"Applied Science, Faculty of"@en . "Biomedical Engineering, School of"@en . "DSpace"@en . "UBCV"@en . "Kheirkhah Dehkordi, Parastoo"@en . "2018-07-26T16:07:38Z"@* . "2018"@en . "Doctor of Philosophy - PhD"@en . "University of British Columbia"@en . "We developed novel algorithms for monitoring sleep, sleep breathing disorder (SBD)\r\nand instantaneous respiratory rate (IRR) in children using the characterization of\r\npulse oximetry photoplethysmogram (PPG). To evaluate the algorithms, we recorded\r\nthe oxygen saturation (SpO\u00E2\u0082\u0082) and PPG signals from 160 children using a phone-based\r\noximeter consisting of a microcontroller-based pulse oximeter module interfacing\r\na smartphone. This mobile oximeter was further developed to perform all\r\nprocessing on the smartphone through the audio interface.\r\nWe evaluated the relative impact of SBD on sympathetic and parasympathetic\r\nactivity in children through the characterization of PPG and concluded that sympathetic\r\nactivity was higher in 30-second epochs with apnea/hypopnea event(s). We\r\nlater characterized the SpO\u00E2\u0082\u0082 pattern in SDB and then combined SpO\u00E2\u0082\u0082 pattern characterization\r\nand PPG analysis to design a model with two binary logistic classifiers\r\nto identify the epochs with apnea/hypopnea events.\r\nWe developed a novel model for identifying the cycles of random eye movement\r\n(REM) and non-REM of the overnight sleep based on the activity of cardiorespiratory\r\nsystem using the overnight PPG. We extracted the features associated with\r\npulse rate variability (PRV), respiratory rate (RR), vascular tone and movement\r\nfrom PPG to build a model with two binary classifiers to identify wakefulness from\r\nsleep (wake/sleep classifier) and REM from non-REM sleep (non-REM/REM classifier).\r\nWe also developed a novel algorithm for extracting the instantaneous respiratory\r\nrate (IRR) from PPG. The algorithm was performed in three steps: extraction\r\nof three respiratory-induced variation signals from PPG, estimation of IRR from\r\neach extracted respiratory-induced variation signal and fusion of IRR estimates. A time-frequency transform called synchrosqueezing transform (SST) was used\r\nto extract the respiratory-induced variation signals from PPG. Later, a second SST\r\nwas applied to estimate IRR from respiratory-induced variation signals. To fuse\r\nIRR estimates, a novel algorithm was proposed.\r\nThis study would expand the functionality of conventional pulse oximetry beyond\r\nthe measurement of heart rate and SpO\u00E2\u0082\u0082 to monitor sleep, to screen SBDs and\r\nmeasure the respiratory rate continuously and instantly."@en . "https://circle.library.ubc.ca/rest/handle/2429/66591?expand=metadata"@en . "Monitoring Sleep and Sleep Breathing Disorders UsingPulse Oximeter PhotoplethysmogrambyParastoo Kheirkhah DehkordiM.A. Sc., Simon Fraser University, 2012M.A. Sc., Tehran Azad University , 2005B.A. Sc., University of Isfahan, 2001A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Biomedical Engineering)The University of British Columbia(Vancouver)July 2018c\u00C2\u00A9 Parastoo Kheirkhah Dehkordi, 2018The following individuals certify that they have read, and recommend to the Fac-ulty of Graduate and Postdoctoral Studies for acceptance, the thesis entitled:Monitoring Sleep and Sleep Breathing Disorders Using Pulse OximeterPhotoplethysmogramsubmitted by Parastoo Kheirkhah Dehkordi in partial fulfillment of the require-ments for the degree of Doctor of Philosophy in Biomedical Engineering.Examining Committee:Guy A Dumont, Biomedical EngineeringSupervisorJ Mark Ansermino, Anesthesia, Pharmacology and TherapeuticsCo-supervisorJohn Fleetham, Experimental MedicineUniversity ExaminerPurang Abolmaesumi, Electrical and Computer EngineeringUniversity ExaminerAdditional Supervisory Committee Members:Matthew J Yedlin, Electrical and Computer EngineeringSupervisory Committee MemberiiAbstractWe developed novel algorithms for monitoring sleep, sleep breathing disorder (SBD)and instantaneous respiratory rate (IRR) in children using the characterization ofpulse oximetry photoplethysmogram (PPG). To evaluate the algorithms, we recordedthe oxygen saturation (SpO2) and PPG signals from 160 