Pacific Ocean from Siberia (56, 57). Even so, some of the health effects associated with wildfire smoke are likely attributable to the complex mixture, for which PM2.5 is currently the best proxy.  14  1.3.2 Exposure assessment of smoke in epidemiology Data from regulatory air quality monitoring networks is most widely used for assessing population exposure to wildfire smoke in epidemiologic studies. PM2.5 concentrations measured with gravimetric or optical instruments have typically been directly assigned to nearby populations as the exposure (31, 58-64), or used to categorize smoke-impacted periods and compare health outcomes within and outside of those periods (65-67). With routine maintenance and calibration, these monitors generally provide accurate and reliable estimates of wildfire smoke exposure at the surface of the earth, but they are spatially sparse, especially in rural or remote areas, and they can be overloaded during extreme smoke events (68). Satellite remote sensing can measure air pollutants in the atmosphere as well as fires on the surface of the earth. For example, the visible channels of satellite sensors can measure the presence of particles in the atmosphere, called aerosol optical depth (AOD), based on the fact that particles can change the way the atmosphere reflects and absorbs light waves. On the other hand, the thermal infrared channels can measure the surface temperature of the earth, which is used to identify fire hotspots when thermal abnormalities are detected. Remote sensing products of smoke and fire have been increasingly used in recent epidemiologic studies, either alone (28, 31, 69, 70) or in combination with monitoring data (22, 54), to identify smoke impacted time periods or areas. They are also important inputs for deterministic and empirical modelling of wildfire smoke (42, 71), as discussed in the Section 1.3.5. Remote sensing data can cover large areas with different spatial and temporal resolutions, depending on the satellite orbit and capabilities of the instruments. However, these products can be influenced or invalidated by 15  cloud cover, leading to further uncertainty in the measurements. In addition, they measure air pollutants in the total column of the atmosphere rather than at the surface, where most people are exposed. 1.3.3 Measuring the vertical distribution of smoke There are two remote sensing platforms commonly used to provide information about the vertical distribution of atmospheric aerosols. The Multi-angle Imaging Spectro-Radiometer (MISR) aboard the Terra satellite (72) can measure the altitude of the layer of maximum contrast, which can be the land surface, cloud top, or smoke plume top. However, MISR does not always provide the complete vertical profile of the smoke plume, especially for thin smoke or smoke far from the source (73). On the other hand, products derived from the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite can provide both aerosol feature classifications (e.g. smoke, dust, etc.) and a complete picture of the vertical profile. However, data from CALIPSO are limited by its extremely narrow swath of observations. Given that CALIPSO passes over most locations only once every 16 days, it is unable to measure the vertical profile of most smoke plumes. To achieve more spatially and temporally resolved information on the vertical distribution of smoke we need to turn to models rather than measurements. 1.3.4 Modeling vertical distribution of smoke There are two major approaches of modeling: deterministic and empirical. Deterministic models use fundamental mathematical descriptions of the physical and chemical process of fire behavior, 16  smoke emissions, and atmospheric dispersion processes (74) to estimate the output of interest. Deterministic models are the most commonly used in the field of air pollution modeling and forecasting. However, these models tend to require intensive resources and expertise to develop and maintain. For the vertical distribution of wildfire smoke, the initial maximum smoke plume rise over the fire (injection height) and the subsequent vertical distribution of smoke (downwind height) are modelled separately, with the former as an important input for the latter. Such models are currently used for operational smoke forecasting systems, including the BlueSky (75, 76) and FireWork (77) frameworks in Canada. So far, most studies have focused on using CALIPSO and MISR data to evaluate and improve injection height estimates (78-81). However, the more important estimate for public health is the downwind height, which describes smoke dispersion to the surface where human populations can be exposed.  Empirical models estimate the outcome of interest as a function of the relevant measurements that are currently available using the statistical relationship between those measurements and the outcome derived from historical data (74). Empirical approaches typically rely on large volumes of historical data for model training, but they require much less time and expertise to develop and maintain than deterministic models. So far, there are no such models for injection height or downwind height, although some studies have recognized their feasibility and have identified potentially predictive variables related to fire intensity and meteorological conditions (72, 82-86). 17  1.3.5 Modeling wildfire smoke exposure  Both deterministic and empirical models can be used to model air pollution exposure. Deterministic models have been used in epidemiologic studies to make spatially and temporally resolved estimates of PM2.5 exposure during wildfire smoke events (28-31, 87-92). Information about fire location and size, fuel type, and meteorological conditions is entered into the models, which simulate the entire life-span of the particles: how many are emitted, how high and far they reach and, ultimately, the ground-level concentrations in the geographic areas of interest (76, 77, 93). The accuracy of the estimates depends on the availability and quality of the relevant input data and model parameterization. Deterministic models have been used to estimate and forecast smoke impacts in several regions (75-77, 94), but the performance is variable (28, 77, 95).  Empirical models to estimate daily PM2.5 exposure from wildfires have been developed in North America and Australia (42, 71, 96, 97). These models estimate ground-level PM2.5 concentrations using data sources including remotely sensed fires and smoke, meteorological measurements, and output from deterministic models. The Optimized Statistical Smoke Exposure Model (OSSEM-24h) developed to estimate 24-hour average PM2.5 exposure during wildfire seasons for all populated areas in BC has been successfully used for public health surveillance and epidemiologic research (29, 98).  1.3.6 Machine learning and random forests Conventional approaches to empirical air pollution modeling have relied on linear regression, but rapid developments in the fields of data mining, machine learning, and computation have made 18  these alternative methods more widely available in recent years (96, 99). Machine learning is a range of computational methods that produce predictions based on algorithms trained using historical data, in an automated and iterative process without explicit programming (100-102). Many machine learning approaches can accommodate complex relationships that are non-linear, or that are best explained by high-dimensional interactions. These properties have made them attractive for a wide range of applications, and several recent studies have used them to model air quality and pollutant exposures (103). Unlike deterministic models, empirical models derived from machine learning predict exposure based purely on the relationships between the response and predictive variables, without explicit knowledge of the complex and inter-dependent underlying physical processes. Although prior knowledge of the basic physical processes is necessary to offer these models a relevant set of predictive variables, achieving the most predictive model is the primary objective of machine learning approaches, not the interpretability of the output. These models tend to produce estimates that agree better with observations than their deterministic counterparts (104-107).  The random forests model is one commonly used machine learning approach. It is an ensemble of regression trees, each of which was constructed with a random subset of observations and a random subset of predictive variables. For each tree, predictions are made with the subset of data that were not sampled for training, otherwise known as the “out-of-bag” data. Because each data point falls into the out-of-bag subset multiple times, it has multiple predicted values for the response variable. The average of these values is reported as the final out-of-bag prediction for that data point, and the final out-of-bag predictions for all data points can be used to evaluate the model performance. A unique feature of this modeling approach is that the importance of each 19  variable can be calculated using a permutation test on the out-of-bag predictions (108). If a predictive variable is important, a random permutation of the variable will considerably increase the model prediction errors when compared with a model that uses the real values of the variable. Thus, the increase in the error between the permuted and actual model predictions can be used to rank the importance of the variables. This thesis adds to the existing literature by using novel data sources and random forests models to develop the first empirical model of the vertical distribution of wildfire smoke, as well as the first empirical model to estimate 1-hour exposures to PM2.5 during wildfire seasons. 1.4 Review of critical literature on health effects  1.4.1 Short-term exposure to wildfire smoke The epidemiologic evidence on short-term exposure to wildfire smoke, measured at the time scale of days, has been summarized in several recent literature reviews (19-21, 109). These reviews have all come to three similar conclusions. First, there is consistent evidence on the association between wildfire smoke exposure and respiratory morbidity, including measures of sub-clinical symptoms, medication dispensations, physician visits, emergency room visits, and hospital admission. Specifically, exacerbations of asthma and chronic obstructive pulmonary disease (COPD) are the most commonly found impacts, while evidence for the associations with respiratory infections is inconsistent. Second, evidence on the association between wildfire smoke exposure and mortality, as well as cardiovascular morbidity, is inconclusive but growing. Among specific cardiovascular conditions, more evidence supported the association with out-of-20  hospital cardiac arrest (OHCA) and myocardial infarction is available compared with other cardiovascular outcomes. Third, children, the elderly, and those with pre-existing cardiopulmonary diseases appear to be more susceptible to the health effects of wildfire smoke exposure. Recent studies have provided more evidence on the association between wildfire smoke exposure and mortality from all causes. Shaposhnikov et al. estimated that exposure to air pollution from wildfires and its interaction with high temperatures contributed to more than 3000 deaths in Moscow in 2010 (110). Faustini et al. reported increases in all-cause, respiratory, and cardiovascular mortality associated with elevated PM during smoky days (33). More evidence has also emerged for the association with cardiovascular morbidity, such as OHCA, emergency room visits and hospital admissions for ischemic heart diseases (IHD), and hypertension (26, 67, 111-113).  There is also emerging evidence related to other health outcomes. A recent study found that wildfire smoke exposure was associated with paramedic assessments of hypoglycemia, a condition commonly found in people with diabetes (114). This was one of the very few population-based studies able to investigate the exacerbation of diabetic conditions from air pollution, which are not well-captured by many administrative health records. Exposures in utero have been associated with slightly reduced birth weight (70), while early life exposure may be related to immune dysregulation and compromised lung function in adolescence (115). In addition, some of the most recent studies have focused on health effects among vulnerable 21  populations, and identified higher risks in relation to obesity, lower education level, lower income, lower socioeconomic status, and minority ethnicity (30, 88, 116).  1.4.2 Sub-daily exposure to wildfire smoke To date there has been only one study on the health effects of sub-daily exposure to wildfire smoke. Dennekamp et al. (26) examined the association between hourly exposure to multiple air pollutants and OHCA based on ambulance dispatch data during one wildfire season in Melbourne, Australia. They found that the effects of PM2.5 and CO were significantly associated with OHCA only when averaged over the previous 48-hour period, while the effect of elevated O3 was observed within a 4-hour period. This demonstrated the feasibility of conducting population-based epidemiologic research on sub-daily exposure to wildfire smoke using ambulance dispatch data, but the study was limited by its short duration and constrained urban area, which did not include rural populations with higher exposure and less access to care. Furthermore, the study did not consider respiratory outcomes, which may be more sensitive to sub-daily exposures than cardiovascular outcomes, possibly due to more direct physiological impacts with shorter lag times. 1.4.3 Sub-daily exposure to PM2.5 Available reviews of studies on the effects of PM2.5 from other emissions sources have suggested that PM2.5 inhalation can trigger cardiopulmonary events within hours after exposure (117-119). A recent study found all-cause emergency department visits were increased within one hour after elevated levels of PM2.5 in Beijing, with the lagged effects lasting up to 10 hours (120). Other 22  studies have observed increased airway inflammation and decreased lung function in children with asthma and the elderly within one hour following exposure, with lagged effects lasting from 5 to 12 hours (121-124). In addition, results from epidemiologic studies have found that myocardial infarction and cardiac arrest within a few hours following exposure, mainly among the elderly and individuals with pre-existing cardiovascular diseases (125-129). Controlled exposure studies in humans suggest there is an increase in subclinical indications of acute cardiovascular responses within hours of elevated exposure to PM, including increased blood pressure, arterial stiffness and thrombus formation (130-133). There is also an increase in indicators of acute respiratory responses, including increased pulmonary inflammation and decreased peak expiratory flow (134-136). However, the evidence is inconclusive with respect to specific lag times between exposure and response, as well as the durations and magnitudes of exposure needed to trigger the response (117, 119). To further summarize the current evidence on sub-daily exposure to PM2.5 from all sources and cardiovascular conditions, we conducted an evidence review for studies on myocardial infarction (MI) and OHCA that used a case-crossover study design – the same study design that will be used for the analysis in Chapter 4 of this dissertation (Table 1.1). Although some studies had null findings (137-140), more found 11% to 48% increases in odds of cardiac events within a 2-hour window after the exposure (126, 129, 141-145). The lag patterns were inconsistent among these studies. While some found the effects disappeared right away after two hours, others found a gradual decline over time.  23  1.5 Use of ambulance data for epidemiologic studies To study the health effects of sub-daily exposure to wildfire smoke, it is necessary to have health data with resolved spatial and temporal resolution that cover a large population for a long period. The BC Emergency Health Service (BCEHS) database provides a unique opportunity for such a population-based study. The database includes records of all ambulances dispatched across the province over the study period, with the exact time at which the dispatch call was made, the exact location to which the ambulance was dispatched, and the medical reason for the call as assessed by (1) the dispatcher in conversation with the caller and (2) the attending paramedics. These features can overcome the limitations of other administrative data related to temporal and spatial specificity, thus allowing better assessment of the very acute effects of smoke exposures. The BCEHS database has been used in BC studies related to health care services, which made use of the high temporal resolution to study the effects of travel time on patient care outcomes (146).  As mentioned in Section 1.5.2, only one study has examined the effects of sub-daily exposure to wildfire smoke using ambulance dispatches to date (26). Similar databases have also been used in epidemiologic studies of 24-hour average exposure to wildfire smoke in other jurisdictions. Salimi et al. (147) observed a positive association between increased PM2.5 from wildfire smoke and ambulance dispatches for breathing problems on the same day and heart problems at a lag of two days. Haikerwal et al. (111) identified a positive relationship between increased PM2.5 over a 2-day moving average during wildfire events and OHCA. In addition, ambulance dispatches have also been used to study the health effects of ambient air pollution more generally. For 24  example, Johnston et al. (114) found that increased PM2.5 on the same day or at a lag of one day was associated with paramedic assessments of arrhythmia, heart failure, fainting, asthma, COPD, croup and hypoglycemia. Youngquist et al. (148) found associations with diabetic symptoms and fainting for same day exposures. Michikawa et al. (149) and Sajani et al. (150) found positive associations between increased PM exposure and ambulance dispatches for non-traumatic causes, especially those related to respiratory conditions.  Despite the strengths of ambulance dispatch data, there are some important limitations. First, as with any other administrative datasets, the data are not collected for research purposes and their quality may be variable with respect to sensitivity and specificity. For example, missing variables are more common in these data compared with other administrative datasets. Second, dispatch codes include broad categories of symptoms or health conditions that are not strictly defined. For example, the dispatch code “breathing problems” may be assigned to callers reporting shortness of breath, which can be a symptom for many different health conditions. Thus, these codes are more symptomatic than diagnostic. Third, most people calling for an ambulance have no medical training, meaning that their assessment of the medical emergency may differ significantly from that of a trained professional. The uncertainty introduced by all three of these limitations can be reduced by linking ambulance dispatches to any subsequent assessments by attending paramedics, and diagnoses for hospital admissions as I have done in Chapter 4.     25  1.6 Summary Wildfire smoke is a growing public health concern in BC, across Canada, and around the world, and more evidence is needed on the health effects of sub-daily exposures. With recent advancements in technology and methodology, I conducted this thesis to (1) develop tools for estimating population exposure at sufficient spatial and temporal resolutions and (2) assess the very acute health effect from these exposures. To the best of my knowledge, I have conducted the first study to examine sub-daily exposure to wildfire smoke and its effects on wide range of cause-specific health outcomes. The results of this study elucidate the lag structure of the exposure-response relationship and have implications for the appropriate time scale of air quality advisories and other public health actions during wildfire smoke events.  26  Table 1.1 Summary of case-crossover studies on myocardial infarction (MI) and out-of-hospital cardiac arrest (OHCA) Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Bhaskaran et al. 2011 National register of admissions for all hospitals in England and Wales from 2003 to 2006 79288 Discharge diagnosis of MI, both with and without ST segment elevation, and troponin positive acute coronary syndrome PM10: Median = 21 Interquartile range (IQR) = 14-30 PM10: 10  1-6  7-12  13-18  19-24  25-72  single pollutant: 1.012 (1.003-1.021) 0.993(0.983-1.003) 0.997(0.989-1.007) 0.992(0.982-1.002) 0.992(0.982-1.002) Multi-pollutant: 1.010 (1.000-1.020) 0.998(0.986-1.010) 0.997(0.985-1.009) 0.998(0.987-1.008) 0.996(0.982-1.010) Ensor et al. 2013 Non–dead-on- arrival adults from Texas Houston Fire Department emergency calls from 2004 to 2011  11677 OHCA Mean = 11.42 IQR = 7.34-14.37 6 0 1 2 3 4 1.009 (0.986-1.034) 1.011 (0.987-1.035) 1.011 (0.988-1.035) 1.003 (0.980-1.027) 1.009 (0.985-1.033) Evans et al. 2017 Acute coronary syndrome patients treated at one medical centre in Rochester, NY from 2007 to 2012 362 ST elevation MI Mean = 7.62 IQR = 3.20-10.30 Max = 79.20 7.1 1 1-12 1.17 (0.99-1.39) 1.11 (0.93-1.33) Gardner et al. 2014 Acute coronary syndrome patients treated at one medical centre in Rochester, NY from 2007 to 2010 338 ST elevation MI  Mean = 8.0 IQR = 3.9-10.2 Max = 43 7.1 6.8 6.1 5.9 5.4 0 0-2 0-11 0-23 0-47 1.18 (1.01-1.38) 1.15 (0.99-1.35) 1.12 (0.95-1.31) 1.11 (0.93-1.32) 0.94 (0.79-1.13) 27  Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Link et al. 2013 Patients with dual chamber implantable cardioverter-defibrillator recruited at Tufts Medical Center in Boston 176 subjects, 328 events Atrial Fibrillation Daily averages: Mean = 8.4 IQR = 5.3-10.2 6 5 1-2 1-24 1.26 (1.08-1.47) 1.14 (0.97-1.34) Madsen et al. 2012 Deaths in Oslo, Norway from 1992 to 2001 48713 Cardiovascular mortality Daily averages: Mean = 15.09 Range = 0.84-145.09 Daily Peak: Mean = 22.96 Range = 0.93-194.05 10  1-24 Daily average: 1.027 (1.003-1.05) Daily Peak: 1.017(1.003-1.031) Peters et al. 2001 Patients living in greater Boston area between 1995 to 1996 772 MI Mean = 12.1 5-95% = 2.6-29.6 25  1-2  1.48 (1.09-2.02) 28  Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Pradeau et al. 2015 Adults with activation of the medical intensive care unit in Gironde, France from 2007 to 2012. 4558 OHCA with presumptive cardiac etiology Mean = 16.3  Range = 13.3-79 10.5 0 1 2 3 4 5 6 7 8 1.11 (1.02 – 1.19) 0.89 (0.80 – 0.97) 0.96 (0.92 – 0.99) 1.01 (1.00 – 1.02) 1.05 (1.02 – 1.08) 1.05 (1.03 – 1.09) 1.04 (1.02 – 1.06) 0.99 (0.98 – 1.02) 0.94 (0.88 – 0.98) Raza et al. 2014 EMS calls in Stockholm, Sweden from 2000 to 2010 5973 OHCA Daily average: Mean = 8.1 Range = 0.14-161.7 10 1-2 1-24 0.99 (0.93 – 1.04) 0.99 (0.92 – 1.06) 29  Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Rosenthal et al. 2008 EMS in Indianapolis, US from 2002 to 2006 511 Non-dead-on-arrival OHCA witnessed by bystanders Median = 13.8  IQR = 8.8-20.7 10 0 1 2 3 4 5 6 7 0-3 0-7 1.12 (1.01-1.25) 1.09 (0.98-1.20) 1.03 (0.94-1.15) 1.03 (0.94-1.15) 1.03 (0.94-1.15) 0.98 (0.88-1.10) 1.00 (0.90-1.10) 1.03 (0.95-1.15) 1.09 (0.97-1.20) 1.05 (0.94-1.17) Rosenthal et al. 2013 EMS in Helsinki, Finland from 1998 to 2006 2134 All OHCA without secondary signs of death; OHCA caused by MI Mean = 8.7 IQR = 1–16.4 7.7  0 1 2 3 0-7 All OHCA 1.07 (1.01-1.13) 1.06 (1.01–1.12) 1.04 (0.99–1.10) 1.05 (1.00–1.11) 1.06 (1.00–1.13) MI-OHCA 1.14 (1.03–1.27) 1.14 (1.03–1.26) 1.11 (1.00–1.23) 1.08 (0.98–1.19) 1.12 (1.00–1.26) 30  Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Straney et al. 2014 Adults over 35 years old covered by Emergency ambulance services in Perth, Australia from 2000 to 2010 8551 OHCA with presumptive cardiac etiology Median = 6.80 IQR = 4.72-9.80 1  0 1 2 3 0-1 0-2 0-3 0-4 0-8 0-12 Single pollutant 0.999 (0.995-1.004) 1.001 (0.997-1.005) 1.001 (0.997-1.005) 1.002 (0.998-1.006) 1.000 (0.996-1.005) 1.001 (0.996-1.005) 1.001 (0.996-1.006) 1.003 (0.997-1.008) 1.006 (1.000-1.011) 1.007 (1.001-1.013) Multi-pollutant 0.997 (0.991-1.003) 1.001 (0.995-1.006) 1.003 (0.997-1.008) 1.003 (0.998-1.009) 0.999 (0.992-1.005) 1.001 (0.994-1.008) 1.002 (0.994-1.009) 1.005 (0.997-1.012) 1.011 (1.002-1.020) 1.010 (1.000-1.020) Sullivan et al. 2005 A database linking EMS data with hospital outcomes in King County, Washington State, US from 1988 to 1994 5793 Subject with discharge diagnosis of acute MI Mean = 12.8 Median = 8.6 IQR = 5.3-15.9 Range = 2.0-147 10 1 1-2 1-4  1.01 (0.98–1.05) 1.01 (0.97–1.05) 1.02 (0.98–1.04) 31  Study Setting and population Number of cases Outcome Min / mean / max hourly PM2.5 Delta PM2.5 Hours prior to outcome Effect estimate Wichmann et al. 2013 600,000 residents covered by Mobile Emergency Care Unit in Copenhagen from 1994 to 2010 4657 OHCA with activation of ambulance services Mean = 10.31 Median = 8.73 IQR = 6.20-12.10 5.9 0 1 2 3 4 5 6 7 0-4 0-8 1.01 (0.97-1.04) 0.98 (0.94-1.02) 0.98 (0.95-1.02) 0.97 (0.93-1.01) 0.97 (0.93-1.01) 1.00 (0.96-1.03) 0.98 (0.94–1.02) 0.95 (0.91–0.99) 0.98 (0.94-1.02) 0.97 (0.94–1.02) 32  Chapter 2: Predicting the minimum height of wildfire smoke within the atmosphere using machine learning and data from the CALIPSO satellite 2.1 Introduction To improve the applicability of remote sensing data for assessing population exposure to wildfire smoke, it is important to have information about the vertical distribution of smoke in the atmosphere with high spatial and temporal resolution. In this chapter, we applied random forests models to predict the minimum height of the smoke layer observed by the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) using information about fire activity in the vicinity, geographic location, and meteorological conditions. These data are typically available in near-real-time, ensuring that the resulting model can be operationally applied. The predictions can be 1) directly applied to smoke detections from the existing remote sensing products; 2) incorporated into statistical models with inputs from remote sensing products; or 3) used to inform estimates of injection height and downwind dispersion in deterministic models. These potential applications are expected to improve the assessment of ground-level population exposure to wildfire smoke for both epidemiologic research and public health surveillance.  