children using a phone-based oximeter consisting of a microcontroller-based pulse oximeter module inter-facing a smartphone. This mobile oximeter was further developed to perform allprocessing on the smartphone through the audio interface.We evaluated the relative impact of SBD on sympathetic and parasympatheticactivity in children through the characterization of PPG and concluded that sympa-thetic activity was higher in 30-second epochs with apnea/hypopnea event(s). Welater characterized the SpO2 pattern in SDB and then combined SpO2 pattern char-acterization and PPG analysis to design a model with two binary logistic classifiersto identify the epochs with apnea/hypopnea events.We developed a novel model for identifying the cycles of random eye move-ment (REM) and non-REM of the overnight sleep based on the activity of cardiores-piratory system using the overnight PPG. We extracted the features associated withpulse rate variability (PRV), respiratory rate (RR), vascular tone and movementfrom PPG to build a model with two binary classifiers to identify wakefulness fromsleep (wake/sleep classifier) and REM from non-REM sleep (non-REM/REM clas-sifier).We also developed a novel algorithm for extracting the instantaneous respira-tory rate (IRR) from PPG. The algorithm was performed in three steps: extractionof three respiratory-induced variation signals from PPG, estimation of IRR fromeach extracted respiratory-induced variation signal and fusion of IRR estimates.iiiA time-frequency transform called synchrosqueezing transform (SST) was used toextract the respiratory-induced variation signals from PPG. Later, a second SSTwas applied to estimate IRR from respiratory-induced variation signals. To fuseIRR estimates, a novel algorithm was proposed.This study would expand the functionality of conventional pulse oximetry be-yond the measurement of heart rate and SpO2 to monitor sleep, to screen SBDs andmeasure the respiratory rate continuously and instantly.ivLay SummarySleep apnea is defined as a short pause of breathing during a normal overnightsleep. Each pause typically lasts from 10 to 90 seconds. In children with sleep ap-nea syndrome, these pauses can be repeated several times during a night, resultingin a low level of oxygen in the blood and a low quality of sleep. Untreated sleepapnea in children can be linked to impairments in memory, attention, learning,and behavior. Polysomnography, also called a sleep study, is a common test usedto diagnose sleep apnea syndrome. Polysomnography is a very complicated andexpensive test and requires an overnight stay at a very equipped sleep laboratory.In this study we designed and developed a simple and low cost mobile tech-nology for screening sleep and sleep apnea in children using a pulse oximeter con-nected to a smartphone.vPrefaceChapter 2 is based on the work conducted in UBC\u00E2\u0080\u0099s Electrical & Computer En-gineering for Medicine (ECEM) group by Parastoo Dehkordi, Dr. Ainara Garde,Dr. J. Mark Ansermino, and Dr. Guy A. Dumont. I was responsible for develop-ing an algorithm for extracting pulse rate variability from the Phone oximeterTMphotoplethysmogram (PPG) to assess the cardiac modulation in children with sleepbreathing disorders. Some aspects of Chapters 2 have been published:-Dehkordi P et al., \u00E2\u0080\u009DEvaluation of Cardiac Modulation in Children in Re-sponse to Apnea/Hypopnea using the Phone Oximeter,\u00E2\u0080\u009D Physiological Measure-ment. 2016. 37(2): 187-202.-Dehkordi P et al., \u00E2\u0080\u009DPulse rate variability in children with sleep disorderedbreathing in different sleep stages,\u00E2\u0080\u009D Proceeding of International Conference ofComputing in Cardiology. 2015: 1015-18.-Dehkordi P et al., \u00E2\u0080\u009DDetrended Fluctuation Analysis of PhotoplethysmogramPulse Rate Intervals in Sleep Disordered Breathing,\u00E2\u0080\u009D Conference Proceeding ofIEEE Health Innovations and Point-of-Care Technologies. 2014: 323-6.-Dehkordi P et al., \u00E2\u0080\u009DPulse Rate Variability Compared with Heart Rate Vari-ability in Children with and without Sleep Disordered Breathing,\u00E2\u0080\u009D Conference Pro-ceeding of IEEE Engineering in Medicine and Biology Society. 2013: 6563-6.I performed all of the data analysis and wrote the manuscripts.Chapter 3 is based on the work conducted in UBC\u00E2\u0080\u0099s Electrical & ComputerEngineering for Medicine (ECEM) group by Parastoo Dehkordi, Dr. Ainara Garde,Dr. Behnam Molavi, Dr. J. Mark Ansermino, and Dr. Guy A. Dumont. I wasresponsible for proposing and developing an algorithm for extracting instantaneousrespiratory rate from the Phone oximeterTM PPG. Some aspects of Chapters 3 havevibeen published:-Dehkordi P et al., \u00E2\u0080\u009DEstimating Instantaneous Respiratory Rate from the PhoneOximeter Photoplethysmogram using the Synchrosqueezing transform\u00E2\u0080\u009D, Acceptedto the Frontier in physiology - Computational Physiology and Medicine journal.-Dehkordi P et al., \u00E2\u0080\u009DEstimating Instantaneous Respiratory Rate from the Pho-toplethysmogram\u00E2\u0080\u009D, Proceeding of IEEE Engineering in Medicine and Biology So-ciety Conference. 2015: 6150-3.I performed all of the data analysis and wrote the manuscripts.Chapter 4 is based on the work conducted in UBC\u00E2\u0080\u0099s Electrical & ComputerEngineering for Medicine (ECEM) group by Parastoo Dehkordi, Dr. Ainara Garde,Dr. J. Mark Ansermino and Dr. Guy A. Dumont. I proposed to monitor overnightsleep using the PPG recordings. I was responsible for designing and developinga model for classifying different sleep stages using the features extracted from thePhone oximeterTM PPG. Some aspects of Chapters 4 have been published:-Dehkordi P et al., \u00E2\u0080\u009DExtracting Paediatric Hypnogram from Photoplethysmo-gram,\u00E2\u0080\u009D (in preparation).-Dehkordi P et al., \u00E2\u0080\u009DSleep/Wake Classification Using Cardiorespiratory Fea-tures Extracted from Photoplethysmogram,\u00E2\u0080\u009D Proceeding of International Confer-ence of Computing in Cardiology. 2016: 294-147.-Dehkordi, P et al., \u00E2\u0080\u009DSleep-Wake Classification Using PhotoplethysmogramPulse Interval Variability,\u00E2\u0080\u009D Proceeding of International Conference of Computingin Cardiology. 2014: 297-300.I performed all of the data analysis and wrote the manuscripts.Chapter 5 is based on the work conducted in UBC\u00E2\u0080\u0099s Electrical & ComputerEngineering for Medicine (ECEM) group and BC children\u00E2\u0080\u0099s hospital by ParastooDehkordi, Dr. Ainara Garde, Dr. J. Mark Ansermino and Dr. Guy A. Dumont.I was responsible for designing and developing a model for identifying the sleepapnea/hypopnea events in children using the features of PPG and SpO2. I wasresponsible for extracting and analysing the PPG features and Dr. Ainara Gardewas responsible for extracting and analysing the SpO2 features. Some aspects ofChapters 5 have been published:-Dehkordi, P et al., \u00E2\u0080\u009DScreening Sleep and Sleep Disordered Breathing in Chil-dren Using the Phone Oximeter,\u00E2\u0080\u009D (in preparation).vii-Garde, A., Dehkordi, P., et al., \u00E2\u0080\u009DDevelopment of a Screening Tool for SleepDisordered Breathing in Children Using the Phone Oximeter,\u00E2\u0080\u009D PloS One. 2014:9(11).-Garde, A., Dehkordi, P., et al., \u00E2\u0080\u009DIdentifying individual Sleep Apnea/HypopneaEpochs Using Smartphone-based Pulse Oximetry,\u00E2\u0080\u009D Proceeding of IEEE Engineer-ing in Medicine and Biology Society Conference. 2016: 3195-98.To evaluate the proposed algorithms, we designed and conducted a study torecorded the SpO2 and PPG signals from 160 children using the Phone OximeterTMin the standard setting of overnight polysomnography (PSG) in BC Children\u00E2\u0080\u0099shospital in Vancouver. This research was approved by the University of BritishColumbia and Children\u00E2\u0080\u0099s & Womens Health Centre of British Columbia ResearchEthics Board (H11-01769). I was responsible for part of designing and running thestudy, part of recruiting patients and collecting data and the all of the post hoc datacleaning and storing. I developed an algorithm for detecting the PPG pulse peakand PPG segmentation and also an algorithm to automatically reject the segmentsof PPG contaminated by noise or motion artifact.viiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Sleep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Sleep Breathing Disorder . . . . . . . . . . . . . . . . . . . . . . 