2.2 Methods 2.2.1 The CALIPSO data and the response variable The CALIPSO satellite was launched in April 2006, as part of the US National Aeronautics and Space Administration (NASA) Afternoon Constellation (A-train), in a polar orbit with a 16-day 33  repeat cycle (Figure 2.1A). The primary instrument carried by CALIPSO is the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP). The CALIOP instrument generates laser pulses at two wavelengths (532 nm and 1064 nm) every 333 metres (m) along the satellite ground track, and then collects the returned backscatter signals. Analysis of these signals allows for discrimination between cloud and aerosol, and provides insight into the vertical distribution, shape, size, and type of detected aerosols (151). We obtained the CALIPSO Level 2 lidar vertical feature mask data (Version 3) from the CALIPSO Search and Subsetting Web Application for all overpasses in British Columbia (BC) during the study period. These data were stored in a block for each 5-kilometre (km) segment of the footprint along the swath (Figure 2.1A-B), with data cells at different vertical and horizontal resolutions depending on the altitude (Figure 2.1B-C). The vertical resolution is 30 m for all altitudes below 8.2 km, which are the most relevant to human activity and exposure. A detailed description of the algorithm used to classify features of the CALIPSO data can be found elsewhere (152-156). In brief, each data cell is classified as cloud, aerosol, or other, with a quality assessment (QA) of low, medium, or high. The distinction between cloud and aerosol is primarily based on the scattering strength and the spectral dependence of backscattering. If a data cell is classified as aerosol, its type is then further classified into clean marine, dust, polluted continental, clean continental, polluted dust, smoke, or other (Figure 2.1), with a binary QA of confident or not confident. Aerosol subtype classifications are primarily based on indicators such as depolarization, geophysical location, and backscatter intensity.  34  We defined our Minimum Height response variable as the minimum height of detected smoke above the land surface. To construct this variable, the first step was to identify all the CALIPSO data blocks that were impacted by smoke. A data block was considered to be smoke-impacted when the following two conditions were met: 1) any data cell within the data block was classified as aerosol with a high QA value, and as subtype smoke with a confident QA value; and 2) there were fire hotspots observed within 500 km of the data block within a 24-hour window of the smoke classification, spanning 12 hours on either side of the overpass. The fire hotspots were detected using data from the Moderate Resolution Imaging Spectroradiometer (MODIS), as described in the following section. The second step was to find the minimum altitude of the smoke layer. In each smoke-impacted data block, the Minimum Altitude variable was defined as the altitude of the lowest cell classified as aerosol above sea level, regardless of its subtype classification (Figure 2.1C). This approach was used because the subtype classification algorithm requires aerosol to be lofted before it can be classified as smoke (153). As a result, any smoke in the surface layer would have been classified as another type of aerosol. Given the limited sources of summertime aerosol pollution in most of the province and the fire hotspots requirement, we assumed that any aerosol underlying smoke in the 5-km data block was also smoke or a mixture of smoke and other aerosols. Finally, the Minimum Height variable was created by subtracting the elevation of the land surface at the location of the smoke detection from the Minimum Altitude variable to reflect the distance of the smoke layer from the land surface instead of the sea level. 35   Figure 2.1 Illustration of the CALIPSO data. An example of the footprint (A), the aerosol classification and vertical profile (B), and a data block (C) of the CALIPSO data. Panel C also shows the definition of smoke impacted block and the minimum altitude of aerosol in the block. Adapted from the CALIPSO product description (157).  2.2.2 Other data and the potentially predictive variables Because the horizontal resolution of the CALIPSO data blocks was 5 km, all of the data for deriving the potentially predictive variables were first rasterized into a base grid of 5 km2 over the province, and then the value of the cell in which the specific smoke detection fell was 36  extracted for analysis. We selected variables that were expected to be associated with the outcome, commonly available near real-time, and easy to understand even for users outside field of atmospheric science or meteorology. These variables were classified into four categories, as described below.  Fire Activity:  Information about active fires was retrieved from the MODIS active fire product (MCD14ML, Collection 6), distributed by the file transfer protocol (FTP) server maintained by the University of Maryland  (158). The MODIS instruments measure thermal infrared brightness temperature of the land surface, and active fires are identified where thermal abnormalities are detected in 1 km pixels (159). There are MODIS instruments aboard the Terra and Aqua satellites, each of which overpasses the province twice daily. The MODIS fire product provides the latitude and longitude of the fire pixels, the date and time of the fire detection, and its fire radiative power (FRP). The FRP is a quantitative measure of radiant heat output commonly used to approximate fire intensity, which is proportional to its combustion rate and smoke emissions (160-162). Fire hotspots were extracted from the database and rasterized to the 5km base grid if they were (1) observed 12 hours before or after each CALIPSO smoke detection and (2) classified as presumed vegetation fire in the fire product. Both the scan angle and the bowtie effect can erroneously increase FRP near the edge of the satellite image swath (163, 164). To account for this, we first calculated the scan angle and pixel size for each fire detection using the number of sampled pixels provided in the fire product (165). The summed FRP assigned to each 5km grid was then adjusted for the scan angle and pixel areas using previously described methods (166). The detailed methods for these adjustments are presented in 0.  37  Five potentially predictive variables were derived from the rasterized FRP data for each smoke detection (Table 2.1). The FRP Within 500km was calculated as summed FRP values for all fire cells within 500 km of the smoke detection. This variable was also used to evaluate the second criterion for defining smoke-impacted data blocks described in the previous section, where only observations with FRP Within 500km larger than zero were included in the analysis. The Closest FRP Cluster was the summed FRP value for the fire cell cluster closest to the smoke detection. A fire cell cluster was defined as a patch of neighboring fire cells that shared any boundary using the Queen’s case criterion (167). The IDW FRP applied the inverse distance weighted (IDW) average to all fire cells within 500 km of the smoke detection, where each fire cell was given a weight as the inverse of its distance to the smoke observation (168). This variable was created to reflect the fact that closer fires may have a larger impact on local smoke than more distant fires. The Minimum Distance and Direction variables indicate the distance and the direction in degrees bearing from the smoke detection to the nearest fire cell, respectively.  Meteorology: Meteorology plays an important role in smoke emissions and dispersion. As we were interested in developing an empirical model that could be operationalized in near real-time with readily available data, we did not attempt to retrospectively simulate the complex processes of fire behavior and pollutant transport as has been done in deterministic models. Instead we used a limited set of meteorological parameters accessible in near real-time and applicable in our machine learning approach.  The planetary boundary layer (PBL) is the lowest part of the atmosphere that is directly influenced by the surface of the earth, where turbulent air flow and vertical mixing are usually 38  strong. The height of the planetary boundary layer (PBLH) above the surface varies in space and time, with a strong diurnal cycle over land (169). We retrieved hourly PBLH estimates from the two-dimensional surface turbulent flux diagnostics produced by the NASA Modern Era Retrospective-analysis for Research and Applications (MERRA) program (170) at a spatial resolution of 2/3-degree longitude by 1/2-degree latitude. The raw data were averaged to the 5km base grid to generate a PBLH variable value in each cell.  A similar approach was used for the eastward and northward wind components at 50 m above the surface (E50m and N50m), respectively, to represent the wind patterns close to the surface. We also created variables from the eastward and northwards components of winds at three different air pressure levels: 850 hPa (E850hPa, N850hPa); 500 hPa (E500hPa, N500hPa); and 250 hPa (E250hPa, N250hPa). These variables can represent wind patterns at the lower, middle and upper troposphere, respectively. All wind data were retrieved from the two-dimensional atmospheric single-level diagnostics in the MERRA product, also at a spatial resolution of 2/3-degree longitude by 1/2-degree latitude. Fire-Meteorology combined variable: The Wind-Weighted Minimum Distance and Wind-Weighted Closest FRP Cluster variables were created to link fire, smoke, and meteorological conditions together. Because the movement of smoke can be facilitated or impeded by the wind field, these two variables calculated (1) the minimum distance to the closest fire cluster and (2) the FRP of that fire cluster, accounting for whether the relative locations of fire and smoke were aligned with the wind direction. The calculations included the following steps: (1) for each smoke observation, the direction from each fire cluster detected 12 hours before and/or after was 39  calculated in degrees bearing; (2) trigonometry was used to construct a 5km × 5km wind direction raster (in degrees bearing) using the two 850 hPa Cartesian wind components (E850hPa and N850hPa) at the time of the smoke observation; (3) a cost surface was created based on the difference between the fire-to-smoke direction and the wind direction in each grid cell, where larger distances had a higher cost (i.e. differences of 0-30, 30-60, 60-90, 90-120, 120-150, 150-180 degrees were assigned cost values of 1 through 6, respectively); and (4) the least-cost distance between the smoke observation and each fire was calculated using the costDistance function in the gdistance package in R (171). The ultimate result is that FRP clusters upwind of the smoke observation are closer in weighted distance than those that are downwind but at the same geographic distance. The FRP of the fire cluster with the smallest cost-weighted distance was assigned as Wind-Weighted Closest FRP Cluster, and the value of the cost distance was assigned as the Wind-Weighted Minimum Distance for the smoke observation. Daytime and Month variables were created to capture further diurnal and seasonal patterns in fire activity and meteorology not covered by existing variables. Because CALIPSO only passes over BC from 01:00 – 03:00 (night) or 12:00 – 13:00 (day) local time, the Daytime variable cannot capture the full range of nighttime or daytime conditions under which smoke and fires occur. Geographic Location: The elevation of the land surface was obtained from the GTOPO30 product, a global digital elevation model with a horizontal grid spacing of 30 arc seconds (approximately 1 km at the equator), developed by the US Geological Survey Earth Resources Observation and Science (EROS) Center (172). The raw data were assigned to the 5km base grid as the average to generate values for the variable Elevation in each cell. The latitudes and 40  longitudes of the smoke-impacted CALIPSO data blocks were used to create the Latitude and Longitude variables, which could reflect other important geospatial information, such as terrain features and land use. Because our model was developed with the intention of future operational use in near-real-time, we ensured that all variables we included could be derived in near-real-time. Although we used meteorological variables from retrospective MERRA project for the model described here, comparable variables could easily be derived from operational weather forecasting systems. Similarly, the FRP data used in our model were derived from a product with retrospective quality assurance procedures, typically available months after the actual retrieval date, but the Fire Information for Resource Management System (FIRMS) maintained by NASA also provides a near-real-time FRP product. To operationalize this model, we could simply replace the retrospective data sources with their near-real-time substitutes, although further evaluation would be needed to assess model performance with data of different quality.      41  Table 2.1 Description and summary statistics of variables for minimum height model CONTINUOUS VARIABLES Variables Median [Quartiles] Description Minimum Height (m) * 293 [59, 1416] Response variable. Minimum height of aerosol above land surface in smoke-impacted CALIPSO data block  PBLH (m) 458 [138, 1310] Planetary boundary layer height above land surface Elevation (m) * 1016 [749, 1293] Elevation of the land surface above sea level Latitude (degree) * 54.21 [50.795, 57.13] Latitude of the smoke-impacted data block. Longitude (degree) * -123.4 [-126.1, -121.1] Longitude of the smoke-impacted data block FRP Within 500km (MW) * 161 [33, 769] Summed FRP for all fire cells within 500 km Closest FRP (MW) * 15 [7, 38] Summed FRP for the closest fire cell cluster IDW FRP (MW) * 29 [13, 77] The inverse distance weighted averaged FRP for fire cells within 500 km Minimum Distance (km) 214 [115, 340] Distance to the closest fire cell  Wind-Weighted Closest FRP (MW) * 13 [7, 35] Summed FRP for the fire cell cluster with the shortest wind-direction cost-weighted distance Wind-Weighted Minimum Distance (cost unit) 51 [28, 84] The accumulative cost distance from the closet FRP weighted to smoke E50m (m/s) * 1.37 [-0.72, 2.93] Eastward wind at 50 meter above surface N50m (m/s) * 0.22 [-1.20, 1.70] Northward wind at 50 meter above surface E850hPa (m/s) 1.84 [-0.63, 4.10] Eastward wind at 850 hPa pressure level N850hPa (m/s) * 0.15 [-2.00, 2.37] Northward wind at 850 hPa pressure level E500hPa (m/s) 5.65 [1.35, 10.40] Eastward wind at 500 hPa pressure level N500hPa (m/s) * 0.52 [-4.49, 6.53] Northward wind at 500 hPa pressure level E250hPa (m/s) * 9.74 [1.73, 19.26] Eastward wind at 250 hPa pressure level N250hPa (m/s) -0.94 [-10.48, 8.49] Northward wind at 250 hPa pressure level 42  CATEGORICAL VARIABLES Variables N (%) Description Direction†  [0, 90] 4160 (26.6) Direction as degrees bearing to the closest fire cell  (90, 180] 5129 (32.9) (180, 270] 3263 (20.9) (270, 360] 3065 (19.6) Daytime Day 3592 (23.0) Binary variable indicating whether smoke was detected in daytime or nighttime overpass  Night 12025 (77.0) Month* April 1392 (8.9) Categorical variable indicating the month of the smoke detection May 1867 (12.0) June 1405 (9.0) July 3892 (24.9) August 4746 (30.4) September 2315 (14.8) * Variables included in the final model. † Direction was calculated as degrees bearing (0 to 360) and used as a continuous variable in the model. It is categorized here for a more sensible descriptive summary.  2.2.3 Statistical modelling All data cleaning and analysis was conducted in the R Statistical Computing Environment (R Core Team, Vienna, Austria). The data were fitted with a random forests model with 500 regression trees and a subset of four predictive variables sampled for each tree, using the randomForest package in R (173). Each tree was trained with a selected subset of data, and predictions were made with the subset of data that were not sampled for training, otherwise known as the “out-of-bag” data.  Because each data point fell into the out-of-bag subset multiple times out of the 500 trees, it had multiple predicted values for the Minimum Height response variable. The average of these values 43  was reported as the final out-of-bag prediction for that data point, and the final out-of-bag predictions for all data points were used to calculate the root mean square error (RMSE) and the R-squared (percentage of variance explained by the model) against the observations. These values were used to evaluate the model performance.  The importance of each variable was calculated using a permutation test on the out-of-bag predictions (108). If a predictive variable is important, a random permutation of the variable will considerably increase the model prediction errors when compared with a model that uses the real values of the variable. Thus, the increase in the RMSE between the permuted and actual model predictions, scaled by the standard deviation of the difference, was used to rank the importance of the variables. A backward selection process was applied to select the final model. All variables were included in the initial model, then a reduced model excluding the least important variable would be fitted. If the RMSE and R-squared values for the reduced model were not significantly different from those for the initial model, the variable would be permanently excluded. The process was then repeated with the least important variable in the new model, which was the reduced model from the previous step. The process stopped, and the final model was determined when the removal of the least important variable resulted in any reduction of R-squared values larger than 1% of the highest R-squared achieved in the previous models. This process was aimed to obtain the most predictive model with as few variables as possible, which will be useful for operationalizing the model in near real-time.  44  The out-of-bag predictions from the final model were also used to evaluate its accuracy, sensitivity (true positive rate), and specificity (true negative rate) in predicting whether Minimum Height was below or above thresholds of 100, 300, 500 and 1000 m. These thresholds cover the range of possible human exposure. Accuracy, sensitivity and specificity were calculated as the percentage of correct predictions among all observations, among observations below the threshold, and among observations above the threshold, respectively.  2.3 Results A total of 15,617 CALIPSO data blocks met our criteria for classification as smoke-impacted and had complete information for all the potentially predictive variables. They were identified on 887 of the 1761 dates in the study period, and more than half (55.3%) occurred in July and August (Table 2.1), which is the peak wildfire season in the province. Furthermore, more than half (52.7%) had Minimum Height values below the PBLH at the corresponding time and location, meaning that smoke was more likely to be mixed in the boundary layer and to be affecting surface air quality.  The final random forests model after the backward selection process included 13 of the 21 variables (Table 2.1). The final model explained 82.1 % of the variance in the out-of-bag Minimum Height observations, and the RMSE between the out-of-bag predictions and the observations was 560 m (Figure 2.2), which was large compared with the vertical resolution of the CALIPSO data (30 m at minimum). There were generally more over-predictions when the observed Minimum Height values were below 500 m. A descriptive analysis of the data in the cluster of over-predictions with observed Minimum Height below 500 m found that they tended 45  to be at lower elevation, to be farther from fires, and to have lower PBLH values when compared with the other observations (Table 2.2). On the other hand, under-predictions were more uniform across the range of observed values.  In addition, the model predicted whether Minimum Height was below the 100 m, 300 m and 500 m thresholds with low to moderate sensitivity (0.31-0.80) and high specificity (0.90-0.96). This means that when the observed Minimum Height was below the respective thresholds, the model predicted correctly 31% to 80% of the time. Meanwhile, when the observed Minimum Height was above the thresholds, the model predicted correctly 90 to 96% of the time. In general, the accuracy and specificity increased considerably as the threshold increased, while the sensitivity decreased slightly (Table 2.3). In other words, the model became more accurate at predicating smoke below the threshold as the threshold height increased, and the model was quite accurate at predicting smoke above the threshold for all threshold values. This indicates that the model is especially useful for ruling out smoke that is not relevant to surface exposures. The most important variable in the model was the speed of eastward wind at 50 m above the surface. There was a 97 % increase in the normalized mean squared error when predictions were made using the randomly scrambled version of this variable compared with predictions using the actual values. The three most important variables in the model were all related to wind pattern, followed by variables related to fire activity, including Month and FRP within 500km (Figure 2.3). 46  Table 2.2 Descriptive analysis of the low altitude observations in the cluster of over-prediction This table summaries statistics of the variables for observations in the cluster of over-prediction where the Minimum Height values were smaller than 500 m. Variable Observations in the over-prediction cluster All other observations P value PBLH (m) 628 [599, 657] 787 [775, 800] <0.05 Elevation (m) 996 [976, 1015] 1020 [1014, 1027] <0.05 IDW FRP (MW) 67 [63, 72] 65 [63, 66] 0.24 FRP Within 500km (MW) 707 [643, 771] 769 [739, 800] 0.08 Minimum Distance (km) 247 [241, 254] 227 [225, 230] <0.05 Closest FRP (MW) 54 [47, 62] 61 [58, 65] 0.09 Wind-Weighted Closest FRP (MW) 43 [39, 47] 51 [48, 54] <0.05 Wind-Weighted Minimum Distance (cost unit) 67 [64, 69] 62 [61, 63] <0.05 E50m (m/s) 0.97 [0.86, 1.08] 1.08 [1.04, 1.12] 0.05 N50m (m/s) 0.18 [0.09, 0.27] 0.25 [0.21, 0.28] 0.20 E850hPa (m/s) 1.87 [1.71, 2.03] 1.77 [1.71, 1.84] 0.29 N850hPa (m/s) 6.34 [5.97, 6.71] 6.18 [6.05, 6.30] 0.41 E500hPa (m/s) 11.47 [10.8, 12.12] 10.92 [10.69, 11.14] 0.12 N500hPa (m/s) 0.33 [0.18, 0.49] 0.19 [0.13, 0.24] 0.08 E250hPa (m/s) 1.25 [0.87, 1.62] 0.76 [0.62, 0.90] <0.05 N250hPa (m/s) -0.42 [-1.12, 0.28] -1.59 [-1.84, -1.33] <0.05  47   Figure 2.2 Scatterplot comparing observations and predictions for minimum height model. This figure shows the comparison between observations of minimum height from CALIPSO and the out-of-bag predictions from the model. The red line represents the 1:1 perfect match. When Minimum Height observations were below 500 m, there were generally more over-predictions (data points above the red line), while under-predictions (data points below the red line) dominate in other cases.  Table 2.3 Model performance for predictions of Minimum Height below thresholds  Minimum Height <  100 m 300 m 500 m 1 km Accuracy  0.75 0.78 0.84 0.91 Sensitivity 0.31 0.65 0.80 0.94 Specificity 0.96 0.91 0.90 0.85 *Accuracy = Percent of correction predictions among all observations   Sensitivity = Percent of correct predictions among observations over the threshold   Specificity = Percent of correct predictions among observations below the threshold  48    Figure 2.3 Dotchart of variable importance of minimum height model. Variable importance was measured by the normalized percent increase in mean squared error (MSE) when the variable was randomly permutated. Variables are grouped into four categories related to fire activity, geographic location, meteorology and fire-meteorology combined. Only variables in the final model after the model selection procedures were included in this figure.  