21.3 Pulse Oximetry . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.4 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.5 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.6 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.7 Organization of the Dissertation . . . . . . . . . . . . . . . . . . 13ix2 Evaluation of Cardiac Modulation in Children in Response to Ap-nea/Hypopnea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2 Materials and methods . . . . . . . . . . . . . . . . . . . . . . . 172.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . 172.2.2 Data collection . . . . . . . . . . . . . . . . . . . . . . . 172.2.3 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . 182.2.4 Sleep and apnea analysis . . . . . . . . . . . . . . . . . . 192.2.5 Parameter extraction . . . . . . . . . . . . . . . . . . . . 192.2.6 Data analysis . . . . . . . . . . . . . . . . . . . . . . . . 212.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3.1 Intra-individual event analysis . . . . . . . . . . . . . . . 232.3.2 Inter-groups analysis . . . . . . . . . . . . . . . . . . . . 232.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.4.1 Intra-individual event analysis . . . . . . . . . . . . . . . 282.4.2 Inter-groups analysis . . . . . . . . . . . . . . . . . . . . 302.4.3 Limitations and future work . . . . . . . . . . . . . . . . 342.4.4 Clinical relevance . . . . . . . . . . . . . . . . . . . . . . 343 Extracting Instantaneous Respiratory Rate from Photoplethysmo-gram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . 373.2 Algorithm Description . . . . . . . . . . . . . . . . . . . . . . . 393.2.1 Instantaneous Frequency (IF) . . . . . . . . . . . . . . . 393.2.2 Synchrosqueezing Transform (SST) . . . . . . . . . . . . 403.3 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . 413.3.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . 413.3.2 Estimation of IRR from PPG . . . . . . . . . . . . . . . . 423.3.3 Algorithm Evaluation . . . . . . . . . . . . . . . . . . . . 473.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.4.1 Capnobase data base . . . . . . . . . . . . . . . . . . . . 483.4.2 Sleep database . . . . . . . . . . . . . . . . . . . . . . . 50x3.5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 524 Extracting the Pediatric Hypnogram from Photoplethysmogram . . 554.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574.2.1 Pulse Rate Variability . . . . . . . . . . . . . . . . . . . . 574.2.2 Vascular tone . . . . . . . . . . . . . . . . . . . . . . . . 574.2.3 Respiratory rate . . . . . . . . . . . . . . . . . . . . . . . 574.2.4 Movement . . . . . . . . . . . . . . . . . . . . . . . . . 584.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 584.3.1 PPG Preprocessing . . . . . . . . . . . . . . . . . . . . . 584.3.2 Sleep Labelling . . . . . . . . . . . . . . . . . . . . . . . 584.3.3 Feature extraction . . . . . . . . . . . . . . . . . . . . . . 594.4 Statistical Learning . . . . . . . . . . . . . . . . . . . . . . . . . 624.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 624.4.2 Multivariate model development and validation . . . . . . 664.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.5.1 Wake/sleep classifier . . . . . . . . . . . . . . . . . . . . 684.5.2 non-REM/REM classifier . . . . . . . . . . . . . . . . . 694.6 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 714.6.1 Limitation of study and future work . . . . . . . . . . . . 735 Development of a Monitoring Tool for Sleep Disordered Breathingin Children Using the Phone Oximeter . . . . . . . . . . . . . . . . . 745.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 765.2.1 Apnea/Hypopnea Labelling . . . . . . . . . . . . . . . . 765.2.2 PPG Features Extraction . . . . . . . . . . . . . . . . . . 775.2.3 SpO2 Features Extraction . . . . . . . . . . . . . . . . . . 795.2.4 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . 825.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855.3.1 Univariate Analysis . . . . . . . . . . . . . . . . . . . . . 855.3.2 Multivariate Model Validation . . . . . . . . . . . . . . . 85xi5.4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 876 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . 936.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946.1.1 Sleep Solution . . . . . . . . . . . . . . . . . . . . . . . 946.2 Limitation of the research . . . . . . . . . . . . . . . . . . . . . . 94Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97xiiList of TablesTable 1.1 Rules for scoring obstructive apnea/hypopnea events in adultsand children defined by American Academy of Sleep Medicine(AASM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5Table 2.1 Demographics and AHI index of studied population expressedas mean \u00C2\u00B1 standard deviation. In this table: 1Rapid Eye Move-ment; 2Body Mass Index; 3Total Sleep Time; 4Total Bed Time;\u00E2\u0088\u0097p < 0.001; \u00E2\u0088\u0097\u00E2\u0088\u0097p < 0.0001 comparing SDB and non-SDB; \u00E2\u0080\u00A0p-value < 0.001 comparing AHI in REM and non-REM sleep stages 18Table 2.2 Descriptive results (median) of estimated parameters for chil-dren with and without SDB during the entire sleep period . . . 27Table 2.3 Descriptive results (median) of estimated parameters for chil-dren with and without SDB during non-REM sleep . . . . . . . 27Table 2.4 Descriptive results (median) of estimated parameters for chil-dren with and without SDB during REM sleep . . . . . . . . . 28Table 3.1 The performance of different method for estimation IRR fromPPG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51Table 4.1 Description of the features extracted from PPG . . . . . . . . . 63Table 4.2 The performance of a binary classifier is summarized by a 2 \u00C3\u00972 confusion matrix for a given decision threshold \u00CF\u0084 . TN: num-ber of true negative, FP: number of false positive, FN: numberof false negative, TP: number of true positive, PN: number ofpredicted negative, PP: number of predicted positive . . . . . . 66xiiiTable 4.3 Estimated coefficient and error for 15 features selected withLASSO as the significant features for wake/sleep model . . . . 69Table 4.4 Estimated coefficient and error for 16 features selected withLASSO as the significant features for non-REM/REM model . 71Table 5.1 Description of the features extracted from PPG . . . . . . . . . 81Table 5.2 Description of the features extracted from SpO2 . . . . . . . . 82Table 5.3 Distribution of features extracted from PPG and SpO2 for A/Hand non-A/H epochs . . . . . . . . . . . . . . . . . . . . . . . 86Table 5.4 Distribution of features extracted from SpO2 for A/H and non-A/H epochs . . . . . . . . . . . . . . . . . . . . . . . . . . . 87Table 5.5 Estimated coefficient and error for 12 features selected withLASSO as the significant features for A/H model trained overdatabase1 (including PPG and SpO2 features) . . . . . . . . . 88Table 5.6 Estimated coefficient and error for 5 features selected with LASSOas the significant features for A/H model trained over database2(including SpO2 features) . . . . . . . . . . . . . . . . . . . . 88Table 5.7 Classification results from test set represented by the mean and95% CI of the quartiles of Accuracy(Acc), Sensitivity (Sn),Specificity(Sp)and the area of the ROC curve (AUC) . . . . . . 89xivList of FiguresFigure 1.1 Hypnogram, provided by PSG, depicts the structure of an overnightsleep, organization of cycles and timing of different sleep stages. 3Figure 1.2 A conventional transmitted pulse oximeter sensor has two light-emitting diodes (LEDs) and a light detector mounted on theopposite side of the LEDs (inspired by [75]) . . . . . . . . . . 8Figure 1.3 Light transmitted through the living tissue (inspired by [75]) . 8Figure 1.4 Absorption spectra of hemoglobin (from [75]) . . . . . . . . . 9Figure 1.5 The Phone OximeterTM interfacing a microcontroller-based pulseoximeter module with a smartphone. . . . . . . . . . . . . . . 10Figure 1.6 (a) Sensor and (b) the user interface of iOS App of Kenek O2Pulse Oximeter . . . . . . . . . . . . . . . . . . . . . . . . . 11Figure 1.7 Organization of the Dissertation. . . . . . . . . . . . . . . . . 14Figure 2.1 In double logarithmic plot, the fluctuation function of PPIs,F(n), is plotted as a function of n (the number of pulses) for achild without SDB during non-REM (blue squares) and REMsleep (red stars). The slopes of the curves correspond to thefluctuation scaling exponent \u00CE\u00B1 . For n > 100, the fluctuationfunction of PPI during REM and non-REM are distinguishable. 