2.4 Discussion A machine learning approach was applied to predict the minimum height of smoke in the atmosphere using variables that reflected fire activity, geographic location, and meteorology, and the combination of fire activity and meteorology. The model explained 82% of the variability in the observations. The comparison of predictions and observations with binary categories for smoke under 100, 300, 500, and 1000 m (Table 2.3) indicated high specificity but low to moderate sensitivity in predicting smoke height below a certain threshold. The most important 49  variables in the model were wind components, followed by month of observations and fire intensity within 500 km.  Several studies have used deterministic models of plume rise and evaluated them against remote sensing measurements from MISR or CALIPSO (174) but, to the best of our knowledge,  ours is the first statistical model to be proposed. The R2 value for our model was 0.82 compared with values from 0.10 to 0.30 for deterministic models of plume rise that were evaluated using observations from MISR (78-80, 83). Most of these studies predicted the initial smoke injection height over specific fires, so do not have directly comparable objectives. However, Raffuse et al. also compared the downwind plume height with CALIPSO observations, and found weak correlation for both the injection height (R2 = 0.10) and the downwind height (R2 = 0.22). The improved performance in our model may be due to the following factors: (1) increased predictive power due to the large dataset; (2) the application of the flexible random forests machine learning approach; and (3) the inclusion of a wide variety of fire activity, geographic location, and meteorological variables. While the model explained much of the variability in the Minimum Height observations, the RMSE for the out-of-bag estimates was 560 m, which was high relative to the zone of human exposure given that the tallest building in the study area was less than 300 m. However, the results for the 100, 300, 500, and 1000 m thresholds suggest that the continuous predictions from the model also perform reasonably well when dichotomized to make binary predictions. To further evaluate this, we also built a random forests model with the same input variables to predict the binary outcome for the 300 m threshold, which resulted in more balance between the 50  sensitivity (0.84) and specificity (0.83) when compared with those reported in here (0.65 and 0.91, respectively). This machine learning approach can be easily adapted to different types of outcomes to achieve the optimal performance depending on the purpose.  The most important variables in the model were wind patterns at different altitudes, indicating the critical influence of meteorological conditions on the vertical distribution of smoke. In this study the smoke observed by CALIPSO could have been at any stage in the dispersion process (initial emission, local dispersion, or regional transport) because the smoke-impacted data blocks were not causally associated with specific fires prior to analysis. In fact, only 73 out of the 15617 observations were within 10 km of the closest MODIS detected fire, suggesting that most of the smoke was observed during downwind dispersion instead of initial emission. It follows that atmospheric conditions are more influential than fire intensity in our model, which is not consistent with previous studies focusing only on initial injection height (81, 82). Another contributor to the difference between this study and the previous studies might be that we modeled the minimum height of the smoke layer, while most previous studies modeled the average or the maximum height. The Month variable may be important because it reflects seasonal changes in both fire behavior and meteorological conditions. Our dataset included much larger proportion of observations made during nighttime overpasses compared with daytime overpasses (Table 2.1). This seems inconsistent with previous studies that found that either (1) fire intensity and smoke emissions peaked during the day (27, 175, 176), or (2) there was no obvious diurnal cycle during extreme forest fire events (166). However, this imbalance between nighttime and daytime smoke observations has also been reported by 51  other studies using CALIPSO data (177), and might be attributable to differences in data quality between the nighttime and daytime products. Backscatter retrieval during daytime is substantially influenced by solar background light, resulting in sub-optimal data quality in the subsequent high-level products, such as feature classification (178). Because we included smoke observations flagged with high data quality, a larger proportion of daytime data may have been omitted.  Although the random forests approach offers good predictive power, the interpretability of models is limited when compared with general linear regression. It is difficult to illustrate the relationship between the predictors and the outcome within the model. The partial dependence plot (179) is a common way to investigate that relationship in random forests, where the marginal effect of a single predictor is shown by averaging the model predictions at each value of the predictor. Partial dependence plots (Figure 2.4) for our model show some intuitive relationships. For example, the minimum height tends to be lower when wind speed is small, indicating more stable atmospheric conditions. However, these plots cannot depict any high-dimensional interactions between predictive variables in the model (179) and any interpretation should be cautious. 52   Figure 2.4 Partial dependence plots for minimum height model. These plots show the marginal effects of variables on predicting minimum height of the smoke, arranged in order of variable importance. The x-axis indicates the values of the predictive variables, and the y-axis displays the average Minimum Height of smoke observations corresponding to each value of the predictive variable.  53  We imagine the results from this model being applied in three different ways. First, they could be directly applied to existing smoke-related remote sensing products to provide another dimension of information on ground-level impacts. For example, the widely used Hazard Mapping System produces smoke plume outlines drawn by trained analysts after considering images from multiple satellites near real-time (180). Our predictive model could add another layer to these data and give the users a perspective on the relevance of the smoke in terms of population exposure (Figure 2.5). Second, the model predictions or the variables in the model could be incorporated into operational empirical models such OSSEM, which estimate or forecast ground-level smoke concentrations using remote sensing data as the primary input. Several these models have been developed over the past few years (42, 71, 96), and better information about smoke plume height may improve their performance for public health and other applications. Finally, the predictions can be used to validate or calibrate the estimates of injection height or downwind vertical distribution from deterministic models such as BlueSky and FireWork.   There are several limitations with this study. First, the model included some meteorological variables that are readily accessible in near-real-time, but we were unable to account for the complete temporal and spatial aspects of atmospheric transport processes. However, the high predictive power, simplicity of the modelling procedure, and easy access to the variables suggest that it can be valuable for operational purposes and that it can complement conventional modelling efforts. This approach is especially useful when resources and expertise are lacking to develop and maintain a complex deterministic smoke model.  It can also provide valuable information from a purely data-driven perspective when our current understanding is limited or inconclusive on some physical processes of smoke emission and transport. Second, the 54  CALIPSO vertical feature mask product has associated uncertainties and errors. Any error in its aerosol or smoke classifications would be translated into the minimum height estimates, and thus propagated into our final model. However, several studies have suggested reasonable performance of the CALIPSO classification algorithm, where the rate of correctly distinguishing aerosols and clouds is approximately 90%, with occasional misclassification of dense smoke as cloud (155, 181). Third, we used the minimum height of any aerosol in the smoke-impacted data blocks as the minimum height of the smoke, regardless of its subtype classification. The assumption that this aerosol included smoke may not have been valid in all cases, and thus may have biased our minimum height estimates. However, this decision affected only 15% of the observations used in the model, and excluding these observations resulted in a similar model with respect to variance explained and variable importance ranking. Fourth, the training data were restricted to smoke observations with MODIS fires detected within 500 km, so the model could not account for smoke from continental transport or for smoke from small fires, smoldering fires, or fires not detected by MODIS. Although BC has been affected by smoke from continental transport, a recent study reported that most smoke in the province originated from fires in the province (182). Small fires have always posed a challenge to models that rely on remote sensing data, but recent developments in fire detection instruments and algorithms may address this limitation in the near future (183). The limitations of MODIS fire detection could also affect the quality of the variables involving FRP in this study. Finally, by using remote sensing data to identify fires, we could not distinguish between different types of landscape fires, such as forest fires and pile burning, which might have different dynamics with respect to smoke emissions and dispersion. Hazard reduction burning of waste wood from the forest sector is 55  common in the province, especially in the early spring and late fall, and the air quality impacts can be pronounced under some conditions (184).   Figure 2.5  Illustration of potential application of the minimum height model. The three smoke plumes outlines (1, 2, and 3) from the Hazard Mapping System on July 6, 2015, can be combined with the minimum smoke height model estimates to better contextualize the vertical profile. The grey line indicates the boundary of British Columbia.  We have presented a statistical model to estimate the minimum height of forest fire smoke in the atmosphere using data on fire activity, geographic location, and meteorology. The model estimates can be used to improve the relevance of operational smoke-related remote sensing products and statistical or deterministic models intended for population exposure assessment. The machine learning approach performed well, and could easily be applied in other smoke-affected regions or for predicting other measures of the vertical distribution, such as binary indicators of smoke within the surface layer.   56  Chapter 3: A machine learning approach to estimate hourly exposure to fine particulate matter for urban, rural, and remote populations during wildfire seasons 3.1 Introduction To study the health effects of sub-daily exposure to wildfire smoke, there is a need to assess population exposure at high spatial and temporal resolutions. The Optimized Statistical Smoke Exposure Model (OSSEM) is an empirical model previously developed to estimate 24-hour average concentrations of fine particulate matter (PM2.5) for the entire population of British Columbia (BC). It applied linear regression to integrate data from multiple sources, including remote sensing images of fire and smoke, routine air quality monitoring, and meteorological models (42). It has been operationalized for public health surveillance since 2012 (98). Here, an expansion of the 24-hour model (OSSEM-24h) is developed to build a 1-hour model (OSSEM-1h) that estimates PM2.5 exposures for the same population at the same resolution using more advanced statistical modeling techniques and data sources with higher temporal resolution. The study in Chapter 2 provided a thorough assessment of factors associated with the presence of smoke at the surface level, where populations are exposed, and identified relevant variables for this work. The OSSEM-1h model will be useful for conducting epidemiologic studies on sub-daily exposure to wildfire smoke, as well as informing public health actions if operationalized in near-real-time. 57  3.2 Methods 3.2.1 Data sources and variables The following data sources were used to derive the response variable and potentially predictive variables for the OSSEM-1h model, based on previous work to develop the OSSEM-24h (42) and vertical height models in Chapter 2. The spatial resolutions of the input datasets described below varied between 1 and 50 km, so a prediction grid with a resolution of 5 km2 was created for the province. Values for each of the potentially predictive variables were assigned to each cell. If more than one value was available for a single cell, the mean value was assigned for continuous variables and the closest value to the cell centroid was assigned for categorical variables. If no value was available within a cell, the value closest to the cell centroid was assigned. PM2.5 measurements: We obtained 1-hour average PM2.5 measurements from 72 air quality monitoring stations from the Provincial Air Data Archive Website maintained by the BC Ministry of Environment and Climate Change Strategy. The total number of 1-hour observations available at each station during the study period ranged from 2263 to 19325 out of the 26352 hours. These 1-hour average data were used as the PM 1-hour response variable for model training, as well as the observations against which we compared the predictions for model evaluation. The response variable PM 1-hour was log-transformed due to its right-skewed distribution. The calendar month of the observation (Month) was included as a potentially predictive variable based on results in Chapter 2.  58  Ecozone: The provincial digital Biogeoclimatic Ecosystem Classification map (Version 10.0) was obtained from the GeoBC Data Discovery Service. The province was classified into 16 ecozones named by the major vegetation species present in each (Figure 3.1), which have similar growing conditions due to a broad, homogeneous macroclimate (185, 186). While not temporally variable over the study period, these data may indicate how smoke emissions factors vary by fuel type and climate (50, 187-189). The potentially predictive variable Ecozone was assigned to each grid cell as the ecozone of the nearest fire.  Figure 3.1 Areal extent of Biogeoclimatic Zones in British Columbia.  This figure shows the percentage of area of each type of Biogeoclimatic Zones in BC, reported by Austin et al. in 2008 (190).  59  Fire activity: Data from the Moderate Resolution Spectroradiometer (MODIS) instruments aboard the polar orbit Aqua and Terra satellites were used to assign fire locations and intensity. Fire locations were taken as the centroids of all 1 km by 1 km fire pixels identified in the active fire product (MCD14ML, Collection 6), and fire intensities were approximated by the fire radiative power (FRP) variable. Because these two satellites only passed over the study area four times daily, information was not available for every hour. As a result, hotspots caused by presumed vegetation fires observed within the 12-hour window before or after the 1-hour prediction period were rasterized to the prediction grid cells. Five potentially predictive variables were derived from these data (Table 3.1): (1) FRP Within 500km as summed FRP values for all fires within 500 km of the grid cell; (2) Closest FRP Cluster as the summed FRP value for the closest fire cluster, where a cluster was defined as any group of grid cells with fire detections that shared any boundaries using the Queen’s case criterion (167); (3) IDW FRP as the inverse distance weighted (IDW) average to the centroids of all fires within 500 km, which would reflect larger impacts from closer fires; and (4) Minimum Fire Distance as the distance to the nearest detected fire; and (5) Fire Direction as the direction in degrees bearing from the grid cell to the nearest fire. More details on the derivation of these variables can be found in Chapter 2 (191). Meteorology: Hourly meteorological information was retrieved from the NASA Modern Era Retrospective-analysis for Research and Applications (MERRA) program, available at a spatial resolution of 2/3-degree longitude by 1/2-degree latitude. We used these data to generate potentially predictive variables for the planetary boundary layer height (PBLH) and the eastward and northward wind components at 50 m above the surface (E50m, N50m). We also extracted eastward and northward winds at the 500 hPa (E500hPa, N500hPa) and 250 hPa (E250hPa, 60  N250hPa) pressure levels. These variables were predictive of the dispersion of wildfire smoke in previous studies (96, 191). More details about these variables can be found in Chapter 2 (191). Elevation: The average elevation of the grid cell was calculated using data from the GTOPO30 product from the US Geological Survey Earth Resources Observation and Science Center. This is a global digital elevation model with a horizontal grid spacing of approximately 1km by 1km.  3.2.2 Model building and evaluation 3.2.2.1 Data reduction Given the large number of 1-hour PM2.5 observations available, using all observations for the modeling would be computationally expensive. In addition, a large proportion of the observations reflect low, background PM2.5 concentrations, which are not suited to the objective of modeling the air quality impacts of wildfire smoke. In consideration of these factors, we chose to train and evaluate OSSEM-1h using a reduced dataset. This included all observations with PM2.5 concentration > 15 µg/m3, approximately the 95th percentile of all 1-hour PM2.5 observations, and twice as many randomly-selected observations with PM2.5 concentrations  15 µg/m3 (Figure 3.2).     61  Table 3.1 Data sources and variables for OSSEM-1h. Variable Description Data source Response variable PM 1-hour 1-hour PM2.5 from 72 monitoring stations  BC Ministry of Environment Potentially predictive variables Month Categorical variable indicating the calendar month   Ecozone Ecozone from the Biogeoclimatic Ecosystem Classification map, where the closest fire is observed BC Ministry of Forests and Range FRP Within 500km Summed fire radiative power (FRP) for all fire within 500 km US National Aeronautics and Space Administration (NASA) data from the Moderate Resolution Imaging Spectroradiometer (MODIS) Closest FRP  Summed FRP for the closest fire cluster IDW FRP  The inverse distance weighted averaged FRP for fire within 500 km Minimum Fire Distance  Distance to the closest fire  Fire Direction Direction as degrees bearing to the closest fire   PBLH  Planetary boundary layer height above land surface  NASA Modern Era Retrospective-analysis for Research and Applications (MERRA) program E50m  Eastward wind at 50 meters above surface N50m  Northward wind at 50 meters above surface E500hPa  Eastward wind at 500 hPa pressure level N500hPa  Northward wind at 500 hPa pressure level E250hPa  Eastward wind at 250 hPa pressure level N250hPa  Northward wind at 250 hPa pressure level Elevation Elevation of the land surface above sea level US Geological Survey Earth Resources Observation and Science Center 62   Figure 3.2 Flowchart of data reduction for OSSEM-1h model training.  3.2.2.2 Model training and out-of-bag evaluation  The reduced data were fitted with a Random Forests model, which is an ensemble of regression trees (108). Each tree was trained with a random subset of the predictive variables on a random subset of the reduced data. We constructed the Random Forests models with a total of 500 trees using five predictive variables in each tree. To evaluate the model performance, predictions were made with the unsampled data from each tree, otherwise known as the “out-of-bag” data. Because a single observation could fall into the out-of-bag subset multiple times, all values predicted for the same PM 1-hour observation were averaged to obtain the final out-of-bag prediction. Four statistics were calculated to compare the out-of-bag predictions against the observations: (1) Pearson’s correlation coefficient; (2) root mean squared error (RMSE); (3) mean fractional bias (MFB); and (4) mean fractional error (MFE) (Table 3.2). The out-of-bag 63  predictions were also examined on days with different fire activity levels. The provincial sum of FRP was calculated for each day of the study period, and days with FRP in <40 percentile, 40-80 percentile and >80 percentile were categorized as low, moderate and high fire days, respectively. Out-of-bag predictions from the model were compared with observations stratified by these groupings. Table 3.2 Statistics used for evaluation of the OSSEM-1h model.  These statistics evaluate the model predicted values (P) compared with the observations (O) over the entire dataset (n). Statistics Equation Range Pearson’s correlation coefficient (r) r = 𝑛 ∑ 𝑃𝑂−∑ 𝑂 ∑ 𝑃 √[𝑛 ∑ 𝑂2−(∑ 𝑂)2][𝑛 ∑ 𝑃2−(∑ 𝑃)2] [-1, 1] Root mean squared error (RMSE) RMSE = √∑(𝑃−𝑂)2𝑛 [0, +∞] Mean fractional bias (MFB) MFB = 1𝑛 ∑(𝑃−𝑂)(𝑃+𝑂2) [-200%, 200%] Mean fractions error (MFE) MFE = 1𝑛 ∑|𝑃−𝑂|(𝑃+𝑂2) [0, 200%]  The importance of each model variable was assessed by generating a random permutation of the variable, calculating the increase in model prediction error with the new values, and comparing it with the error in the model that used the real values. If a variable is important for prediction, the error will be dramatically increased when that variable is replaced with random values. A backward variable selection process was used based on variable importance. If a reduced model without the least important variable performed better or the same as the full model in out-of-bag prediction, that variable would be permanently removed from the final model. Otherwise, the variable would be kept, and the full model was regarded as the final model. 64  3.2.2.3 Leave-region-out cross-validation To further test the model performance in areas away from the locations of the training data, a leave-region-out cross-validation was conducted. For the 72 monitoring stations used in the study, those within 50km of each other were clustered as a regional group. The 50km criterion was chosen based on a previous study that found the median distance from the populated grid cells of the model to the nearest monitor was approximately 50km (42). The clustering produced a total of 30 groups, each of which included 1-4 stations, except for the regional group in the greater Vancouver area, where 13 stations were clustered into one group. We then trained 16 models, 15 of which excluded data from a random selection of two regional groups, and one of which excluded data from the greater Vancouver regional group. Then we used the models to make predictions for the excluded stations. Using this spatial cross-validation design, we were able to test model performance when none of the training stations were within 50km of the testing locations. The same four evaluation statistics (Table 3.2) were calculated for each station and overall. 3.2.2.4 Case studies In addition to the province-wide quantitative evaluation, the final model was qualitatively evaluated using two case studies. The first one examined an extreme fire and smoke event that occurred in August 2010 across the central interior region of BC (28), which is frequently affected by seasonal wildfires and smoke (54) (Figure 3.3). The second one focused on the southwestern region around greater Vancouver, where more than half of the population of the province resides (Figure 3.3). This region was more moderately impacted by smoke from 65  wildfires burning north of the area in early July 2015 (192). The monitoring stations used for evaluation in these case studies were chosen to represent impacts from the smoke events at different extents and timing. Predictions were made for populated grid cells (Figure 3.3) in the case study area during the case study period.  Figure 3.3 Map of British Columbia, Canada.  