22xvFigure 2.2 Comparison of spectral parameters in segments with and with-out apnea/hypopnea events for children with SDB (AHI > 5)during the entire period of sleep. Blue (thin) and red (thick)lines show the mean increase and decrease of parameters, re-spectively. The nLF parameter increased in apnea/hypopneaevents for 96% of the children with SDB (Blue lines). TheLF/HF ratio increased in apnea/hypopnea events for 96% ofthe children with SDB (Blue lines), while nHF decreased in94% of children with SDB during apnea/hypopnea events (Bluelines). The VLF parameter increased during apnea/hypopneaevents for almost 92% of the children with SDB (Blue lines). . 24Figure 2.3 Comparison of spectral parameters in segments with and with-out apnea/hypopnea events for children with SDB (AHI > 5)during the non-REM sleep. Blue (thin) and red (thick) linesshow the mean increase and decrease of parameters, respec-tively. For 95% of children with SDB, higher nLF, higherLF/HF ratio, and lower nHF were recognized in segments withapnea/hypopnea events compared to segments without SDB.The VLF parameter increased during apnea/hypopnea eventsfor almost 90% of the children with SDB. . . . . . . . . . . . 25Figure 2.4 Comparison of spectral parameters in segments with and with-out apnea/hypopnea events for children with SDB (AHI > 5)during the REM sleep. Blue (thin) and red (thick) lines showthe mean increase and decrease of parameters respectively. Dur-ing REM sleep, for 73% of the children with SDB, the nLFand LF/HF ratio increased in apnea/hypopnea events. In addi-tion, for 68% of the children with SDB, nHF decreased in theapnea/hypopnea events. The VLF parameter increased duringapnea/hypopnea events for almost 90% of the children with SDB. 26xviFigure 2.5 Frequency domain parameters in children with and withoutSDB during (a) the entire sleep period, (b) non-REM sleepand (c) REM sleep. Significant differences between the SDBand non-SDB groups are marked by one star (*) when p-value< 0.05 and by two stars (**) when p-value < 0.01. Quartilevalues are displayed as the bottom, middle and top horizon-tal line of the boxes. Whiskers are used to represent the mostextreme values within 1.5 times the interquartile range fromthe median. Outliers (data with values beyond the ends of thewhiskers) are displayed as (+). . . . . . . . . . . . . . . . . . 29Figure 2.6 The fluctuation function F(n) during non-REM sleep for chil-dren with SDB children (blue squares) and non-SDB children(red stars) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30Figure 2.7 The fluctuation function F(n) during REM sleep for childrenwith SDB children (blue squares) and non-SDB children (redstars) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31Figure 2.8 \u00CE\u00B1S and \u00CE\u00B1L in children with and without SDB during (a) entiresleep period, (b) non-REM sleep and (c) REM sleep. Signifi-cant differences between SDB and non-SDB group are repre-sented by one star (*) when p-value < 0.05 and by two stars(**) when p-value < 0.01. . . . . . . . . . . . . . . . . . . . 32Figure 3.1 From top, PPG with no modulation, Respiratory-Induced In-tensity Variation (RIIV), Respiratory-Induced Amplitude Vari-ation (RIAV), and Respiratory-Induced Frequency Variation(RIFV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38Figure 3.2 To extract IRR from PPG, the first SST was applied to PPGto extract RIIV, RIAV and RIFV. Later, the second SST wasperformed to estimate IIR from a respiratory-induced variationsignals. The peak-conditioned fusion algorithm was then usedto fuse simultaneous IRR estimates . . . . . . . . . . . . . . 43xviiFigure 3.3 In the STT surface of PPG, two components are identified: astrong cardiac component in the cardiac band (0.5-3 Hz, 30-180 beats/minute) and a respiratory component in the respira-tory band (0.14-0.9 Hz, 8-54 breaths/minute) . . . . . . . . . 44Figure 3.4 The peak-conditioned fusion method combined the IRR esti-mates from three respiratory-induced variations to provide thefinal IRR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48Figure 3.5 Distribution of respiratory rates extracted from the capnogra-phy waveform (IRRCO2) in the capnobase data set. The res-piratory rates ranged from the lowest value of 3.6521 bpm tothe highest value of 44.22 bpm. The mean rate was 15.02 bpmwith standard deviation of 7.66 bpm. . . . . . . . . . . . . . . 49Figure 3.6 Distribution of the respiratory rates extracted from the nasal/o-ral airflow waveform (IRRnas) in the sleep data set. The res-piratory rates ranged from the lowest value of 9.561 bpm tothe highest value of 50.85 bpm. The mean rate was 18.64 bpmwith standard deviation of 5.66 bpm. . . . . . . . . . . . . . . 50Figure 3.7 Bland-Altman plot for comparison of IRRCO2 to IRRre f for allsubjects. The bias and 95% LOA are shown as solid lines. Thebias was 0.28 and the limits of agreement -3.62 to 4.17 . . . . 51Figure 3.8 Bland-Altman plot for comparison of IRRnas to IRRre f for allsubjects. The bias and 95% LOA are shown as solid lines. Thebias was 0.04 and the limits of agreement -5.74 to 5.82 . . . . 52Figure 4.1 The multi-class classifier has two binary classifiers: sleep/wakeclassifier and non-REM/REM classifier. The epochs corre-sponding to the subjects in the training set were used to trainthese two classifiers, and the epochs corresponding to eachsubject in the test set were fed to the trained models. . . . . . 67Figure 4.2 The area under the curve (AUC) of the receiver operating char-acteristic (ROC) curve of a) the wake/sleep classifier and b) thenon-REM/REM classifier . . . . . . . . . . . . . . . . . . . . 70xviiiFigure 5.1 The proposed model has two binary classifiers. The epochscorresponding to the subjects in the training set were used totrain these two classifiers, and the epochs corresponding toeach subject in the test set were fed to the trained classifiers. . 84Figure 5.2 The area under the curve (AUC) of the receiver operating char-acteristic (ROC) curve of a) A/H classifier (PPG + SpO2 fea-tures) and b) A/H classifier (SpO2 features) . . . . . . . . . . 91Figure 5.3 The estimated and observed A/H epochs for a subject with (a)high accuracy (79%), (b)low specificity with low number A/Hevents, and (c) low specificity with high number A/H events. . 92Figure 6.1 Sleep report provides the valuable information about the qual-ity of sleep, variation of heart rate and oxygen saturation dur-ing an overnight sleep . . . . . . . . . . . . . . . . . . . . . 95xixList of AbbreviationsAHI apnea/hypopnea IndexA/H apnea/hypopneaDFA detrended fluctuation analysisECG electrocardiogramEEG electroencephalographyEMG electromyographyEOG electrooculographyIRR instantaneous respiratory rateHRV heart rate variabilityOSA obstructive sleep apneaPPG photoplethysmogramPRV pulse rate variabilityRIAV respiratory-induced amplitude variationRIFV respiratory-induced frequency variationRIIV respiratory-induced intensity variationRIP respiratory inductance plethysmographyxxREM random eye movementRR respiratory rateSBD sleep breathing disorderSPO2 oxygen saturationSST synchrosqueezing transformSQI signal quality indexxxiAcknowledgementsTo my senior supervisor Dr. Guy A. Dumont who gave me the wings to fly and tomy supervisor, Dr. J. Mark Ansermino who showed me the way. I offer my endlessgratitude for their leadership, understanding, and support. And thanks for all theirinspiration as the successful researchers, excellent teachers, and brilliant people.Many thanks to Dr. Matthew Yedlin for his help and input to this research.Special thanks to Dr. Ainara Garde as she made her help and support availablein a number of different ways to this research.Many thanks to Dr. David Wensley for his consultation and input to this re-search.Many thanks to Dr. Walter Karlen and Dr. Christian L. Peterson for all help,support, and innovation they brought to this research.Many thanks to Dr. Kouhyar Tavakolian and Dr. Farzad Khosrow-khavar fortheir help and support.Special thanks to my friends and colleagues in Electrical & Computer Engi-neering for Medicine (ECEM) group at the University of British Columbia fortheir endless help and support.I would like to thank the clinical staff of the sleep laboratory at British ColumbiaChildren\u00E2\u0080\u0099s Hospital for their collaboration and assistance with this study.And finally, I am eternally grateful to Nick Langroudi for his unwavering pa-tience and support. This dissertation