Grey squares indicate grid cells that cover the populated areas in the province, where predictions from the model are made. Orange crosses indicate large fires with high radiative power (>1000 MW) that occurred during the study period (2010-2015). The box with the dashed line indicates the central interior region for case study 1, while the box with solid line indicates the area for case study 2, which included the greater Vancouver area. The locations of the monitoring stations presented in case studies are also labeled. Base map data source: BC Stats and BC Ministry of Health  66  3.2.3 Sensitivity analysis  3.2.3.1 Inclusion of PM lag 24-hour The OSSEM-24h model included PM2.5 concentrations from the previous day as its most important predictor (42), but this limits its utility in areas without dense monitoring networks. Here we chose to exclude lagged PM2.5 concentrations from the main analyses, but to evaluate in a sensitivity analysis whether we could achieve a better model with these data. The 24-hour average of the PM2.5 concentrations on each day were calculated using the 1-hour data. The values from the closest stations on the previous day were assigned to grid cells as the potentially predictive variable PM Lag 24-hour, to describe the general air quality in the time before the prediction was made. Evaluation statistics were calculated for out-of-bag predictions from model with PM Lag 24-hour and compared with those from the primary model. 3.2.3.2 Inclusion of GOES AOD We originally intended to use the aerosol optical depth (AOD) product from the Geostationary Operational Environmental Satellite (GOES) West as a potentially predictive variable, because it was the only AOD product available at the timescale relevant to OSSEM-1h. These data are available at a 30-minute interval and 4 km X 4 km spatial resolution during the sunlit portion of the day, and have been correlated with ground-level PM2.5 concentrations in previous studies (96, 193-195). However, the variable was omitted because a large proportion (64%) of the data were missing. Given that MODIS AOD was an important contributor to OSSEM-24h, we tested OSSEM-1h models developed with and without GOES AOD using the subset of available data. 67  We applied the cloud filter provided with the product to remove invalid retrievals and assigned the average of the two retrievals for each hour to the prediction grid. If the value at a grid cell was missing or invalid, the value from the closest cell with valid value within 50km was assigned, which was the same approach used for OSSEM-24h (42). 3.3 Results There were 789,337 1-hour observations with complete information for the response variable and all potentially predictive variables, among which 31,688 (4%) had PM2.5 concentrations higher than 15 µg/m3. The PM2.5 observations had a median [25%ile, 75%ile] of 4 [2, 7] µg/m3. After the sampling procedure (Figure 3.2), a reduced dataset of 95,064 observations was used to train the final model. The PM2.5 observations in the reduced data had a median of 6 [3, 17] µg/m3.  The final model included all 15 potentially predictive variables after the backward selection process. The eastward wind component at 50 m above surface (E50m) was the most important variable in the model, followed by Ecozone of the closest fire and other meteorological variables, and FRP Within 500km, which reflected the intensity of fire activity in area (Figure 3.4).  The final model had an overall correlation of 0.93, RMSE of 9.34 µg/m3, MFB of -0.68%, and MFE of 45% between the out-of-bag predictions and observations. Some of the evaluation statistics were influenced by data reduction, where low PM2.5 values were under-represented relative to the entire distribution (Figure 3.2). When we applied the model to make predictions to the complete dataset (789,337 observations), the predictions had an overall correlation of 0.94, RMSE of 3.2 µg/m3, MFB of -15.1%, and MFE of 44.7% compared with observations. The model performed best on days with high fire activity (r = 0.94), compared with moderate fire 68  activity (r = 0.90) and low fire activity (r = 0.77). The model tended to underestimate observed concentrations in all of these categories (Figure 3.5).  Figure 3.4 Variable importance plot for the OSSEM-1h model.  Variable importance is ranked by the normalized percent increase in mean squared errors comparing the predictions from the model with the actual values of the variable and the model with the permutated values of the variable. 69   Figure 3.5 OSSEM-1h model performance stratified by fire activity. These scatterplots compare the out-of-bag predictions from the OSSEM-1h model against observations on days with low, moderate and high fire activity in the province. Red line indicates the 1:1 relationship.   70   Figure 3.6 Spatial distribution of evaluation statistics for OSSEM-1h model.  The (a) correlation (r), (b) root mean squared error (RMSE), (c) mean fractional bias (MFB), and (d) mean fractional error (MFE) between model predictions and observations in the leave-region-out cross-validation. The dashed box indicates the central interior area in case study 1, and the solid box with inset indicate the area in case study 2, which includes greater Vancouver. Base map data source: BC Stats and BC Ministry of Health  The leave-region-out cross-validation resulted in an overall correlation of 0.60 between predictions and observations. Station-specific correlations ranged from -0.09 to 0.86, with an 71  interquartile range from 0.48 to 0.70. Some spatial patterns were observed in the distribution of the performance statistics across different stations (Figure 3.6). The model performed well around the greater Vancouver region with high correlations, low RMSE and MFE, and MFB close to zero. Stations in the central interior region also had high correlations, but they tended have large RMSE, negative MFB, and high MFE values (Figure 3.6). This may have been due to the higher PM2.5 concentrations and thus larger variability in the central interior region compared with the greater Vancouver region. Stations with fewer observations and lower variability tended to have lower correlation: all stations with correlation below 0.3 had < 500 observations and a PM2.5 interquartile range (IQR) of < 6 µg/m3, compared with the median of 1326 observations and 12 µg/m3 IQR among all stations.  3.3.1 Case study 1 Several large wildfires began burning near Williams Lake (Figure 3.3) in the central interior region on 28 July 2010. These fires remained partially contained until 17 August 2010, when a wind event caused rapid fire growth and wide smoke dispersion across the region. Three of the PM2.5 monitoring stations in the area had distinctly different time series of observed 1-hour concentrations between 17 and 19 August 2010, which were generally well-predicted by OSSEM-1h (Figure 3.7 a-c). However, the model tended to overestimate when PM2.5 concentrations were low and to underestimate when PM2.5 concentrations were high. The spatial distribution of OSSEM-1h predictions reflected the rapid dispersion of smoke from Williams Lake to the surrounding area farther and farther removed from the fires (Figure 3.7 e-f).     72  3.3.2 Case study 2 Starting 5 July 2015, the greater Vancouver area was affected by smoke from multiple fires at the north of the case study boundary (Figure 3.3). The 24-hour averaged PM2.5 concentrations across greater Vancouver were elevated over the provincial air quality objective (25 µg/m3) for multiple days, which is rare for the region. Stations in both Squamish and Chilliwack observed significant elevations in PM2.5 concentrations, but with the peaks occurring at different times, while the Colwood station observed limited impacts (Figure 3.8). Similar to Case Study 1, OSSEM-1h predictions generally agreed well with observed 1-hour PM2.5 concentrations but consistently underestimated the high concentrations (Figure 3.8).  3.3.3 Sensitivity analyses The model including the PM Lag 24-hour variable had a correlation of 0.93, RMSE of 9.41 µg/m3, MFB of -0.23%, and MFE of 44% between the out-of-bag predictions and observations. The correlations on days with high, moderate and low fire activity were 0.94, 0.98 and 0.76, respectively. This performance was very similar to the performance of the primary model without PM Lag 24-hour. There were 245,464 1-hour observations available across 66 monitoring stations with complete data for GOES AOD and all other predictive variables. Using the same approach described previously (Figure 3.2), 34,758 observations were sampled to train the models with and without AOD. The distribution of the PM2.5 observations in this subset of data was similar to that of the full reduced dataset, with a median of 7 [3, 17] µg/m3. The out-of-bag predictions made with the 73  model including AOD had a correlation of 0.90, RMSE of 10.7 µg/m3, MFB of 0.12% and MFE of 46% when compared with observations, while the model excluding AOD had a correlation of 0.91, RMSE of 9.7 µg/m3, MFB of 0.01% and MFE of 45%.   74   Figure 3.7 Illustration of OSSEM-1h performance in Case study 1.  On the left, the time series of 1-hour PM2.5 model predictions (grey bars) are compared with observations (black line) at monitoring stations in (a) Williams Lake, (b) Prince George, and (c) Burns Lake from 17 – 19 August 2010. The maps on the right show model predictions at (d) 17:00 on 17 August, (e) 10:00 on 18 August, and (f) 17:00 on 18 August in all populated grid cells. The time points correlating to the right-hand panels are labeled on each of the time series plots (color coded), and the locations correlating to the left-hand panels are labeled on each of the maps. Base map data: Google. 75   Figure 3.8 Illustration of OSSEM-1h performance in Case study 2.  On the left, the time series of PM2.5 model predictions (grey bars) are compared with observations (black line) at monitoring stations at (a) Chilliwack, (b) Colwood, and (c) Squamish from 05 – 07 July 2015. Maps on the right show model predictions at (d) 00:00 on 06 July 6, (e) 10:00 on 06 July, and (f) 05:00 on 07 July in all populated grid cells. The time points correlating to the right-hand panels are labeled on each of the time series plots (color coded), and the locations correlating to the left-hand panels are labeled on each of the maps. Base map data: Google. 76  3.4 Discussion A statistical model was developed to estimate 1-hour PM2.5 concentrations at 5 km2 resolution during wildfire seasons over the entire province of BC, Canada using a Random Forests model and multiple data sources. The two case studies also showed that the model identified the temporal and spatial peaks in smoke exposures. Predictions from the final model had a high correlation with observations, while spatial cross-validation revealed some variability in prediction performance. Overall, the model performed best closer to the coast, especially around the densely populated greater Vancouver area, where concentrations were generally lower. Although correlations remained high in the interior region, the error and bias were increased further from the coast. To the best of our knowledge, this is the first statistical model developed for smoke-related PM2.5 at a temporal resolution of 1-hour. All other models that produce 1-hour estimates of smoke-related PM2.5 use a deterministic approach. One example is the BlueSky framework, which is an operational smoke forecasting system run in both the US and Canada. An evaluation of BlueSky performance at the 1-hour resolution in California reported MFB values of −84.4% and −1.3% and MFE values of 100.8% and 91.8% for two case studies (95). Another example is the FireWork forecasting system developed by Environment Canada. An evaluation of its performance during the 2015 fire season found a correlation of 0.49 and RMSE of 18 µg/m3 compared with observations in western Canada (77). Although the evaluation statistics for BlueSky and FireWork were less favorable than OSSEM-1hr, these modeling systems were forecasting hours ahead of time while the OSSEM-1hr model was retrospectively estimating concentrations using already-observed data.  77  Deterministic and statistical models have been developed to estimate 1-hour PM2.5 concentrations due to sources other than wildfire smoke. One study examined the PM2.5 estimates from ten deterministic air quality modeling systems in regions of North America and Europe in 2006 and found that their correlations with observations were between 0.53 and 0.66 (196). On the other hand, a statistical regression model with remote sensing data built for southern Ontario, Canada resulted in a correlation of 0.81 (197) in 2004, while a neural network coupled with K-mean clustering approach developed in Auckland, New Zealand, had a correlation of 0.79 (198) between 2008 and 2011. The out-of-bag evaluation suggests that OSSEM-1hr performed well compared with these models, possibly due to the large training dataset, the choice of modeling approach, the selection of relevant predictive variables, and the difference in distribution of observations included in the analysis. The leave-region-out cross-validation analysis showed that model performance can vary when applied to areas more than 50 km from the locations of the training data. However, the model can still provide reasonable estimates for most of the exposed population in BC, because 80% of people reside within 10 km of a monitoring station (42).  The inclusion of PM2.5 measurements from the previous day as a potentially predictive variable did not improve the model performance in the out-of-bag evaluation, suggesting the potential of the model to be applied in regions without large air quality monitoring networks. When the data were subset to observations for which there were complete AOD data, the inclusion of AOD did not improve the model performance. This was surprising, given the importance of AOD in previously-developed 24-hour models (96, 193, 195). However, there are several factors to consider. First, the quality of the GOES AOD data used here may be affected by the location of the study area, where the larger solar zenith angle at high latitude can lead to larger bias in AOD 78  retrieval (199). In addition, BC is at edge of the GOES-West image swath, and the larger satellite viewing angle results in a larger scattering angle, which relates to larger bias compared with data retrieved at the centre of the image (200). Second, the standard cloud filter from the product may not be appropriate for the study period because heavy smoke may have been incorrectly classified as cloud (96). Previous studies have shown that a less restrictive cloud filter derived from local estimates of surface brightness could improve the utility of AOD products from other satellites during smoke events (201, 202). Third, the mountainous terrain in the province may result in more smoke aloft, leading to lower correlation between satellite AOD and surface concentrations. Fourth, the variables used to represent fire activity and meteorological conditions may adequately explain the response variable in the absence of AOD, leaving little room for improvement.  The OSSEM-1h model has some unique strengths. First, all data sources are publicly available and easy to acquire. Second, most of the data have global coverage such that these methods could be locally adapted and expanded to other parts of the world, and could be especially useful in fire-affected areas without sufficient PM2.5 data from existing air quality monitoring networks. Third, the simplicity of the model makes it relatively straightforward to operationalize for near-real-time surveillance. Finally, OSSEM-1hr could also serve as a tool for more spatially comprehensive evaluation of smoke forecasting models. The OSSEM-1h model also has some important limitations. First, errors and uncertainty in the measurement of the predictive variables, such as those from satellite data retrieval or meteorological modeling, will be inherited by the model, affecting the model performance. 79  Second, the fire activity information is currently retrieved from MODIS, which is not updated hourly. As a result, the fire information fed into the model may not reflect the situation at the exact hour of prediction, especially when fire behavior is changing rapidly or there is active and effective effort to suppress fires. We did explore the possibility of using the GOES fire product, which is updated every 30 minutes, but the fire detection sensitivity is affected by the low spatial resolution (203) and FRP measurements are not available in the archived data. However, the launch of the GOES-R series of geostationary satellites since 2016 is expected to improve the capacity of fire detection, potentially useful for developing similar models in the future. The imaging instrument aboard of these satellites can provide three times more spectral information, four times the spatial resolution, and more than five times faster temporal coverage than the previous system (204). Third, this approach cannot fully account for smoke generated well beyond the BC study area. Although smoke can travel to BC from other regions, especially the western United States and Siberia, a recent study found that most smoke affecting the province originated from fires within the province (182). Fourth, while the machine learning approach can make good estimates of the response variable, the precise nature of the relationships between the response and predictive variables are impossible to ascertain. As such, users must simply trust the output in context of its performance, in the absence of specific information about the model structure. In addition, the model output may not be interpreted with the physical process of smoke dispersion. Finally, the model selection process based on variable importance may not be adequate to select the optimal model, as the permutation-based variable importance calculation may be influenced by correlated variables in the model (205).  80  The OSSEM-1h model presented here can be used to temporally resolved estimates of population exposure to PM2.5 during episodes of wildfire smoke. These can be useful for epidemiologic studies on the very acute health effects of sub-daily exposure, given health outcome measures with sub-daily time stamps such as ambulance dispatch data (26, 147). It can also be used as a tool for public health surveillance of the exposure in near-real-time, which can be used inform timely actions to mitigate the adverse population health impacts of wildfire smoke exposure.      81  Chapter 4: Association between sub-daily exposure to fine particulate matter and ambulance dispatches during wildfire seasons 4.1 Introduction The first challenge of assessing population exposure at high temporal and spatial resolution was addressed in Chapter 3. The second challenge is to examine the health effects of these exposures using British Columbia (BC) ambulance dispatch data, which uniquely provide the exact time and location of emergency health events. Linking the ambulance dispatch data with subsequent paramedic reports and hospital admissions allows more complete examination of the relationship between wildfire smoke and all ambulances dispatches, as well as those subsets most likely to be due to cardiovascular, respiratory, and diabetic conditions.  4.2 Methods 4.2.1 Health outcome data We obtained data for all emergency ambulance dispatches during the study period from BC Emergency Health Services, which is the sole provider of ambulance and emergency health services across the province. The data included the date and time of the call, geographic coordinates of the event, and the reason for the call recorded as one of 33 codes (Appendix B.1) assigned by the dispatcher using the Medical Priority Dispatch System (MPDS). The MPDS is a standardized set of protocols produced by the International Academies of Emergency Dispatch (IAED). Calls without a dispatch location or calls from callers who made more than four calls 82  during the study period (5% of all unique callers) were excluded (Figure 4.1). The latter was done to minimize the occurrence of multiple calls within a short period of time, which may violate the assumption in a case-crossover study. If there were multiple calls from the same caller within a 24-hour period, only the first call was included in the analysis. Each call in the dispatch database was provided with a linked patient care report as completed by the attending paramedics. Key information retrieved from these reports included the Personal Health Number (PHN, a lifetime unique identifier for health care in the province), the age and sex of the patient, and any assessment of medical conditions by the paramedics, assigned as one of the 180 Paramedic Impression (PI) codes (Appendix B.2). Although each patient had a care report, not all patients had an impression code as it is not a mandatory field for the paramedics to fill in. We also obtained hospital discharge data from the BC Ministry of Health (206), which included the date of hospital admission and the primary diagnosis, coded according to the International Classification of Diseases, 10th Revision (ICD-10). The primary diagnosis reflects the primary reason for the total length of the hospital stay, and so may or may not reflect the initial reason for admission. Hospital diagnoses were included if they linked to a dispatch call by PHN if (1) the admission occurred within a 7-day period of the ambulance dispatch; and (2) it was the admission closest to the date of the dispatch for cases where multiple admissions were found within the 7-day period.  83   Figure 4.1 Flowchart of analytic data selection.  Given this chain of data linkage, we could have up to three measures of health outcome for each dispatch call: (1) the MPDS code assigned by the dispatcher at the call centre (Ambulance Dispatch); (2) the PI code assigned by the paramedics at the dispatched location (Paramedic Assessment); and (3) the primary ICD-10 code associated with the hospital admission record (Hospital Diagnosis). These three measures have different advantages and disadvantages for the purposes of our study. The Ambulance Dispatch code was assigned to every single dispatch (no missing data) and it was assessed at the time closest to the onset of the emergency event. However, it was generally based on information self-reported by a lay caller and recorded as broad categories of health problems. The Paramedic Assessment code was assigned by professionals with medical training after a physical examination the patient, but the assessment 84  can be constrained by time, equipment, and demand, and as noted is not available on every record. The Hospital Diagnosis code provides the most robust medical assessment among the three, but it was only available for the most severe cases (i.e. that were admitted), and it could be made hours or even days after the initial call. Considering these different features, we decided to first provide a summary of the relationship between these three measures to assess consistency, and then to examine the dispatches related to cardiovascular, respiratory and diabetic conditions, as identified by each of these three measures (Table 4.1).  4.2.2 Exposure assessment and assignment Hourly exposures to fine particulate matter (PM2.5) during the study period for all subjects were estimated with the 1-hour Optimized Statistical Smoke Exposure Model (OSSEM-1h) developed in Chapter 3 (207). Exposure for each dispatch call was assigned based on the date and hour of the call, as well as the dispatch location (latitude and longitude) which was matched to the exposure model grid.       85  Table 4.1 Definitions and number of cases for each health outcome measure. Case groups Definition Number of Cases All  676,401  Cause-specific cases identified by Ambulance Dispatch codes  based on the Medical Priority Dispatch System (MPDS) Breathing Problems MPDS = 6  46,277 Chest Pain MPDS = 10  51,996 Arrest MPDS = 9  3,527 Stroke MPDS = 28  21,173 Heart Problems MPDS = 19  12,039 Diabetic Problems MPDS = 13  5,987  Cause-specific cases identified by Paramedic Assessment codes  based on Paramedic Impressions (PI) Circulatory PI starts with 08 or 1830 44,122 Respiratory PI starts with 09 23,392 Stroke PI = 0615 17,495 Asthma/COPD PI = 0930 5,824 Myocardial Infarction   PI = 0860 1,724 Diabetic PI = 0305 or 0315 4,722  Cause-specific cases identified by Hospital Diagnosis codes based on  International Classification of Diseases, 10th Revision (ICD-10) Circulatory ICD-10 = I00 to I99 37,078 Respiratory ICD-10 = J00 to J99 22,038 Ischemic Heart Diseases  ICD-10 = I20 to I25 10,653 Stroke ICD-10 = I60 to I69, G45, H341 10,373 Asthma/COPD ICD-10 = J40 to J45 9,084 Lower Respiratory Infection  ICD-10 = J13 to J22 7,708 Diabetic ICD-10 = E10 to E14 2,921 86  4.2.3 Statistical analysis A time-stratified case-crossover study design (208) was used to assess the association between ambulance dispatches and estimated PM2.5 exposure during wildfire seasons. Exposure during the case window was compared with exposures during a series of control windows. The case window was defined as the hour immediately before the ambulance was dispatched, while the control windows were defined as the same hour on the same day-of-week in the same calendar month of the ambulance dispatch, to control for day-of-week effects and seasonal trends. Control window exposures were assigned at the same location as the case window exposure. Using conditional logistic regression, individual factors that do not vary over a short time period (i.e. age, smoking status) can be controlled, because the exposures during the case and control windows are compared within the same individual.  To examine the lag structure of the association between exposure and outcome, a distributed lag non-linear model (DLNM) (209) was used. This type of model can simultaneously describe complex exposure-response and lag-response relationships by combining the functions for both relationships in the same model. This approach has been applied in studies on the acute health effects of air pollution and ambient temperature (210-213). We allowed for delayed effects up to 48-hours (lag 1-48h) because most of previous studies using 24-hour average exposures to wildfire smoke found the strongest association or best model fit at lags of 0 to 2 days (31, 54, 58, 62, 91, 214). A natural cubic B-spline with two or three degrees of freedom, depending on the health outcome, was used for the lag-response relationship based on exploratory analyses to minimize the Akaike Information Criterion (AIC). Other functions including polynomial 87  functions and penalized splines, with varying degrees of freedom, were also tested to describe the lag structure in the exploratory analyses, which produced similar results and less desirable model fit compared with cubic splines based on AIC. Both the lag-specific and cumulative odds ratios were calculated to evaluate the time course of the effect and the overall association, respectively. A linear exposure-response relationship was assumed in the analyses, after preliminary evaluation of linear and non-linear options found the linear models fit the data best for most health outcomes. This assumption could also simplify the presentation of the results, and allowed us to focus on the lag-response relationship.  A sensitivity analysis was conducted with the subset of ambulance dispatches that were (1) not linked hospital diagnosis, (2) linked to any hospital diagnosis, and (3) linked to hospital diagnosis that categorically matched with the ambulance dispatch codes (Table 4.2). This analysis was aimed to examine whether the lag-response relationship is different for cases of different severity.  Table 4.2 Matched Hospital Diagnosis for Ambulance Dispatch codes. Ambulance Dispatch Codes Matched Hospital Diagnosis Breathing Problems ICD-10 = J00 to J99 Chest Pain ICD-10 = I00 to I99 Arrest ICD-10 = I00 to I99 Stroke ICD-10 = I60 to I69, G45, H341 Heart Problems ICD-10 = I00 to I99 Diabetic Problems ICD-10 = E10 to E14 .  88  All models were adjusted for the same-day and previous-day maximum apparent temperatures, obtained from Environment and Climate Change Canada, using a natural cubic B-spline with three degrees of freedom. All data preparation and statistical analyses were conducted using R software (version 3.5.1). The dlnm package was used to fit DLNM (215). Cox regression with Breslow ties was used to fit conditional logistic regression models, adopting the example code provided in a previous publication (213). The study was approved by the Behavioural Research Ethics Board at the University of British Columbia (H15-02269). 4.3 Results A total of 676,401 dispatch calls from 500,302 unique individuals were included in the study, among which 444,189 (65.7%) calls had a Paramedic Assessment code and 244,101 (36.1%) calls were linked to Hospital Diagnosis codes (Figure 4.1). The mean (interquartile range) of PM2.5 exposures during the case and control windows in the 1-hour window prior to dispatch were 5.5 (3.1, 6.5) µg/m3 and 5.4 (3.0, 6.4) µg/m3, respectively. The mean of the maximum apparent temperature on the case days and control days were 17.6 (13.5, 21.7) °C and 17.6 (13.4, 21.7) °C, respectively. Paramedics arrived at the dispatched location within one hour of the call in 99% of the cases, regardless of Paramedic Assessment code group. Hospital admissions occurred within the same calendar day of the dispatch calls for 73% to 81% of cases, depending on the Hospital Diagnosis code group.  89  For each of the six Ambulance Dispatch codes groups (Table 4.1), 66% to 73% of the calls also had a Paramedic Assessment code. For calls with the Ambulance Dispatch code Breathing Problems, the most prevalent (45.8%) Paramedic Assessment code was Asthma/COPD or other Respiratory conditions. Most of the calls with an Ambulance Dispatch code for Chest Pain, Heart Problems, and Arrest (61.0%, 59.2% and 60.0%) had Paramedic Assessment codes for Myocardial Infarction or other Circulatory conditions. On the other hand, only 44.2% of calls with the Ambulance Dispatch code for Stroke had a Paramedic Assessment code for Stroke, and only 51.2% of calls with an Ambulance Dispatch code for Diabetic Problems had the Paramedic Assessment code for Diabetic (Figure 4.2A). For each of the six Ambulance Dispatch codes groups assessed (Table 4.1), 24% to 52% were linked to Hospital Diagnosis codes. Compared with the linkage between Ambulance Dispatch and Paramedic Assessment codes, the linkage between Ambulance Dispatch and Hospital Diagnosis codes had a larger proportion of cases in the Other Codes category (Figure 4.2B), indicating a somewhat weaker correspondence. The strongest relationship was observed between Paramedic Assessment codes and Hospital Diagnosis codes (Figure 4.2C). For example, there were 14129 cases with a Paramedic Assessment code for Myocardial Infarction and a subsequent hospital admission, of which almost 80% also had a Hospital Diagnosis code for Ischemic Heart Disease, of which Myocardial Infarction was a primary subtype. 90   Figure 4.2 Corresponding relationship between health outcome measures.  These figures show the distribution of (a) Paramedic Assessment code for each Ambulance Dispatch code; (b) Hospital Diagnosis code for each Ambulance Dispatch code; (c) Hospital Diagnosis code for each Paramedic Assessment code. Numbers at the top row indicate the number of calls included in the analysis and percentages at the second row indicate the percentage of calls without missing data. 91  The lag-response relationship varied by health outcome. There was a small increase in the odds of any ambulance dispatch immediately following increased PM2.5 exposure (Figure 4.3). A greater increase in the immediate effect was observed for respiratory conditions identified by the Ambulance Dispatch code Breathing Problems and the Paramedic Assessment codes Asthma/COPD and Respiratory. In all cases the increase was largest at the 1-hour lag interval. However, the odds ratios for respiratory outcomes identified by the Hospital Diagnosis codes (Respiratory, Asthma/COPD, and Lower Respiratory Infection) increased over time (Figure 4.4). On the other hand, the Ambulance Dispatch codes for Chest Pain and Heart Problems did not show any increase associated with PM2.5 exposure, but there was an immediate effect in the Paramedic Assessment code for Myocardial Infarction and in the Hospital Diagnosis codes for Ischemic Heart Disease. In addition, the Ambulance Dispatch code for Arrest suggests increased risk over time (Figure 4.5). The lag-response relationship for Stroke varied by the outcome measures: an immediate effect was observed in Ambulance Dispatch code, while the odds ratios increased over time and became positive after 24-hours of exposure for the Paramedic Assessment and Hospital Diagnosis codes (Figure 4.5). Finally, the odds of outcomes for the Ambulance Dispatch and Paramedic Assessment codes for Diabetic conditions both began to increase at approximately 24-hour after the exposure, but the same was not observed for the Hospital Diagnosis codes (Figure 4.6). Many outcomes reached their maximum cumulative odds ratios [95% confidence interval] at a 48-hour lag, including: 1.04 [1.01, 1.07] for the Ambulance Dispatch code Breathing Problems; 1.05 [1.00, 1.10] for the Hospital Diagnosis code Respiratory and 1.10 [1.01, 1.19] for Lower Respiratory Infection; 1.07 [1.00, 1.15] for the Ambulance Dispatch code Diabetic Problems; 92  and 1.10 [1.01, 1.20] for the Paramedic Assessment code Diabetic. On the other hand, the Hospital Diagnosis code Ischemic Heart Disease reached a cumulative maximum of 1.07 [0.98,1.16] at a 24-hour lag, and the Paramedic Assessment codes for Myocardial Infarction and Asthma/COPD both reached maximums of 1.19 [0.98, 1.44] and 1.10 [1.01, 1.20], respective, at a 12-hour lag (Figure 4.4 – 4.6).  Results of the sensitivity analysis showed some interesting differences in lag-response relationship for cases of different severity (Figure 4.7). For the Ambulance Dispatch code Breathing Problems, the effect in cases without any hospital diagnosis was the largest in the first hour after exposure and declined over time. However, the opposite trend was observed in cases with matched Hospital Diagnosis. For Stroke, the elevated effect in the first few hours were not observed in cases that had a matched Hospital Diagnosis, compared with those who did not. Plausible explanations for these differences include: (1) cases that were more severe and thus required hospital admission might take a longer time between exposure and the onset of the conditions; and (2) cases in which the patients waited longer to call ambulance for medical attention may lead to more severe outcomes. All results for calls with any linked Hospital Diagnosis (the middle column in Figure 4.7) were null, possibly due to noise resulting from the mix of different and non-specific reasons for the calls.    93   Figure 4.3 Results for PM2.5 exposure and all ambulance dispatch calls during wildfire season The plots show the lag-specific (red line) and cumulative (blue line) odds ratios for a 10 µg/m3 increase in PM2.5 for all ambulance dispatch calls during wildfire season.   Figure 4.4 Results for PM2.5 exposure and respiratory health outcomes during wildfire season These plots show the lag-specific (red line) and cumulative (blue line) odds ratios for a 10 µg/m3 increase in PM2.5 for respiratory health outcomes during wildfire season. Panels are color coded by outcome measures. 94   Figure 4.5 Results for PM2.5 exposure and circulatory health outcomes during wildfire season These plots show lag-specific (red line) and cumulative (blue line) odds ratios for a 10 µg/m3 increase in PM2.5 for circulatory health outcomes during wildfire season. Panels are color coded by outcome measures. 95   Figure 4.6 Results for PM2.5 exposure and diabetic outcomes during wildfire season These plots show lag-specific (red line) and cumulative (blue line) odds ratios for a 10 µg/m3 increase in PM2.5 for diabetic outcomes during wildfire season. Panels are color coded by outcome measures.  96   Figure 4.7 Results of lag-response relationship in sensitivity analysis with cases of different severity.  This figure shows the lag-response relationship for ambulance dispatches that were not linked to hospital diagnosis (red), linked to any hospital diagnosis (green), and linked to hospital diagnosis that matched the dispatch codes (blue).   97  4.4 Discussion In this study, we found that: (1) cause-specific Ambulance Dispatch codes matched to subsequent Paramedic Assessment and Hospital Diagnosis codes agreed reasonably well, providing more confidence in ambulance dispatches as a measure of health; (2) exposure to elevated PM2.5 during wildfire seasons was associated with increased odds of ambulance dispatches related to respiratory and cardiovascular conditions, and the strongest effects were observed in the hour immediately after the exposure; and (3) exposure to elevated PM2.5 during wildfire seasons was also associated with ambulance dispatches related to diabetic conditions, with significant effects observed after a 24-hour lag in exposure.    The associations between respiratory outcomes and PM2.5 estimates were consistent with previous reports. Studies using ambulance data in Australia found an association between daily PM2.5 and breathing problems (RR=1.04, 95% CI 1.02 to 1.05) (147), as well as asthma/COPD calls (OR = 1.06, 95% CI 1.01 to 1.11) (114), similar to the 24-hour and 48-hour cumulative effects we report. These cumulative effects were also consistent with effects estimated for respiratory medication dispensations (29, 54),  physician visits (29), and hospital admissions (31) from studies using daily PM2.5 measurements during wildfire events in the same region. Increased airway inflammation and decreased lung function have been observed in children with asthma and the elderly immediately following exposure to ambient PM2.5, with lagged effects lasting from 5 to 12 hours (121-124). We also found an immediate increase in ambulance dispatches for reparatory codes following exposure, with lagged effects lasting from 12 to 24 hours. 98  Odds of myocardial infarction as measured by paramedic assessment were elevated immediately following exposure, as were odds of ischemic heart disease as measured by hospital diagnosis. Neither result was reported by the previous Australian study (114). However, a few studies using a similar case-crossover design found immediate effects of similar magnitude for myocardial infarction following exposure to elevated ambient PM2.5 (126, 129, 141, 142). Similar effects on myocardial infarction have also been observed in epidemiologic studies with different designs as well as  experimental studies (119). While previous studies have reported significant associations between out-of-hospital cardiac arrests and wildfire smoke exposure (26, 147), we found elevated effects with wide confidence intervals. Results from studies on ambient PM2.5 and out-of-hospital cardiac arrests have also been inconsistent, where some found immediate effects following exposure (129, 144, 145) and others found no association (137, 138, 140). These inconsistencies could be due to different risk factors for cardiac arrest in different regions. For example, drug overdose was one of the major paramedic assessments for cardiac/respiratory arrest calls in this study, but not in the Australian study (147). Although previous studies have found association between long-term exposure to PM2.5 and diabetes incidence and prevalence (216, 217), few have looked at the association between short-term exposure and acute diabetic clinical syndromes. Here we found that the association between diabetic problems and smoke exposure increased over time and became statistically significant at a 24-hour lag. Another recent study in Australia found daily PM2.5 exposure was associated with increased same-day and next-day ambulance dispatches for hypoglycemia (114). When we 99  modelled paramedic assessments for hypoglycemia and hyperglycemia separately, we also found hypoglycemia to be the driver of the association for diabetic conditions (Figure 4.8).  This study has some unique strengths. The temporal resolution of the exposure and ambulance dispatch data allowed the examination of the exposure-response lag structure on an hourly scale. The linkage between ambulance dispatches, paramedic assessments, and hospital admissions provided the opportunity to evaluate the quality of the data and the internal consistency of the study results. In addition, having a single provider of ambulance services in BC enabled us to conduct a population-based study over a very large geographic area.  Figure 4.8 Results for subtypes of diabetic conditions.  The figure shows odds ratio for 10µg/m3 increase in PM2.5 by lag for paramedic assessment code of hypoglycemia (left) and hyperglycemia (right).  There were also several limitations. First, exposure misclassification was possible due to (1) error in the exposure model, (2) uncaptured variability within the 5 km2 prediction grid, and (3) the assumption that each subject was exposed at the dispatch location during the control windows. As described in Chapter 3 (Figures 3.7 and 3.8), the exposure model tended to 100  underestimate PM2.5 concentrations when exposures were high, which could lead to overestimation of the effect estimates. In addition, the exposure model may misclassify exposure in time, which could result in errors in the lag structure estimates. Based on the case studies in Chapter 3 (Figures 3.7 and 3.8), the increase in exposure was sometimes predicted ahead of time compared with observations. If such error was non-differential, it could lead to systematic overestimation of the exposures around the time of the health outcomes, which would produce underestimations of the effect estimates in the first hours. It could also affect the shape of the overall 48-hour lag structure, given that concentrations in the early part of the window may have been overestimated.  Second, some populations may be more likely to call ambulance services than others (218, 219). With the case-crossover study design, we were able to control for the confounding effect of this factor, but the results might not be generalizable to the general population. Third, there is still uncertainty about the temporal relationship between PM2.5 exposure, symptom onset, ambulance dispatch, and subsequent care for any given study subject. Although the ambulance dispatch data have finer temporal resolution than many other administrative datasets, there may still be a lag between the onset of symptoms and the action of calling an ambulance. Further, there was also a lag between some ambulance dispatches and hospital admissions, the reasons for which are unclear. In some cases, the subject may have been transported to hospital and held under observation in the emergency room prior to admission. Unfortunately, the emergency room data are disparately collected for each hospital and not available through an integrated database. All of these uncertainties may impact the lag structure of the effect estimates. 101  This study adds to the limited evidence on the acute health effects from sub-daily exposure to PM2.5, especially during wildfire seasons. We found effects on respiratory and cardiovascular codes immediately following exposure, while effects on diabetic codes were associated with cumulative exposure over a 24-hour period. These results warrant further investigation into the health effects of sub-daily exposures and may have implication for the appropriate time scale of air quality standards and public health actions during air pollution events.    102  Chapter 5: Conclusion 5.1 Research summary and contributions This dissertation developed and applied models of sub-daily exposure to PM2.5 during wildfire seasons in British Columbia (BC) to evaluate acute impacts on population health. A novel machine learning approach was used to identify data relevant to the vertical profile of wildfire smoke in the atmosphere, addressing one of the most important issues for improving the application of remote sensing data for population smoke exposure assessment. With relevant data identified, the hourly Optimized Statistical Smoke Exposure Model (OSSEM-1h) was developed to estimate 1-hour exposure to fine particulate matter (PM2.5) for all populated areas in BC during wildfire seasons from 2010 through 2015, including three severe seasons in 2010, 2014, and 2105. This model can be applied in further epidemiologic studies and provincial public health surveillance. It can also be expanded or transferred to other regions impacted by wildfire smoke, as has been done for OSSEM-24h and transformed into a real-time forecasting model across the United States (220). Finally, a unique administrative health dataset was established with data collected and linked from ambulance dispatches, paramedic assessments, and hospital admissions for the entire province. With modeled exposure assigned to subjects in this dataset, the complex lag-response relationship between exposure and respiratory, circulatory, and diabetic conditions was evaluated using the modern statistical approach of distributed lag non-linear modelling.  103  5.1.1 Additions to the health literature Despite the vast body of literature on the health effects of air pollution that has developed over the past decades, epidemiologic evidence is still limited on the effect of sub-daily exposure, especially for air pollution from non-urban sources. The specific lag-response relationships for different health outcomes are largely unknown. Such evidence can provide insight on the mechanism behind the relationship and is important for the design of public health policies. Results from Chapter 4 add to this evidence, and suggest that exposure to PM2.5 during wildfire seasons can lead to respiratory and cardiovascular health responses within one hour, while the specific lag-response relationships differ for different health responses. It also provides evidence that exposure to PM2.5 can exacerbate diabetic symptoms within 24 hours of the exposure, a relationship that has rarely been studied before. 5.1.2 Advancement in exposure assessment  Accurate and precise assessment of exposure is a key component of environmental epidemiology. The lack of adequate tools for exposure assessment is often the primary obstacle for producing evidence on a specific research question. Conventionally, population-based epidemiologic studies on air pollution have largely relied on exposure data from air quality monitors, which are sparsely distributed outside urban areas. For wildfire smoke pollution specifically, the restricted spatial coverage of air quality monitors may also contribute to inconsistencies in the findings for associations with mortality and cardiovascular outcomes. Exposure assessment tools with high spatial and temporal resolution, as well as sufficient spatial coverage, are essential for addressing these issues. 104  Remote sensing data have been used for air pollution studies for many years. Due to broad spatial coverage, remote sensing data are often used as a supplement to air quality monitors for population-based research (19, 221). One major challenge for applying remote sensing data in population health studies is the uncertainty in the relationship between pollutants present in the total column of the atmosphere and those present at the surface level, where populations are exposed. This issue is even more important for wildfire smoke, which can be injected high into the atmosphere, depending on the intensity of the fire and meteorology, whereas traffic and industrial emissions are injected at a fixed height. Some studies have attempted to address this issue by using ground-based or air-borne lidar to provide information about the vertical profile of wildfire smoke events (11, 55, 56, 222, 223), but they could only focus on specific events restricted to certain locations, given the limited spatial coverage of these instruments. The study described in Chapter 2 illustrated a novel approach to address this limitation, generating estimates of vertical distribution of wildfire smoke for a large geographic area and with high temporal resolution. These estimates can be used to improve the use of remote sensing products for public health research and surveillance, and the approach can also be expanded to other regions. They may also be useful for improving the plume rise and injection height components of deterministic models and forecasting systems, such as FireWork and BlueSky, as they rely readily accessible data. Modeling is a common way to obtain estimates of air pollution exposure with high spatial and temporal resolution for population-based epidemiologic studies. Both deterministic and empirical models have been used to estimate daily averaged exposure to wildfire smoke in previous studies (29-31, 89, 96, 114), but none has yet refined these models to the hourly scale. Chapter 3 of this 105  dissertation developed a statistical model using data on fire activity, meteorology, geographic information, and air pollution measurements. It produced a complete set of multi-year hourly exposure estimates for the entire population of BC, which enabled the analyses in Chapter 4 and will also be useful for future studies in the region. With adequate optimization and automation, the model could also be adapted as a surveillance tool for public health decision making, which would complement information from forecasting models and would be especially useful in areas without any air quality monitors.  5.1.3 Application of machine learning for public health Chapters 2 and 3 demonstrated the potential of machine learning for exposure modeling in environmental epidemiologic research. Conventionally, air pollution modeling has relied heavily on deterministic models that simulate the physical and chemical process of pollutant emissions, dispersion, and transportation. The performance of these models depends on our complete understanding of the underlying processes, knowledge of which takes a long time and ample resources to develop and accumulate. Rapid developments in the field of data science have provided exposure scientists and epidemiologists with the option to go beyond conventional regression approaches and into modern machine learning techniques for developing empirical models.  The models developed in Chapters 2 and 3, as well as those developed in other regions (96, 104, 114, 198), demonstrate that machine learning techniques can provide exposure estimates comparable with observations. These models can play an important role in exposure assessment for public health research and surveillance, supplementing deterministic models while they 106  continue to improve and evolve. In addition, studies have found that blending data from different models and measurements tends to produce better predictions than any single metric in isolation (224-226), and thus estimates from models such as OSSEM-1h can be combined with outputs from deterministic models to produce more robust exposure assessment. For regions impacted by wildfire smoke but lacking the resources to develop and maintain an expensive deterministic air quality model, OSSEM-1h can be a template for an affordable and effective approach to provide much-needed information for research and public health interventions. 5.1.4 Using symptomatic outcomes for epidemiologic studies Administrative health databases have been an invaluable source of data for population-based epidemiologic studies. Health outcomes with diagnosis codes (e.g. ICD-9 or ICD-10), including outpatient physician visits, emergency room visits, hospital admissions, and vital statistics, have been commonly used for wildfire smoke studies (19, 20, 214). Although imperfect, they are generally considered to be reliable because the coding procedures are standardized and based on physical examination, although threats to the accuracy of diagnoses still exist (227, 228). Even so, the power achieved through use of such large and oftentimes population-based datasets generally outweighs that lost to inaccuracies (229). In Chapter 4, we examined a range of symptomatic outcomes from the ambulance dispatch database maintained by BC Emergency Health Services. The novelty of our approach was the inclusion of cause-specific outcome measures for dispatch codes, the validity of which were evaluated using subsequent paramedic assessment codes and hospital diagnoses codes. This validation process addressed the concern of data quality for these symptomatic outcomes. Results in Chapter 4 indicated that these different 107  outcome measures agreed reasonably well with each other, but the lag-response relationship presented for the symptomatic and diagnostic measures was not the same, highlighting the value of including symptomatic outcomes in future studies, which may lead to new discoveries missed by conventional diagnostic outcomes, such as those related to diabetic control (114).    5.1.5 Illustrating the lag-response relationship with distributed lag models An important novelty of Chapter 4 was the focus on the lag-response relationship at the hourly time scale, using a modern statistical method. Previous studies have examined such relationships by including the averaged exposure in specific lag hours as multiple independent variables in one regression model (230), or including exposure averaged over different lag windows in separate models (126, 129, 137, 139-142). The biggest limitation of these methods was that they did not account for the correlation between exposures at different lags. Pradeau et al. (144) addressed this limitation by applying conventional distributed lag models, a methodology originally developed in econometrics (231). However, these models relied on the assumption of a linear effect between the exposure and outcome. Analyses in Chapter 4 applied the new modeling framework of distributed lag non-linear models (DLNMs) which can describe non-linear and lagged effects simultaneously (209). This framework has recently been adapted for case-crossover studies, and a few studies have used it for analyses at the hourly time scale (213). The cumulative effects I found for a 24-hour period using DLNM were consistent with those found in previous studies of 24-hour average exposures using conventional methods, which provides additional evidence on the validity of DLNM, and should encourage the adoption of this method for future studies. 108  5.2 Implications for public health policy and future work 5.2.1 Methodological limitations Although the primary goal of this dissertation was to examine the health effects of sub-daily exposure to wildfire smoke, the exposure model developed in Chapter 3 and the subsequent analyses in Chapter 4 were not exclusive to PM2.5 from wildfire smoke. The OSSEM-1h was trained with total PM2.5 measured at air quality monitors regardless of the source, although it was optimized to reflect smoke-related PM2.5 with predictive variables related to fire activity and data restricted to wildfire seasons. To develop a truly smoke-specific empirical model, source-specific PM2.5 measurements with sufficient temporal and spatial coverage would be needed, but these measurements are rarely available. Some options to improve on this limitation include (1) blending the data with outputs from deterministic models that can produce estimates from only wildfire sources, and (2) identifying and restricting analyses to smoke-impacted times and locations with remotely sensed fire and smoke data. However, this limitation may have less impact in BC than in other locations, given that PM2.5 concentrations are typically very low in summer in the absence of wildfire smoke. Although empirical models can provide exposure estimates at much lower cost than deterministic models, they also have limitations. Empirical models are developed based on past experience, using relationships detected from historical data. Thus, they may not perform as well with new data, especially if those data are drastically different from training data. For example, the wildfire seasons of 2017 and 2018 were unprecedented in BC, but this work was complete before they occurred. The exposure model in Chapter 3 was trained with data from the 2010 through 109  2015 seasons, when the maximum annual area burned was approximately 0.35 million hectares (Figure 1.1) and severe air quality impacts in the densely populated area of greater Vancouver were rare. It is uncertain how the model would perform in 2017 and 2018, when total area burned was greater than 1.2 million hectares in both years, and the greater Vancouver area was under an air quality advisory for weeks. This limitation may affect the applicability of the model in public health surveillance applications. An algorithm that continuously updates the model with recent data and dynamically adapts to new patterns in the data could help to address this limitation, as what has been done in a recently developed forecasting model in US based on the OSSEM-24h model (220), but would be computationally intensive. There were also limitations with the data of health outcomes used in Chapter 4. As with any other administrative health data, ambulance dispatch data were not collected for the purpose of research. For example, missing data on paramedic assessments following ambulance dispatches were quite prevalent, and it remains unclear whether these data were missing at random. The missingness may be related to the types or severity of patient conditions, or the requirements or standards of practice for paramedics in different units. This issue could affect the generalizability of study results specific to paramedic assessments, but not those for dispatches or hospital admissions. In addition, there may still be a lag time between the onset of health conditions and the time at which people made calls to the emergency services. This lag time may vary by conditions and specific circumstances. For example, a patient who had a heart attack while accompanied by someone might contact ambulance services immediately, as compared with a patient who was alone at the time of the event. A condition that occurred abruptly might also lead to a call sooner than a condition that progressed over time, while other conditions may have 110  led to temporary loss of consciousness that delayed ambulance attendance. The sensitivity analysis for calls with and without subsequent hospital admissions (Figure 4.7) also suggest that the severity of the conditions may lead to different lag-response relationships. These different lag times contribute to uncertainty in the assessment of the lag-response relationships I report. With these limitations in mind, the results in Chapter 4 should warrant more studies to examine the lag-response relationship at the hourly time scale, including population-based epidemiologic studies and experimental studies.  5.2.2 Implications for public health policy Public health agencies worldwide have developed guidelines and policies to address the health impacts from wildfire smoke exposure (232-235). Many of these guidelines recommend specific interventions based on evidence derived from studies on urban air pollution rather than wildfire smoke. In addition, their utility may be constrained by the lack of quality evidence on sub-daily exposures over short durations. This dissertation adds to the much-needed evidence for developing and improving public health guidelines for wildfire smoke. The lag-response relationship found in Chapter 4 suggests that the cardiopulmonary health effects were triggered within one hour after exposure, indicating that interventions should be implemented quickly based on 1-hour exposures rather than averages over a longer period of time. For example, in BC, air quality advisories are issued by the Ministry of Environment or Metro Vancouver Regional Government, on behalf of a multi-agency partnerships that involves public health authorities (236, 237). Typically, the advisory will be issued when actual or forecasted air pollutant concentrations exceed a trigger value over a 24-hour average period. However, for 111  wildfire smoke, “there is no trigger value; instead a combination of information is used including PM2.5 levels, presence of smoke plumes, and meteorological conditions to issue an advisory” (237). Based on the results from Chapter 4, wildfire smoke forecasts should play a vital role in this process, as actions based on near real-time air quality observations may already be too late. This also highlights the importance of continuously improving our ability to forecast fire and smoke activity in the future. Moreover, interventions such as setting up clean air shelters or distributing portable air cleaners to communities should be planned and prepared before the wildfire season begins. The public should also be aware that the health effects of wildfire smoke can be immediate, and prepared to take action to reduce their exposures promptly once advised. This would mean stocking up on emergency medications, knowing the location of the nearest clean air shelter, and having a portable air cleaner at home if possible, before wildfire season begins. During a wildfire smoke event, people should be encouraged to seek medical attention promptly if feeling unwell.      In Canada, the Air Quality Health Index (AQHI) is the tool most commonly used to communicate with the public about the health risks associated with air pollution. The AQHI is calculated based on a formula that includes the 3-hour rolling averaged concentrations ground-level ozone (O3), nitrogen dioxide (NO2), and PM2.5. The formula was derived from the concentration-response relationship between daily all-cause mortality and these pollutants in Canadian urban centers, and is more weighted towards the NO2 and O3 components than the PM2.5 component, because the primary source of air pollution is motor vehicle exhaust in these regions (238). The results reported in Chapter 4 highlight the limitations of the AQHI during wildfire smoke events. First, health effects were observed within a 1-hour period, whereas the 112  AQHI is calculated based on 3-hour rolling averages. Second, health effects were associated with PM2.5 alone, whereas the AQHI is calculated using concentrations of three pollutants, and PM2.5 has the lowest of the three coefficients. To address these limitations, my related work at the British Columbia Centre for Disease Control (BCCDC) has compared the 3-pollutant 3-hour AQHI with a 1-pollutant 1-hour amendment during BC wildfire seasons and found that the latter provides a better predictor of population respiratory risk (239). This amendment was implemented across BC in May 2018, and was activated over 50% of the time in regions most impacted by wildfire smoke in 2018 (unpublished report).  One important aspect of public health messaging during wildfire smoke events is to provide advice to the most vulnerable populations. So far, these populations include people with pre-existing cardiopulmonary conditions, the elderly, young children and pregnant women. People with diabetes have sometimes, but not always, been included in these messages, but results from our study, along with another recent study (114), suggest that they should be included.   In Chapter 4, we focused on the lag-response relationship between sub-daily exposure to wildfire smoke and different health outcomes, and only presented results that assumed linear exposure-response relationship. However, in the preliminary analysis, we found that some health outcomes exhibited a non-linear exposure-response relationship. For example, the relationship between PM2.5 concentrations and ambulance calls with an Ambulance Dispatch code for Breathing Problems and Paramedic Assessment code of Asthma/COPD appeared to have thresholds in the 100-150 µg/m3 range (Figure 5.1). This result suggests that the health effects of exposure may not increase much after the concentrations have passed a certain threshold, which may support 113  the capping of indices such as the AQHI. Further investigation should be conducted to support this implication, as the precision of the effect estimates is quite low for exposure higher than 100 µg/m3, due to much fewer data available at that range of exposure.  Figure 5.1 Examples of non-linear exposure-response relationship between PM2.5 and ambulance dispatches.   5.2.3 Recommendations for future work A number of further explorations arise from the work described in this dissertation. First, the ambulance dispatch analyses could be restricted to smoke impacted regions and days identified by remote sensing products or other sources of information about the fire activity (29, 54, 240) so that we can examine the health impact of PM2.5 more specifically from wildfire smoke. With sufficiently large datasets, analyses could also be stratified by sex, age, and other demographic factors to assess potential differences in the health effects among different subgroups. In addition, by linking the ambulance dispatch dataset with other administrative health data, such as 114  the physician visit billing data, it would be possible to identify people with pre-existing health conditions, and examine whether the lag-response relationship or magnitude of effects in these subgroups is different.  So far, evidence has been convincing that wildfire smoke exposure is affecting public health. While we need to continue the accumulation of this evidence, it is becoming more pressing to understand what we can do to mitigate the health impacts. More studies on the effectiveness of different public health interventions should be prioritized to support evidence-based advice for the public during wildfire smoke events. We need to explore how the findings from this dissertation can be incorporated into the public health decision-making process.  Based on this work an interactive online tool for public health professionals and the public could be developed, with easy access to near-real-time estimates from OSSEM-1h, smoke forecasts, and health risk assessment. Daily static reports including similar information have been produced by the BCCDC during wildfire seasons via the BC Asthma Monitoring System (BCAMS) (98) to 2017, and the BC Asthma Prediction System (BCAPS) (192) thereafter. These reports are distributed to public health officials, but the implementation of an interactive online tool is underway. Another phase of the tool could allow the public to provide their own assessment or perception of the smoke situation and their health responses in real time, as is being done using smartphone applications elsewhere (241, 242). Such crowd-sourcing information would be especially useful for regions without air quality monitors and can also be collected and used to evaluate and improve the performance of smoke modelling, surveillance, and epidemiology in these regions.    115  In the future, interdisciplinary collaboration will be essential for advancement in understanding and mitigating the health impacts of wildfire smoke. Collaboration between data scientists and atmospheric modelers can improve exposure assessment via two paths: (1) by blending data from empirical and deterministic models, and (2) by sharing insights gained with the two different approaches. For example, blending estimates from OSSEM-1h with outputs from the FireWork smoke forecasting system (77) developed by Environment and Climate Change Canada (ECCC) may produce more accurate forecasts (224). Moreover, the collaboration between the remote sensing communities, forest and wildfire specialists, and health researchers, can better design and produce remote sensing products for wildfire smoke research and surveillance. A recent example of such collaboration is the Multi-Angle Imager for Aerosols (MAIA) satellite instrument that produces data specifically for studying the human health impact of air pollution, a project resulted from the partnership of the US National Aeronautics and Space Adminstration (NASA), epidemiologists, and health organizations (243). Meanwhile, the Canadian Space Agency (CSA), the Canadian Forest Service (CFS), and ECCC are exploring the feasibility of launching a new satellite to better support wildfire management and air quality forecasting in Canada (244). More of these collaborative efforts will provide new opportunities for us to further refine our ability to assess the exposure to wildfire smoke and its health impacts. Wildfires are becoming a new normal in our lives under climate change, and wildfire smoke is a growing source of air pollution and public health burden for more communities, in Canada and globally. Unlike many other sources of air pollution, it is impossible to implement regulatory top-down control measures for wildfire emissions, such as those used to control traffic and industrial emissions. Reducing the health impacts of wildfire smoke will rely on (1) behavioural 116  change in the general public to reduce exposure during smoke events and (2) modifications to the built environment to limit the infiltration of wildfire smoke into indoor environments. Environmental epidemiologists, public health professionals, forest and fire experts, atmospheric scientists, toxicologists, and engineers must produce the necessary scientific evidence, and then take the responsibility to mobilize the knowledge to the public and policy-makers, and advocate for the necessary protective changes in behaviours and the built environment.    117  Bibliography 1. Giglio L, Randerson JT, Werf GR. Analysis of daily, monthly, and annual burned area using the fourth‐generation global fire emissions database (GFED4). Journal of Geophysical Research: Biogeosciences. 2013;118(1):317-28. 2. Andela N, Morton D, Giglio L, Chen Y, Van Der Werf G, Kasibhatla P, et al. A human-driven decline in global burned area. Science. 2017;356(6345):1356-62. 3. Jain P, Wang X, Flannigan MD. Trend analysis of fire season length and extreme fire weather in North America between 1979 and 2015. International Journal of Wildland Fire. 2018;26(12):1009-20. 4. Dennison PE, Brewer SC, Arnold JD, Moritz MA. Large wildfire trends in the western United States, 1984–2011. Geophysical Research Letters. 2014;41(8):2928-33. 5. Aldersley A, Murray SJ, Cornell SE. Global and regional analysis of climate and human drivers of wildfire. Science of the Total Environment. 2011;409(18):3472-81. 6. Westerling AL, Hidalgo HG, Cayan DR, Swetnam TW. Warming and earlier spring increase western US forest wildfire activity. science. 2006;313(5789):940-3. 7. Wotton B, Flannigan M, Marshall G. Potential climate change impacts on fire intensity and key wildfire suppression thresholds in Canada. Environmental Research Letters. 2017;12(9):095003. 8. Barbero R, Abatzoglou J, Larkin N, Kolden C, Stocks B. Climate change presents increased potential for very large fires in the contiguous United States. International Journal of Wildland Fire. 2015;24(7):892-9. 9. Dempsey F. Forest Fire Effects on Air Quality in Ontario: Evaluation of Several Recent Examples. B Am Meteorol Soc. 2013;94(7):1059-64. 10. Jeong JI, Park RJ, Youn D. Effects of Siberian forest fires on air quality in East Asia during May 2003 and its climate implication. Atmos Environ. 2008;42(39):8910-22. 11. Miller DJ, Sun K, Zondlo MA, Kanter D, Dubovik O, Welton EJ, et al. Assessing boreal forest fire smoke aerosol impacts on U.S. air quality: A case study using multiple data sets. J Geophys Res-Atmos. 2011;116:D22. 12. Dirksen RJ, Folkert Boersma K, de Laat J, Stammes P, van der Werf GR, Val Martin M, et al. An aerosol boomerang: Rapid around-the-world transport of smoke from the December 2006 Australian forest fires observed from space. Journal of Geophysical Research: Atmospheres. 2009;114(D21):n/a-n/a. 118  13. Jaffe D, Hafner W, Chand D, Westerling A, Spracklen D. Interannual variations in PM2. 5 due to wildfires in the western United States. Environmental Science & Technology. 2008;42(8):2812-8. 14. Zhang Y, Tao S. Global atmospheric emission inventory of polycyclic aromatic hydrocarbons (PAHs) for 2004. Atmos Environ. 2009;43(4):812-9. 15. Langmann B, Duncan B, Textor C, Trentmann J, van der Werf GR. Vegetation fire emissions and their impact on air pollution and climate. Atmos Environ. 2009;43(1):107-16. 16. Lelieveld J, Evans J, Fnais M, Giannadaki D, Pozzer A. The contribution of outdoor air pollution sources to premature mortality on a global scale. Nature. 2015;525(7569):367-71. 17. Stanaway JD, Afshin A, Gakidou E, Lim SS, Abate D, Abate KH, et al. Global, regional, and national comparative risk assessment of 84 behavioural, environmental and occupational, and metabolic risks or clusters of risks for 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. The Lancet. 2018;392(10159):1923-94. 18. Naeher LP, Brauer M, Lipsett M, Zelikoff JT, Simpson CD, Koenig JQ, et al. Woodsmoke health effects: a review. IInhalation Toxicology. 2007;19(1):67-106. 19. Liu JC, Pereira G, Uhl SA, Bravo MA, Bell ML. A systematic review of the physical health impacts from non-occupational exposure to wildfire smoke. Environ Res. 2015;136:120-32. 20. Reid CE, Brauer M, Johnston FH, Jerrett M, Balmes JR, Elliott CT. Critical Review of Health Impacts of Wildfire Smoke Exposure. Environmental health perspectives. 2016;124(9):1334-43. 21. Youssouf H, Liousse C, Roblou L, Assamoi E-M, Salonen R, Maesano C, et al. Non-Accidental Health Impacts of Wildfire Smoke. International Journal of Environmental Research and Public Health. 2014;11(11):11772-804. 22. Kochi I, Champ PA, Loomis JB, Donovan GH. Valuing mortality impacts of smoke exposure from major southern California wildfires. J For Econ. 2012;18(1):61-75. 23. Richardson LA, Champ PA, Loomis JB. The hidden cost of wildfires: Economic valuation of health effects of wildfire smoke exposure in Southern California. J For Econ. 2012;18(1):14-35. 24. Kochi I, Donovan GH, Champ PA, Loomis JB. The economic cost of adverse health effects from wildfire-smoke exposure: a review. International Journal of Wildland Fire. 2010;19(7):803-17. 119  25. Fann N, Fulcher CM, Baker K. The Recent and Future Health Burden of Air Pollution Apportioned Across U.S. Sectors. Environmental Science & Technology. 2013;47(8):3580-9. 26. Dennekamp M, Straney LD, Erbas B, Abramson MJ, Keywood M, Smith K, et al. Forest Fire Smoke Exposures and Out-of-Hospital Cardiac Arrests in Melbourne, Australia: A Case-Crossover Study. Environmental health perspectives. 2015;123(10):959-64. 27. Hodzic A, Madronich S, Bohn B, Massie S, Menut L, Wiedinmyer C. Wildfire particulate matter in Europe during summer 2003: meso-scale modeling of smoke emissions, transport and radiative effects. Atmos Chem Phys. 2007;7(15):4043-64. 28. Yao J, Brauer M, Henderson SB. Evaluation of a Wildfire Smoke Forecasting System as a Tool for Public Health Protection. Environmental health perspectives. 2013;121(10):1142. 29. Yao J, Eyamie J, Henderson SB. Evaluation of a spatially resolved forest fire smoke model for population-based epidemiologic exposure assessment. Journal of Exposure Science and Environmental Epidemiology. 2016;26(3):233. 30. Reid CE, Jerrett M, Tager IB, Petersen ML, Mann JK, Balmes JR. Differential respiratory health effects from the 2008 northern California wildfires: A spatiotemporal approach. Environ Res. 2016;150:227-35. 31. Henderson SB, Brauer M, Macnab YC, Kennedy SM. Three measures of forest fire smoke exposure and their associations with respiratory and cardiovascular health outcomes in a population-based cohort. Environmental health perspectives. 2011;119(9):1266-71. 32. Kollanus V, Prank M, Gens A, Soares J, Vira J, Kukkonen J, et al. Mortality due to Vegetation-Fire Originated PM 2.5 Exposure in Europe–Assessment for the Years 2005 and 2008. Environmental health perspectives. 2017;125:30-7. 33. Faustini A, Alessandrini ER, Pey J, Perez N, Samoli E, Querol X, et al. Short-term effects of particulate matter on mortality during forest fires in Southern Europe: results of the MED-PARTICLES Project. Occupational and environmental medicine. 2015;72(5):323-9. 34. Paciorek CJ, Liu Y. Assessment and statistical modeling of the relationship between remotely sensed aerosol optical depth and PM2. 5 in the eastern United States. Research report (Health Effects Institute). 2012(167):5-83; discussion 5-91. 35. Liu Y, Wang Z, Wang J, Ferrare RA, Newsom RK, Welton EJ. The effect of aerosol vertical profiles on satellite-estimated surface particle sulfate concentrations. Remote sensing of environment. 2011;115(2):508-13. 36. Statistics Canada. Table  051-0001 -  Estimates of population, by age group and sex for July 1, Canada, provinces and territories, annual (persons unless otherwise noted),  CANSIM (database). Accessed: 2018-01-12. 2017. 120  37. Ministry of Forests and Ministry of Sustainable Resource Management. British Columbia’s forests: a geographical snapshot 2003. Available from: https://www.for.gov.bc.ca/hfd/pubs/Docs/Mr/Mr112/BC_Forests_Geographical_Snapshot.pdf. 38. British Columbia Ministry of Environment. Indicators of Climate Change for British Columbia 2016 Update 2016. Available from: https://www2.gov.bc.ca/assets/gov/environment/research-monitoring-and-reporting/reporting/envreportbc/archived-reports/climate-change/climatechangeindicators-13sept2016_final.pdf. 39. Aukema BH, Carroll AL, Zhu J, Raffa KF, Sickley TA, Taylor SW. Landscape level analysis of mountain pine beetle in British Columbia, Canada: spatiotemporal development and spatial synchrony within the present outbreak. Ecography. 2006;29(3):427-41. 40. Taylor SW, Carroll AL, Alfaro RI, Safranyik L. Forest, climate and mountain pine beetle outbreak dynamics in western Canada. The mountain pine beetle: A synthesis of biology, management, and impacts on lodgepole pine. 2006:67-94. 41. British Columbia Lung Association. State of the Air: 2016. On-line: https://bclungca/sites/default/files/State%20of%20the%20Air%202017%20-%20mergedpdf [cited 1-17-2018]. 2017. 42. Yao J, Henderson SB. An empirical model to estimate daily forest fire smoke exposure over a large geographic area using air quality, meteorological, and remote sensing data. J Expos Sci Environ Epidemiol. 2014;24(3):328-35. 43. Reisen F, Duran SM, Flannigan M, Elliott C, Rideout K. Wildfire smoke and public health risk. International Journal of Wildland Fire. 2015;24(8):1029-44. 44. Kavouras IG, DuBois DW, Etyemezian V, Nikolich G. Spatiotemporal variability of ground-level ozone and influence of smoke in Treasure Valley, Idaho. Atmos Res. 2013;124:44-52. 45. Keane JC. Air quality and visibility in Southwestern British Columbia during forest fire smoke events: University of British Columbia; 2012. 46. Tan WC, Qiu D, Liam BL, Ng TP, Lee SH, van Eeden SF, et al. The human bone marrow response to acute air pollution caused by forest fires. Am J Respir Crit Care Med. 2000;161(4 Pt 1):1213-7. 47. Gilman J, Lerner B, Kuster W, Goldan P, Warneke C, Veres P, et al. Biomass burning emissions and potential air quality impacts of volatile organic compounds and other trace gases from fuels common in the US. Atmos Chem Phys. 2015;15(24):13915-38. 121  48. Vicente A, Alves C, Monteiro C, Nunes T, Mirante F, Evtyugina M, et al. Measurement of trace gases and organic compounds in the smoke plume from a wildfire in Penedono (central Portugal). Atmos Environ. 2011;45(29):5172-82. 49. Reid J, Koppmann R, Eck T, Eleuterio D. A review of biomass burning emissions part II: intensive physical properties of biomass burning particles. Atmos Chem Phys. 2005;5(3):799-825. 50. Kim YH, Warren SH, Krantz QT, King C, Jaskot R, Preston WT, et al. Mutagenicity and Lung Toxicity of Smoldering vs. Flaming Emissions from Various Biomass Fuels: Implications for Health Effects from Wildland Fires. Environmental health perspectives. 2018;126(1):017011-. 51. Kim YH, King C, Krantz T, Hargrove MM, George IJ, McGee J, et al. The role of fuel type and combustion phase on the toxicity of biomass smoke following inhalation exposure in mice. Archives of Toxicology. 2019:1-13. 52. Brunekreef B, Holgate ST. Air pollution and health. The lancet. 2002;360(9341):1233-42. 53. Valavanidis A, Fiotakis K, Vlachogianni T. Airborne particulate matter and human health: toxicological assessment and importance of size and composition of particles for oxidative damage and carcinogenic mechanisms. Journal of Environmental Science and Health, Part C. 2008;26(4):339-62. 54. Elliott CT, Henderson SB, Wan V. Time series analysis of fine particulate matter and asthma reliever dispensations in populations affected by forest fires. Environmental health : a global access science source. 2013;12:11. 55. McKendry I, Strawbridge K, Karumudi M, O'Neill N, Macdonald A, Leaitch R, et al. Californian forest fire plumes over Southwestern British Columbia: lidar, sunphotometry, and mountaintop chemistry observations. Atmos Chem Phys. 2011;11(2):465-77. 56. Cottle P, Strawbridge K, McKendry I. Long-range transport of Siberian wildfire smoke to British Columbia: Lidar observations and air quality impacts. Atmos Environ. 2014;90:71-7. 57. Jaffe D, Bertschi I, Jaeglé L, Novelli P, Reid JS, Tanimoto H, et al. Long‐range transport of Siberian biomass burning emissions and impact on surface ozone in western North America. Geophysical Research Letters. 2004;31(16):L16106. 58. Johnston FH, Bailie RS, Pilotto LS, Hanigan IC. Ambient biomass smoke and cardio-respiratory hospital admissions in Darwin, Australia. BMC public health. 2007;7:240. 59. Hanninen OO, Salonen RO, Koistinen K, Lanki T, Barregard L, Jantunen M. Population exposure to fine particles and estimated excess mortality in Finland from an East European 122  wildfire episode. Journal of exposure science & environmental epidemiology. 2009;19(4):414-22. 60. Lee TS, Falter K, Meyer P, Mott J, Gwynn C. Risk factors associated with clinic visits during the 1999 forest fires near the Hoopa Valley Indian Reservation, California, USA. International journal of environmental health research. 2009;19(5):315-27. 61. Tham R, Erbas B, Akram M, Dennekamp M, Abramson MJ. The impact of smoke on respiratory hospital outcomes during the 2002-2003 bushfire season, Victoria, Australia. Respirology. 2009;14(1):69-75. 62. Morgan G, Sheppeard V, Khalaj B, Ayyar A, Lincoln D, Jalaludin B, et al. Effects of bushfire smoke on daily mortality and hospital admissions in Sydney, Australia. Epidemiology. 2010;21(1):47-55. 63. Vora C, Renvall MJ, Chao P, Ferguson P, Ramsdell JW. 2007 San Diego wildfires and asthmatics. The Journal of asthma : official journal of the Association for the Care of Asthma. 2011;48(1):75-8. 64. Crabbe H. Risk of respiratory and cardiovascular hospitalisation with exposure to bushfire particulates: new evidence from Darwin, Australia. Environ Geochem Hlth. 2012;34(6):697-709. 65. Kolbe A, Gilchrist KL. An extreme bushfire smoke pollution event: health impacts and public health challenges. New South Wales public health bulletin. 2009;20(1-2):19-23. 66. Schranz CI, Castillo EM, Vilke GM. The 2007 San Diego Wildfire impact on the Emergency Department of the University of California, San Diego Hospital System. Prehospital and disaster medicine. 2010;25(5):472-6. 67. Resnick A, Woods B, Krapfl H, Toth B. Health outcomes associated with smoke exposure in Albuquerque, New Mexico, during the 2011 Wallow fire. Journal of public health management and practice : JPHMP. 2015;21 Suppl 2:S55-61. 68. Yao J. Evaluation of the BlueSky smoke forecasting system and its utility for public health protection in British Columbia: University of British Columbia; 2012. 69. Rappold AG, Stone SL, Cascio WE, Neas LM, Kilaru VJ, Carraway MS, et al. Peat bog wildfire smoke exposure in rural North Carolina is associated with cardiopulmonary emergency department visits assessed through syndromic surveillance. Environmental health perspectives. 2011;119(10):1415-20. 70. Holstius DM, Reid CE, Jesdale BM, Morello-Frosch R. Birth Weight following Pregnancy during the 2003 Southern California Wildfires. Environmental health perspectives. 2012;120(9):1340-5. 123  71. Price OF, Williamson GJ, Henderson SB, Johnston F, Bowman DM. The relationship between particulate pollution levels in Australian cities, meteorology, and landscape fire activity detected from MODIS hotspots. PloS one. 2012;7(10):e47327. 72. Kahn RA, Li WH, Moroney C, Diner DJ, Martonchik JV, Fishbein E. Aerosol source plume physical characteristics from space‐based multiangle imaging. Journal of Geophysical Research: Atmospheres. 2007;112(D11). 73. Kahn RA, Chen Y, Nelson DL, Leung FY, Li Q, Diner DJ, et al. Wildfire smoke injection heights: Two perspectives from space. Geophysical Research Letters. 2008;35(4): L04809. 74. Zannetti P, Anfossi D. Air Quality Modeling: Theories, Methodologies, Computational Techniques, and Available Databases and Software: EnviroComp; 2003. 75. Sakiyama S. The bluesky western canada wildfire smoke forecasting system. Victoria, BC: British Columbia Ministry of Environment. 2013. 76. Larkin NK, O’Neill SM, Solomon R, Raffuse S, Strand T, Sullivan DC, et al. The BlueSky smoke modeling framework. International Journal of Wildland Fire. 2010;18(8):906-20. 77. Pavlovic R, Chen J, Anderson K, Moran MD, Beaulieu P-A, Davignon D, et al. The FireWork air quality forecast system with near-real-time biomass burning emissions: Recent developments and evaluation of performance for the 2015 North American wildfire season. Journal of the Air & Waste Management Association. 2016;66(9):819-41. 78. Raffuse SM, Craig KJ, Larkin NK, Strand TT, Sullivan DC, Wheeler NJ, et al. An evaluation of modeled plume injection height with satellite-derived observed plume height. Atmosphere. 2012;3(1):103-23. 79. Sofiev M, Ermakova T, Vankevich R. Evaluation of the smoke-injection height from wild-land fires using remote-sensing data. Atmos Chem Phys. 2012;12(4):1995-2006. 80. Sessions W, Fuelberg H, Kahn R, Winker D. An investigation of methods for injecting emissions from boreal wildfires using WRF-Chem during ARCTAS. Atmos Chem Phys. 2011;11(12):5719-44. 81. Val Martin M, Logan JA, Kahn RA, Leung F-Y, Nelson DL, Diner DJ. Smoke injection heights from fires in North America: analysis of 5 years of satellite observations. Atmos Chem Phys. 2010;10(4):1491-510. 82. Peterson D, Hyer E, Wang J. Quantifying the potential for high‐altitude smoke injection in the North American boreal forest using the standard MODIS fire products and subpixel‐based methods. Journal of Geophysical Research: Atmospheres. 2014;119(6):3401-19. 124  83. Val Martin M, Kahn RA, Logan JA, Paugam R, Wooster M, Ichoku C. Space‐based observational constraints for 1‐D fire smoke plume‐rise models. Journal of Geophysical Research: Atmospheres. 2012;117(D22). 84. Mazzoni D, Logan JA, Diner D, Kahn R, Tong L, Li Q. A data-mining approach to associating MISR smoke plume heights with MODIS fire measurements. Remote Sensing of Environment. 2007;107(1–2):138-48. 85. Freitas S, Longo K, Andreae M. Impact of including the plume rise of vegetation fires in numerical simulations of associated atmospheric pollutants. Geophysical Research Letters. 2006;33(17):L17808. 86. Freitas SR, Longo KM, Chatfield R, Latham D, Silva Dias M, Andreae M, et al. Including the sub-grid scale plume rise of vegetation fires in low resolution atmospheric transport models. Atmos Chem Phys. 2007;7(13):3385-98. 87. Rappold AG, Cascio WE, Kilaru VJ, Stone SL, Neas LM, Devlin RB, et al. Cardio-respiratory outcomes associated with exposure to wildfire smoke are modified by measures of community health. Environmental health : a global access science source. 2012;11:71. 88. Liu JC, Wilson A, Mickley LJ, Ebisu K, Sulprizio MP, Wang Y, et al. Who Among the Elderly Is Most Vulnerable to Exposure to and Health Risks of Fine Particulate Matter From Wildfire Smoke? Am J Epidemiol. 2017;186(6):730-5. 89. Liu JC, Wilson A, Mickley LJ, Dominici F, Ebisu K, Wang Y, et al. Wildfire-specific Fine Particulate Matter and Risk of Hospital Admissions in Urban and Rural Counties. Epidemiology. 2017;28(1):77-85. 90. Hanigan IC, Johnston FH, Morgan GG. Vegetation fire smoke, indigenous status and cardio-respiratory hospital admissions in Darwin, Australia, 1996-2005: a time-series study. Environmental health : a global access science source. 2008;7:42. 91. Delfino RJ, Brummel S, Wu J, Stern H, Ostro B, Lipsett M, et al. The relationship of respiratory and cardiovascular hospital admissions to the southern California wildfires of 2003. Occupational and environmental medicine. 2009;66(3):189-97. 92. Thelen B, French NH, Koziol BW, Billmire M, Owen RC, Johnson J, et al. Modeling acute respiratory illness during the 2007 San Diego wildland fires using a coupled emissions-transport system and generalized additive modeling. Environmental Health. 2013;12(1):94. 93. Goodrick SL, Achtemeier GL, Larkin NK, Liu Y, Strand TM. Modelling smoke transport from wildland fires: a review. International Journal of Wildland Fire. 2013;22(1):83-94. 125  94. Rolph GD, Draxler RR, Stein AF, Taylor A, Ruminski MG, Kondragunta S, et al. Description and Verification of the NOAA Smoke Forecasting System: The 2007 Fire Season. Weather Forecast. 2009;24(2):361-78. 95. Strand TM, Larkin N, Craig KJ, Raffuse S, Sullivan D, Solomon R, et al. Analyses of BlueSky Gateway PM2.5 predictions during the 2007 southern and 2008 northern California fires. J Geophys Res-Atmos. 2012;117:D17301. 96. Reid CE, Jerrett M, Petersen ML, Pfister GG, Morefield PE, Tager IB, et al. Spatiotemporal Prediction of Fine Particulate Matter During the 2008 Northern California Wildfires Using Machine Learning. Environmental Science & Technology. 2015;49(6):3887-96. 97. Preisler HK, Schweizer D, Cisneros R, Procter T, Ruminski M, Tarnay L. A statistical model for determining impact of wildland fires on Particulate Matter (PM2.5) in Central California aided by satellite imagery of smoke. Environmental Pollution. 2015;205(0):340-9. 98. McLean KE, Yao J, Henderson SB. An Evaluation of the British Columbia Asthma Monitoring System (BCAMS) and PM2.5 Exposure Metrics during the 2014 Forest Fire Season. Int J Environ Res Public Health. 2015;12(6):6710-24. 99. Lassman W, Ford B, Gan RW, Pfister G, Magzamen S, Fischer EV, et al. Spatial and temporal estimates of population exposure to wildfire smoke during the Washington state 2012 wildfire season using blended model, satellite, and in situ data. GeoHealth. 2017;1(3):106-21. 100. Mohri M, Rostamizadeh A, Talwalkar A. Foundations of machine learning: MIT press; 2012. 101. Langley P. Elements of machine learning: Morgan Kaufmann; 1996. 102. Samuel AL. Some studies in machine learning using the game of checkers. IBM Journal of research and development. 1959;3(3):210-29. 103. Bellinger C, Jabbar MSM, Zaïane O, Osornio-Vargas A. A systematic review of data mining and machine learning for air pollution epidemiology. BMC public health. 2017;17(1):907. 104. Lightstone SD, Moshary F, Gross B. Comparing CMAQ Forecasts with a Neural Network Forecast Model for PM2. 5 in New York. Atmosphere. 2017;8(9):161. 105. Ma D, Zhang Z. Contaminant dispersion prediction and source estimation with integrated Gaussian-machine learning network model for point source emission in atmosphere. Journal of hazardous materials. 2016;311:237-45. 106. Xi X, Wei Z, Xiaoguang R, Yijie W, Xinxin B, Wenjun Y, et al., editors. A comprehensive evaluation of air pollution prediction improvement by a machine learning 126  method. Service Operations And Logistics, And Informatics (SOLI), 2015 IEEE International Conference on; 2015: IEEE. 107. Pelliccioni A, Cristofari A, Silibello C, Gherardi M, Cecinato A, Lamberti M, editors. Estimation of PAHs concentration fields in an urban area by means of Support Vector Machines. 7th Intl Congress on Env Modelling and Software; 2014; San Diego, CA. 108. Breiman L. Random forests. Machine learning. 2001;45(1):5-32. 109. Adetona O, Reinhardt TE, Domitrovich J, Broyles G, Adetona AM, Kleinman MT, et al. Review of the health effects of wildland fire smoke on wildland firefighters and the public. Inhalation toxicology. 2016;28(3):95-139. 110. Shaposhnikov D, Revich B, Bellander T, Bedada GB, Bottai M, Kharkova T, et al. Mortality related to air pollution with the moscow heat wave and wildfire of 2010. Epidemiology. 2014;25(3):359-64. 111. Haikerwal A, Akram M, Del Monaco A, Smith K, Sim MR, Meyer M, et al. Impact of Fine Particulate Matter (PM2.5) Exposure During Wildfires on Cardiovascular Health Outcomes. J Am Heart Assoc. 2015;4(7). 112. Tinling MA, West JJ, Cascio WE, Kilaru V, Rappold AG. Repeating cardiopulmonary health effects in rural North Carolina population during a second large peat wildfire. Environmental health : a global access science source. 2016;15:12. 113. DeFlorio-Barker S, Crooks J, Reyes J, Rappold AG. Cardiopulmonary Effects of Fine Particulate Matter Exposure among Older Adults, during Wildfire and Non-Wildfire Periods, in the United States 2008–2010. Environmental health perspectives. 2019;127(3):037006. 114. Johnston FH, Salimi F, Williamson GJ, Henderson SB, Yao J, Dennekamp M, et al. Ambient particulate matter and paramedic assessments of acute diabetic, cardiovascular, respiratory conditions. Epidemiology. 2018;Publish Ahead of Print. 115. Black C, Gerriets JE, Fontaine JH, Harper RW, Kenyon NJ, Tablin F, et al. Early life wildfire smoke exposure is associated with immune dysregulation and lung function decrements in adolescence. American journal of respiratory cell and molecular biology. 2017;56(5):657-66. 116. Tse K, Chen L, Tse M, Zuraw B, Christiansen S. Effect of catastrophic wildfires on asthmatic outcomes in obese children: breathing fire. Annals of allergy, asthma & immunology : official publication of the American College of Allergy, Asthma, & Immunology. 2015;114(4):308-11.e4. 117. Ghio AJ. Mechanism of asthmatic exacerbation by ambient air pollution particles. Expert Review of Respiratory Medicine. 2008;2(1):109-18. 127  118. Brook RD, Rajagopalan S, Pope CA, 3rd, Brook JR, Bhatnagar A, Diez-Roux AV, et al. Particulate matter air pollution and cardiovascular disease: An update to the scientific statement from the American Heart Association. Circulation. 2010;121(21):2331-78. 119. Burgan O, Smargiassi A, Perron S, Kosatsky T. Cardiovascular effects of sub-daily levels of ambient fine particles: a systematic review. Environmental Health. 2010;9(1):26. 120. Chen D, Zhang F, Yu C, Jiao A, Xiang Q, Yu Y, et al. Hourly associations between exposure to ambient particulate matter and emergency department visits in an urban population of Shenzhen, China. Atmos Environ. 2019. 121. Delfino RJ, Staimer N, Gillen D, Tjoa T, Sioutas C, Fung K, et al. Personal and Ambient Air Pollution is Associated with Increased Exhaled Nitric Oxide in Children with Asthma. Environmental health perspectives. 2006;114(11):1736-43. 122. Adamkiewicz G, Ebelt S, Syring M, Slater J, Speizer FE, Schwartz J, et al. Association between air pollution exposure and exhaled nitric oxide in an elderly population. Thorax. 2004;59(3):204-9. 123. Mar TF, Jansen K, Shepherd K, Lumley T, Larson TV, Koenig JQ. Exhaled Nitric Oxide in Children with Asthma and Short-Term PM(2.5) Exposure in Seattle. Environmental health perspectives. 2005;113(12):1791-4. 124. Yamazaki S, Shima M, Ando M, Nitta H, Watanabe H, Nishimuta T. Effect of hourly concentration of particulate matter on peak expiratory flow in hospitalized children: a panel study. Environmental health : a global access science source. 2011;10:15. 125. Murakami Y, Ono M. Myocardial infarction deaths after high level exposure to particulate matter. J Epidemiol Commun Health. 2006;60:262 - 6. 126. Peters A, Dockery D, Muller J, Mittleman M. Increased particulate air pollution and the triggering of myocardial infarction. Circulation. 2001;103:2810 - 5. 127. Peters A, von Klot S, Heier M, Trentinaglia I, Cyrys J, Hormann A, et al. Particulate air pollution and nonfatal cardiac events. Part I. Air pollution, personal activities, and onset of myocardial infarction in a case-crossover study. Res Rep Health Eff Inst. 2005;124:1 - 66. 128. Straney L, Finn J, Dennekamp M, Bremner A, Tonkin A, Jacobs I. Evaluating the impact of air pollution on the incidence of out-of-hospital cardiac arrest in the Perth Metropolitan Region: 2000-2010. Journal of epidemiology and community health. 2014;68(1):6-12. 129. Rosenthal F, Carney J, Olinger M. Out-of-hospital cardiac arrest and airborne fine particulate matter: a case-crossover analysis of emergency medical services data in Indianapolis, Indiana. Environmental health perspectives. 2008;116:631 - 536. 128  130. Urch B, Silverman F, Corey P, Brook J, Lukic K, Rajagopalan S, et al. Acute blood pressure responses in healthy adults during controlled air pollution exposures. Environmental health perspectives. 2005;113:1052 - 5. 131. Soppa VJ, Schins RP, Hennig F, Nieuwenhuijsen MJ, Hellack B, Quass U, et al. Arterial blood pressure responses to short-term exposure to fine and ultrafine particles from indoor sources–A randomized sham-controlled exposure study of healthy volunteers. Environmental research. 2017;158:225-32. 132. Lucking A, Lundback M, Mills N, Faratian D, Barath S, Pourazar J, et al. Diesel exhaust inhalation increases thrombus formation in man. Eur Heart J. 2008;29:3043 - 51. 133. Lundbäck M, Mills NL, Lucking A, Barath S, Donaldson K, Newby DE, et al. Experimental exposure to diesel exhaust increases arterial stiffness in man. Particle and fibre toxicology. 2009;6(1):7. 134. Salvi S, Blomberg A, Rudell B, Kelly F, Sandstrom T, Holgate S, et al. Acute inflammatory responses in the airways and peripheral blood after short-term exposure to diesel exhaust in healthy human volunteers. Am J Respir Crit Care Med. 1999;159:702 - 9. 135. Nightingale JA, Maggs R, Cullinan P, Donnelly LE, Rogers DF, Kinnersley R, et al. Airway inflammation after controlled exposure to diesel exhaust particulates. American journal of respiratory and critical care medicine. 2000;162(1):161-6. 136. Xu Y, Barregard L, Nielsen J, Gudmundsson A, Wierzbicka A, Axmon A, et al. Effects of diesel exposure on lung function and inflammation biomarkers from airway and peripheral blood of healthy volunteers in a chamber study. Particle and fibre toxicology. 2013;10(1):60. 137. Ensor KB, Raun LH, Persse D. A case-crossover analysis of out-of-hospital cardiac arrest and air pollution. Circulation. 2013:CIRCULATIONAHA. 113.000027. 138. Raza A, Bellander T, Bero-Bedada G, Dahlquist M, Hollenberg J, Jonsson M, et al. Short-term effects of air pollution on out-of-hospital cardiac arrest in Stockholm. European heart journal. 2013;35(13):861-8. 139. Sullivan J, Sheppard L, Schreuder A, Ishikawa N, Siscovick D, Kaufman J. Relation between short-term fine-particulate matter exposure and onset of myocardial infarction. Epidemiology. 2005;16:41 - 8. 140. Wichmann J, Folke F, Torp-Pedersen C, Lippert F, Ketzel M, Ellermann T, et al. Out-of-hospital cardiac arrests and outdoor air pollution exposure in Copenhagen, Denmark. PloS one. 2013;8(1):e53684. 129  141. Evans KA, Hopke PK, Utell MJ, Kane C, Thurston SW, Ling FS, et al. Triggering of ST-elevation myocardial infarction by ambient wood smoke and other particulate and gaseous pollutants. Journal of Exposure Science and Environmental Epidemiology. 2017;27(2):198. 142. Gardner B, Ling F, Hopke PK, Frampton MW, Utell MJ, Zareba W, et al. Ambient fine particulate air pollution triggers ST-elevation myocardial infarction, but not non-ST elevation myocardial infarction: a case-crossover study. Particle and fibre toxicology. 2014;11(1):1. 143. Link MS, Luttmann-Gibson H, Schwartz J, Mittleman MA, Wessler B, Gold DR, et al. Acute exposure to air pollution triggers atrial fibrillation. Journal of the American College of Cardiology. 2013;62(9):816-25. 144. Pradeau C, Rondeau V, Lévèque E, Guernion P-Y, Tentillier E, Thicoipé M, et al. Air pollution and activation of mobile medical team for out-of-hospital cardiac arrest. The American journal of emergency medicine. 2015;33(3):367-72. 145. Rosenthal FS, Kuisma M, Lanki T, Hussein T, Boyd J, Halonen JI, et al. Association of ozone and particulate air pollution with out-of-hospital cardiac arrest in Helsinki, Finland: evidence for two different etiologies. Journal of Exposure Science and Environmental Epidemiology. 2013;23(3):281. 146. Panagiotoglou D. The effects of British Columbia hospital closures on delivery of health care services and the population's health: University of British Columbia; 2016. 147. Salimi F, Henderson SB, Morgan GG, Jalaludin B, Johnston FH. Ambient particulate matter, landscape fire smoke, and emergency ambulance dispatches in Sydney, Australia. Environment international. 2017;99:208-12. 148. Youngquist ST, Hood CH, Hales NM, Barton CC, Madsen TE, Pope CA. Association between EMS calls and fine particulate air pollution in Utah. Air Quality, Atmosphere & Health. 2016;9(8):887-97. 149. Michikawa T, Ueda K, Takeuchi A, Kinoshita M, Hayashi H, Ichinose T, et al. Impact of short-term exposure to fine particulate matter on emergency ambulance dispatches in Japan. Journal of epidemiology and community health. 2015;69(1):86-91. 150. Sajani SZ, Alessandrini E, Marchesi S, Lauriola P. Are day-to-day variations of airborne particles associated with emergency ambulance dispatches? International journal of occupational and environmental health. 2014;20(1):71-6. 151. Winker DM, Hostetler C, Vaughan M, Omar A. CALIOP Algorithm Theoretical Basis Document, Part 1: CALIOP Instrument, and Algorithms Overview. Release. 2006;2:29. 152. Vaughan M, Winker DM, Powell K. CALIOP algorithm theoretical basis document, part 2: Feature detection and layer properties algorithms. Rep PC-SCI. 2005;202:87. 130  153. Liu Z, Omar A, Hu Y, Vaughan M, Winker D, Poole L, et al. CALIOP algorithm theoretical basis document. Part 3: Scene classification algorithms. NASA-CNES document PC-SCI-203. 2005. 154. Omar AH, Winker DM, Vaughan MA, Hu Y, Trepte CR, Ferrare RA, et al. The CALIPSO automated aerosol classification and lidar ratio selection algorithm. Journal of Atmospheric and Oceanic Technology. 2009;26(10):1994-2014. 155. Liu Z, Vaughan M, Winker D, Kittaka C, Getzewich B, Kuehn R, et al. The CALIPSO lidar cloud and aerosol discrimination: Version 2 algorithm and initial assessment of performance. Journal of Atmospheric and Oceanic Technology. 2009;26(7):1198-213. 156. Vaughan MA, Powell KA, Winker DM, Hostetler CA, Kuehn RE, Hunt WH, et al. Fully automated detection of cloud and aerosol layers in the CALIPSO lidar measurements. Journal of Atmospheric and Oceanic Technology. 2009;26(10):2034-50. 157. US NASA. Data User's Guide - Data Product Descriptions - Lidar Level 2 5 km Vertical Feature Mask (VFM) Product 2016 [Available from: https://www-calipso.larc.nasa.gov/resources/calipso_users_guide/data_summaries/vfm/index.php. 158. LANCE FIRMS. NASA Near Real-Time and MCD14DL MODIS Active Fire Detections (SHP format). In: NASA, editor. 2016. 159. Giglio L, Descloitres J, Justice CO, Kaufman YJ. An enhanced contextual fire detection algorithm for MODIS. Remote sensing of environment. 2003;87(2):273-82. 160. Wooster M, Zhukov B, Oertel D. Fire radiative energy for quantitative study of biomass burning: Derivation from the BIRD experimental satellite and comparison to MODIS fire products. Remote Sensing of Environment. 2003;86(1):83-107. 161. Wooster MJ, Roberts G, Perry G, Kaufman Y. Retrieval of biomass combustion rates and totals from fire radiative power observations: FRP derivation and calibration relationships between biomass consumption and fire radiative energy release. Journal of Geophysical Research: Atmospheres. 2005;110(D24). 162. Ichoku C, Kaufman YJ. A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE transactions on Geoscience and Remote Sensing. 2005;43(11):2636-49. 163. Freeborn PH, Wooster MJ, Roberts G. Addressing the spatiotemporal sampling design of MODIS to provide estimates of the fire radiative energy emitted from Africa. Remote Sensing of Environment. 2011;115(2):475-89. 164. Freeborn PH, Wooster MJ, Roy DP, Cochrane MA. Quantification of MODIS fire radiative power (FRP) measurement uncertainty for use in satellite‐based active fire 131  characterization and biomass burning estimation. Geophysical Research Letters. 2014;41(6):1988-94. 165. Giglio L. MODIS Collection 6 Active Fire Product User’s Guide Revision A. Unpublished manuscript, Department of Geographical Sciences, University of Maryland[ftp://fuocogeogumdedu/modis/docs/MODIS_C6_Fire_User_Guide_Apdf]. 2015. 166. Kaiser J, Heil A, Andreae M, Benedetti A, Chubarova N, Jones L, et al. Biomass burning emissions estimated with a global fire assimilation system based on observed fire radiative power. Biogeosciences. 2012;9(1):527. 167. Lloyd C. Spatial data analysis: an introduction for GIS users: Oxford university press; 2010. 57 p. 168. Shepard D, editor A two-dimensional interpolation function for irregularly-spaced data. Proceedings of the 1968 23rd ACM national conference; 1968: ACM. 169. Stull RB. An introduction to boundary layer meteorology: Springer Science & Business Media; 2012. 170. Rienecker MM, Suarez MJ, Gelaro R, Todling R, Bacmeister J, Liu E, et al. MERRA: NASA's modern-era retrospective analysis for research and applications. Journal of Climate. 2011;24(14):3624-48. 171. van Etten J. gdistance: Distances and routes on geographical grids. URL http://CRAN R-project org/package= gdistance R package version. 2012:1.-4. 172. EROS Data Center. GTOPO30 Documentation Universal Resourse Locator, Global Land Information System1996 [Available from: https://lta.cr.usgs.gov/GTOPO30. 173. Liaw A, Wiener M. Classification and regression by randomForest. R news. 2002;2(3):18-22. 174. Paugam R, Wooster M, Freitas S, Val Martin M. A review of approaches to estimate wildfire plume injection height within large-scale atmospheric chemical transport models. Atmos Chem Phys. 2016;16(2):907-25. 175. Roberts G, Wooster M, Lagoudakis E. Annual and diurnal African biomass burning temporal dynamics. Biogeosciences. 2009;6(5):849-66. 176. Konovalov IB, Beekmann M, Kaiser JW, Shudyaev AA, Yurova A, Kuznetsova IN, editors. Diurnal variations of wildfire emissions in Europe: analysis of the MODIS and SEVIRI measurements in the framework of the regional scale air pollution modelling. EGU General Assembly Conference Abstracts; 2013. 132  177. Huang J, Guo J, Wang F, Liu Z, Jeong MJ, Yu H, et al. CALIPSO inferred most probable heights of global dust and smoke layers. Journal of Geophysical Research: Atmospheres. 2015;120(10):5085-100. 178. Hunt WH, Winker DM, Vaughan MA, Powell KA, Lucker PL, Weimer C. CALIPSO lidar description and performance assessment. Journal of Atmospheric and Oceanic Technology. 2009;26(7):1214-28. 179. Friedman JH. Greedy function approximation: a gradient boosting machine. Annals of statistics. 2001:1189-232. 180. Schroeder W, Ruminski M, Csiszar I, Giglio L, Prins E, Schmidt C, et al. Validation analyses of an operational fire monitoring product: The Hazard Mapping System. International Journal of Remote Sensing. 2008;29(20):6059-66. 181. Amiridis V, Marinou E, Kazadzis S, Gerasopoulos E, Mamouri R, Kokkalis P, et al. Evaluation of CALIPSO’s Aerosol Classification Scheme During the ACEMED Experimental Campaign Over Greece: The Case Study of 9th of September 2011.  Advances in Meteorology, Climatology and Atmospheric Physics: Springer; 2013. p. 865-71. 182. Brey SJ. Using operational hms smoke observations to gain insights on north american smoke transport and implications for air quality. Fort Collins, Colorado: Colorado State University; 2016. 183. Zhang T, Wooster M, editors. Small Fire Detection Algorithm Development using VIIRS 375m Imagery: Application to Agricultural Fires in Eastern China. EGU General Assembly Conference Abstracts; 2016. 184. Ainslie B, Jackson P. The use of an atmospheric dispersion model to determine influence regions in the Prince George, BC airshed from the burning of open wood waste piles. Journal of environmental management. 2009;90(8):2393-401. 185. Delong C, Griesbauer H, Mackenzie W, Foord V. Corroboration of biogeoclimatic ecosystem classification climate. Journal of Ecosystems and Management. 2010;10(3):49–64. 186. Pojar J, Klinka K, Meidinger D. Biogeoclimatic ecosystem classification in British Columbia. Forest Ecology and Management. 1987;22(1-2):119-54. 187. Wiedinmyer C, Quayle B, Geron C, Belote A, McKenzie D, Zhang X, et al. Estimating emissions from fires in North America for air quality modeling. Atmos Environ. 2006;40(19):3419-32. 188. Andreae MO, Merlet P. Emission of trace gases and aerosols from biomass burning. Global biogeochemical cycles. 2001;15(4):955-66. 133  189. Kistler M, Schmidl C, Padouvas E, Giebl H, Lohninger J, Ellinger R, et al. Odor, gaseous and PM10 emissions from small scale combustion of wood types indigenous to Central Europe. Atmos Environ. 2012;51:86-93. 190. Austin M, Buffett D, Nicolson D, Scudder G, Stevens V. Taking nature’s pulse: the status of biodiversity in British Columbia. Biodiversity BC, Victoria. British Columbia, Canada. 2008. 191. Yao J, Raffuse SM, Brauer M, Williamson GJ, Bowman DM, Johnston FH, et al. Predicting the minimum height of forest fire smoke within the atmosphere using machine learning and data from the CALIPSO satellite. Remote Sensing of Environment. 2018;206:98-106. 192. Henderson SB, Morrison KT, McLean KE, Yao J, Shaddick G, Buckeridge DL. Staying ahead of the epidemiologic curve: Demonstration of the British Columbia Asthma Prediction System (BCAPS) during a wildfire smoke event. Environmental Health Perspectives (submitted). 2018. 193. Green M, Kondragunta S, Ciren P, Xu C. Comparison of GOES and MODIS aerosol optical depth (AOD) to aerosol robotic network (AERONET) AOD and IMPROVE PM2. 5 mass at Bondville, Illinois. Journal of the Air & Waste Management Association. 2009;59(9):1082-91. 194. Liu Y, Paciorek CJ, Koutrakis P. Estimating regional spatial and temporal variability of PM2. 5 concentrations using satellite data, meteorology, and land use information. Environmental health perspectives. 2009;117(6):886. 195. Paciorek CJ, Liu Y, Moreno-Macias H, Kondragunta S. Spatiotemporal associations between GOES aerosol optical depth retrievals and ground-level PM2.5. Environmental Science & Technology. 2008;42(15):5800-6. 196. Solazzo E, Bianconi R, Pirovano G, Matthias V, Vautard R, Moran MD, et al. Operational model evaluation for particulate matter in Europe and North America in the context of AQMEII. Atmos Environ. 2012;53:75-92. 197. Tian J, Chen D. A semi-empirical model for predicting hourly ground-level fine particulate matter (PM2. 5) concentration in southern Ontario from satellite remote sensing and ground-based meteorological measurements. Remote Sensing of Environment. 2010;114(2):221-9. 198. Elangasinghe M, Singhal N, Dirks K, Salmond J, Samarasinghe S. Complex time series analysis of PM10 and PM2. 5 for a coastal site using artificial neural network modelling and k-means clustering. Atmos Environ. 2014;94:106-16. 199. Prados AI, Kondragunta S, Ciren P, Knapp KR. GOES aerosol/smoke product (GASP) over North America: comparisons to AERONET and MODIS observations. Journal of Geophysical Research: Atmospheres. 2007;112:D15201. 134  200. Zhang H, Lyapustin A, Wang Y, Kondragunta S, Laszlo I, Ciren P, et al. A multi-angle aerosol optical depth retrieval algorithm for geostationary satellite data over the United States. Atmos Chem Phys. 2011;11(23):11977. 201. Raffuse SM, McCarthy MC, Craig KJ, DeWinter JL, Jumbam LK, Fruin S, et al. High‐resolution MODIS aerosol retrieval during wildfire events in California for use in exposure assessment. Journal of Geophysical Research: Atmospheres. 2013;118(19):11-242. 202. van Donkelaar A, Martin RV, Levy RC, da Silva AM, Krzyzanowski M, Chubarova NE, et al. Satellite-based estimates of ground-level fine particulate matter during extreme events: A case study of the Moscow fires in 2010. Atmos Environ. 2011;45(34):6225-32. 203. Schroeder W, Prins E, Giglio L, Csiszar I, Schmidt C, Morisette J, et al. Validation of GOES and MODIS active fire detection products using ASTER and ETM+ data. Remote Sensing of Environment. 2008;112(5):2711-26. 204. NOAA & NASA. Advanced Baseline Imager (ABI) Fact Sheet 2017 [Available from: https://www.goes-r.gov/education/docs/Factsheet_ABI.pdf. 205. Nicodemus K, Shugart Y, editors. Impact of linkage disequilibrium and effect size on the ability of machine learning methods to detect epistasis in case-control studies. Genetic Epidemiology; 2007: WILEY-LISS DIV JOHN WILEY & SONS INC, 111 RIVER ST, HOBOKEN, NJ 07030 USA. 206. Canadian Institute for Health Information. Discharge Abstract Database (Hospital Separations). V2 ed. Population Data BC: Data Extract. MOH (2017). http://www.popdata.bc.ca/data; 2017. 207. Yao J, Brauer M, Raffuse S, Henderson SB. Machine Learning Approach To Estimate Hourly Exposure to Fine Particulate Matter for Urban, Rural, and Remote Populations during Wildfire Seasons. Environmental science & technology. 2018;52(22):13239-49. 208. Maclure M. The case-crossover design: a method for studying transient effects on the risk of acute events. American journal of epidemiology. 1991;133(2):144-53. 209. Gasparrini A, Armstrong B, Kenward MG. Distributed lag non‐linear models. Statistics in medicine. 2010;29(21):2224-34. 210. Gasparrini A, Guo Y, Hashizume M, Lavigne E, Zanobetti A, Schwartz J, et al. Mortality risk attributable to high and low ambient temperature: a multicountry observational study. The Lancet. 2015;386(9991):369-75. 211. Buteau S, Goldberg MS, Burnett RT, Gasparrini A, Valois M-F, Brophy JM, et al. Associations between ambient air pollution and daily mortality in a cohort of congestive heart 135  failure: Case-crossover and nested case-control analyses using a distributed lag nonlinear model. Environment international. 2018;113:313-24. 212. Guo Y, Barnett AG, Pan X, Yu W, Tong S. The impact of temperature on mortality in Tianjin, China: a case-crossover design with a distributed lag nonlinear model. Environmental health perspectives. 2011;119(12):1719. 213. Guo Y. Hourly associations between heat and ambulance calls. Environmental pollution. 2017;220:1424-8. 214. Youssouf H, Liousse C, Roblou L, Assamoi EM, Salonen RO, Maesano C, et al. Non-accidental health impacts of wildfire smoke. Int J Environ Res Public Health. 2014;11(11):11772-804. 215. Gasparrini A. Distributed lag linear and non-linear models in R: the package dlnm. Journal of statistical software. 2011;43(8):1. 216. Eze IC, Hemkens LG, Bucher HC, Hoffmann B, Schindler C, Künzli N, et al. Association between ambient air pollution and diabetes mellitus in Europe and North America: systematic review and meta-analysis. Environmental health perspectives. 2015;123(5):381-9. 217. Pearson JF, Bachireddy C, Shyamprasad S, Goldfine AB, Brownstein JS. Association between fine particulate matter and diabetes prevalence in the US. Diabetes care. 2010;33(10):2196-201. 218. Kerr D, Holden D, Smith J, Kelly A-M, Bunker S. Predictors of ambulance use in patients with acute myocardial infarction in Australia. Emergency medicine journal. 2006;23(12):948-52. 219. Rucker DW, Edwards RA, Burstin HR, O'neil AC, Brennan TA. Patient-specific predictors of ambulance use. Annals of emergency medicine. 1997;29(4):484-91. 220. Marsha A, Larkin NK. A statistical model for predicting PM2. 5 for the western United States. Journal of the Air & Waste Management Association. 2019(just-accepted). 221. Yao J, Roberecki S, DuTeaux S, Rideout K. Evidence Review: Exposure measures for wildfire smoke surveillance. 2014. 222. Lavrov A, Utkin AB, Vilar R, Fernandes A. Evaluation of smoke dispersion from forest fire plumes using lidar experiments and modelling. Int J Therm Sci. 2006;45(9):848-59. 223. Mei L, Xue Y, de Leeuw G, Guang J, Wang Y, Li Y, et al. Integration of remote sensing data and surface observations to estimate the impact of the Russian wildfires over Europe and Asia during August 2010. Biogeosciences. 2011;8(12):3771-91. 136  224. Yuchi W, Yao J, McLean KE, Stull R, Pavlovic R, Davignon D, et al. Blending forest fire smoke forecasts with observed data can improve their utility for public health applications. Atmos Environ. 2016;145:308-17. 225. Fuentes M, Raftery AE. Model evaluation and spatial interpolation by Bayesian combination of observations with outputs from numerical models. Biometrics. 2005;61(1):36-45. 226. McMillan NJ, Holland DM, Morara M, Feng J. Combining numerical model output and particulate data using Bayesian space–time modeling. Environmetrics: The official journal of the International Environmetrics Society. 2010;21(1):48-65. 227. O'malley KJ, Cook KF, Price MD, Wildes KR, Hurdle JF, Ashton CM. Measuring diagnoses: ICD code accuracy. Health services research. 2005;40(5p2):1620-39. 228. Surján G. Questions on validity of International Classification of Diseases-coded diagnoses. International Journal of Medical Informatics. 1999;54(2):77-95. 229. Lucyk K, Lu M, Sajobi T, Quan H. Administrative health data in Canada: lessons from history. BMC medical informatics and decision making. 2015;15(1):69. 230. Bhaskaran K, Wilkinson P, Smeeth L. Cardiovascular consequences of air pollution: what are the mechanisms? Heart. 2011;97(7):519-20. 231. Almon S. The distributed lag between capital appropriations and expenditures. Econometrica: Journal of the Econometric Society. 1965:178-96. 232. California Department of Health. Wildfire Smoke, A Guide for Public Health Officials. 2008. 233. Manitoba Health. Smoke exposure from wildland fires -- interim guidelines for protecting commnunity health and wellbeing. 2012. 234. Franklin P GM. Bushfires and other vegetative fires: protecting community health and well being from smoke exposure. In: Department WAH, editor. 2012. 235. Elliott CT. Guidance for BC Public Health Decision Makers During Wildfire Smoke Events. In: BC Centre for Disease Control, editor. 2014. 236. Tanaka HL, Kanetaka M. Realtime Prediction System of Forest-Fire Smoke using Satellite Data and the PUFF Model: A Case Study for May 2003. Sola. 2005;1:9-12. 237. Barn P, Kosatsky T. Introduction to air quality advisories. National Collaborating Centre for Envrionmental Health; 2010. 137  238. Stieb DM, Burnett RT, Smith-Doiron M, Brion O, Shin HH, Economou V. A new multipollutant, no-threshold air quality health index based on short-term associations observed in daily time-series analyses. Journal of the Air & Waste Management Association. 2008;58(3):435-50. 239. Yao J, Stieb DM, Taylor E, Henderson SB. Assessment of the Air Quality Health Index (AQHI) and four alternate AQHI-Plus amendments for wildfire seasons in British Columbia. Canadian Journal of Public Health. 2019:1-11. 240. Johnston F, Hanigan I, Henderson S, Morgan G, Bowman D. Extreme air pollution events from bushfires and dust storms and their association with mortality in Sydney, Australia 1994-2007. Environ Res. 2011;111(6):811-6. 241. Johnston F, Wheeler A, Williamson G, Campbell S, Jones P, Koolhof I, et al. Using smartphone technology to reduce health impacts from atmospheric environmental hazards. Environmental Research Letters. 2018;13(4):044019. 242. Prince S, Wei L, Corrigan A, Rappazzo K, Baghdikian C, Hubbell B, et al., editors. Symptoms and Behaviors Related to Wildland Fire Smoke Exposure: Data from the EPA Smoke Sense Citizen Science Smartphone Application. ISEE Conference Abstracts; 2018. 243. Liu Y, Diner DJ. Multi-angle imager for aerosols: a satellite investigation to benefit public health. Public Health Reports. 2017;132(1):14-7. 244. Canadian Space Agency. Canadian Space Agency Space Utilization -- WildFireSat Statement of Work. 2019.     138  Appendices Appendix A  Procedure for calculating summed fire radiative power (FRP) in each 5km grid cell, adjusting for the scan angle and bow-tie effects. 1. (Giglio 2015) Calculate scan angle (θ) from sample number for each fire detection: θ = s × (sample number − 676.5) where s = 0.0014184397  2. (Giglio 2015) Calculate pixel area (A) for each fire detection: ∆S = Res (cos θ√(𝑅𝑒/r)2 − sin2 θ  – 1)   ∆T = rs (cos θ − √(𝑅𝑒/r)2  − sin2 θ ),  A = ∆S × ∆T  where Re = 6378.137 km (Earth radius), r = Re + h, h = 705 km (satellite altitude), s = 0.0014184397, and θ is the scan angle  3. (Kaiser et al. 2012) For each 5-km grid cell j, within which there were i observed FRP (Fi), and the total observed FRP <F>j  and total satellite observed pixel area <A>j  can be expressed as <F>j  =  ∑  𝐹𝑖 𝑐𝑜𝑠2𝜃𝑖𝑖∈𝑗∑  𝑐𝑜𝑠2𝜃𝑖𝑖∈𝑗  <A>j  =  ∑  𝐴𝑖 𝑐𝑜𝑠2𝜃𝑖𝑖∈𝑗∑  𝑐𝑜𝑠2𝜃𝑖𝑖∈𝑗  4. (Kaiser et al. 2012) The adjusted FRP areal density for grid cell j is calculated as  <D>j = <𝐹>𝑗<𝐴>𝑗  Giglio, L. (2015). MODIS Collection 6 Active Fire Product User’s Guide Revision A.Unpublishedmanuscript, Department of Geographical Sciences, University ofMaryland.[ftp://fuoco.geog.umd.edu/modis/docs/MODIS_C6_Fire_User_Guide_A.pdf] Kaiser, J., Heil, A., Andreae, M., Benedetti, A., Chubarova, N., Jones, L., Morcrette, J.-J., Razinger, M.,Schultz, M., & Suttie, M. (2012). Biomass burning emissions estimated with a global fire assimilationsystem based on observed fire radiative power.Biogeosciences, 9, 527  139  Appendix B  Codes for ambulance dispatches and paramedic assessments. B.1 Medical Priority Dispatch System (MPDS) codes 1 Abdominal Pain 12 Convulsions / Seizures 23 Overdose / Poisoning (Ingestion) 2 Allergies (Reactions) / (Stings, Bites) 13 Diabetic Problems 24 Pregnancy / Childbirth / Miscarriage 3 Animal Bites / Attacks 14 Drowning (Near) / Diving / SCUBA Accident 25 Psychiatric / Abnormal Behavior / Suicide Attempt 4 Assault / Sexual Assault 15 Electrocution / Lightning 26 Sick Person (specific diagnosis) 5 Back Pain (Non-Traumatic or Non Recent Trauma) 16 Eye Problems / Injuries 27 Stab / Gunshot / Penetrating Trauma 6 Breathing Problems 17 Falls 28 Stroke (CVA) 7 Burns (Scalds) / Explosions (Blasts) 18 Headache 29 Traffic / Transportation Incidents 8 Carbon Monoxide / Inhalation / HAZMAT / CBRN 19 Heart Problems / AICD 30 Traumatic Injuries (Specific) 9 Cardiac or Respiratory Arrest / Death 20 Heat / Cold Exposure 31 Unconscious / Fainting (Near) 10 Chest Pain (Non Traumatic) 21 Hemorrhage / Lacerations 32 Unknown Problem (man down) 11 Choking 22 Inaccessible Incident / Other Entrapment (Non-Veh) 33 Transfer / Interfacility / Palliative Care    140  B.2 Paramedic impression codes Allergy Genitourinary System (GU) 1722 Isolated Spinal Fracture  Respiratory System 1706 Allergic Reaction - Anaphylaxis 1125 Renal Colic   without Deficits 0905 Near Drowning 1754  Allergic Reaction - Sensitivity 1130 Urinary retention 1723 Isolated laceration 0910 Upper Airway Obstruction  Circulatory System 1135  Urinary Tract Infection 1719 Isolated sprain 0915 Pneumothorax (spontaneous) 0805 Suspected Cardiac Ischemia 1115 GU/Renal-Other 1711 Isolated dislocation 0925 Respiratory Failure 0815 Other Shock Gynecologic/Pregnancy/Childbirth 1724 Isolated Head Trauma 0930 Asthma/COPD 0825 Cardiac Arrest:Treated 1105 GYN - Vaginal Hemorrhage 1715 Isolated fracture 0935 Acute Bronchitis 0830 Cardiac Rhythm Disturbance 1205 Obstetrical - Childbirth/Labour 1738 Other Fractures 0955 Respiratory Arrest 0835 Hypovolemia 1210 Obstetrical - Complication of 1735 Peripheral Injury (limbs only) 0970 Mechanical Airway Obstruction 0840 Cardiogenic Shock  Pregnancy 1740 Soft Tissue Injury 0975 Infectious Pneumonia 0845 Aortic Aneurysm 1220 Obstetrical - Complication  1710 Sexual Assault 0940 Respiratory Distress - Other 0850 Congested Heart Failure (CHF)  of Labour 1716 Electrocution  Psychiatric / Behavioral 0855 Peripheral Vascular Disease 1225 Eclampsia 1717 Hanging 0510 Depression 0860 Acute Ml 1230 Post partum Hemorrhage 1714 Injury-Other 0515 Psychosis 0880 Pericardial tamponade 1110 GYN - Other Neonate 0520 Agitated Delirium 0885 Chest Pain NYD Infectious Disease 1215 Neonate - Uncomplicated 0505 Psychiatric/Behavioral - Other 0890 Pulmonary Embolus 0100 Infectious Disease 1605 Neonate-Complication of Birth Non Specific Symptoms 1830 Cardiac Arrest: Untreated 0105 Localized Infection Neoplasm/Cancer 1815 Sick/Illness NYD 0820 Circulatory System - Other 0115 Sepsis 0200 Neoplasm / Cancer 1820 Abnormal Vital Signs NYD Digestive System (Gl) 0120 Septic Shock Nervous System 1845 Nausea/Vomiting NYD 1005 Gl - Bleeding  Injury 0615 CVA 1850 Abdominal Pain NYD 1020 Diarrhea of unknown cause 1712 Burn - Minor 0620 Altered Level of Consciousness 1855 Weakness NYD 1060 Foreign body: Esophagus 1744 Burn - Major  NYD 1860 Headache NYD 1050 Gl-Other 1753 Traumatic cardiac arrest: treated 0625 Intracerebral Hemorrhage 1865 Back Pain NYD Endocrine 1755 Traumatic cardiac arrest: untreated 0630 Sub Arachnoid Hemorrhage 1870 Other Pain NYD 0305 Hyperglycemia 1718 Multitrauma 0655 Seizure 1880 Limb Pain NYD 0315 Hypoglycemia 1736 Abdominal /Chest Trauma - 0660 Syncope / Near Syncope / 1885 Malaise/Fatigue NYD 0310 Endocrine-Other  Blunt  Vertigo 1890 Rash/Itching NYD ENT/Eyes/Dental 1750 Abdominal /Chest Trauma  0665 Peripheral Nervous System Miscellaneous 1905 Pharyngitis  -Penetrating   Disorder 1805 Medical Device Problem 1910 Foreign Body: throat 1721 Amputation above wrist or 0610 Nervous System - Other 1810 Public Assist 1915 Foreign Body: ear  ankle Poisoning / Overdose 1811 Post Operative Pain 1920 Foreign Body: eye 1725 Amputation foot or hand 1746 Alcohol Intoxication 1812 Post Operative Bleeding 1925 Ear Pain  (partial or full) 1747 Alcohol Withdrawal 1875 Medication Request 1930 Isolated Eye trauma 1758 Bleeding - Controlled 1780 Recreational Drug Overdose 1800 Other 1935 Dental Pain 1760 Bleeding - Uncontrolled 1785 Recreational Drug Withdrawal   1940 ENT/Eye/Dental-Other 1732 Spinal Cord Injury with deficits 1765 Medication Overdose   Environmental 1752 Spinal Cord Injury without 1770 Medication Reaction   1728 Exposure - Smoke Inhalation  deficits 1742 Sting / Envenomation   1835 Hyperthermia 1720 Brain Injury 1790 Toxic Inhalation   1840 Hypothermia 1730 Pneumothorax 1708 Other toxic Ingestion   1704 Environmental - Other 1734 Pelvis Fracture 1775 Poisoning / Overdose - Other     

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0383432/manifest

Comment

Related Items