is the least I can dedicate to him.xxiiDedicationTo my husband, for twenty years of love and supportTo my girls, instead of all bedtime stories and lullabiesxxiiiChapter 1Introduction1.1 SleepThe basic structure of human sleep has three major stages: wakefulness, randomeye movement (REM) and non-REM. Later, non-REM sleep is divided into N1, N2,and N3 stages, progressing from stage N1 (light sleep) to stage N3 (deep sleep).non-REM sleep forms about 75 to 80 percent of the total time of sleep and REMsleep forms the remaining 20 to 25 percent.In a regular overnight sleep, non-REM and REM occur cyclically (Figure 1.1).Each cycle, taking about 90 to 120 minutes, starts with stage N1 (light sleep) ofnon-REM sleep, progresses to stage N2 and then to stage N3, repeats stages N2and N3 backward and then proceeds to REM sleep. The first half of an overnightsleep is dominated by deep sleep (stage N3) while the second half is dominated byREM sleep.Polysomnography (PSG) is the gold standard for assessing and scoring sleep.In an overnight PSG, the brain activity (electroencephalography (EEG)), eye move-ment (electrooculography (EOG)) and muscle activity (electromyography (EMG))are recorded according to the recommendation of American Academy of SleepMedicine (AASM).Later the EEG, EOG and EMG recordings are subdivided into 30-epochs. Asleep stage is assigned to each epoch by a sleep technician according to the AASMcriteria:1- Stage W represents alert wakefulness to drowsiness. The EEG signal consistsof Alpha waves (8 - 13 Hz), the EOG activity shows irregular peaked eye move-ments with the frequency of 0.5-2 Hz, and the EMG activity shows normal or highchin muscle tone.- Stage N1 forms 5% of total sleep time. The EEG consists of Theta waves (4-7Hz), the EOG shows slow eye movements, and the EMG is irregular but is oftenless variable than wakefulness.- Stage N2 makes up 50% of total sleep time. The EEG consists of Theta wavesmixed with K-complexes 1 and sleep spindles 2. The EOG recording doesn\u00E2\u0080\u0099t showsignificant eye movement. The EMG has variable amplitude, but usually lowerthan wakefulness.- Stage N3 forms 20-25% of total sleep time. The EEG wave consists of afrequency of a 0.5-3 Hz with high amplitudes (> 75 \u00C2\u00B5V). The EMG has variableamplitude, often lower than in Stage N2 and sometimes as low as in REM sleep.- During REM sleep, the EEG consists of low voltage mixed frequency Thetawaves (4-7 Hz). Alpha waves may be present but will be 1-2 Hz slower than Alphaduring wakefulness. The EOG shows the presence of rapid eye movements. TheEMG is significantly reduced compared to non-REM sleep.For an overnight sleep, PSG provides detailed information about the structureand pattern of sleep stages, overall sleep time, the time spent in different sleepstages and timing and structure of cycles. This information is simplified in a graphcalled hypnogram (Figure 1.1).1.2 Sleep Breathing Disordersleep breathing disorder (SBD) is characterized by abnormalities of respirationduring sleep. The International Classification of Sleep Disorders, second edition(ICSD-2) published by the American Academy of Sleep Medicine further catego-rized SBDs into obstructive sleep apnea/hypopnea syndromes, central sleep apnea1K-complex consists of a brief negative high-voltage peak, usually greater than 100 \u00C2\u00B5V, followedby a slower positive complex around 350 and 550 ms and at 900 ms a final negative peak2A burst of oscillatory brain activity visible on an EEG2Figure 1.1: Hypnogram, provided by PSG, depicts the structure of anovernight sleep, organization of cycles and timing of different sleepstages.syndromes, sleep related hypoventilation 3/hypoxemia 4 disorders [4]. Some pa-tients may have a combination of these disorders, particularly many patients havea combination of obstructive and central sleep apnea.American Academy of Sleep Medicine (AASM) Manual for the Scoring ofSleep and Associated Events [5] provides the terminology and scoring rules forsleep related respiratory events and disorders. It also provides the technical spec-ification for evaluation of a standard sleep test conducted in a sleep laboratory aswell as home sleep testing.Apnea is defined as the complete cessation of breathing during sleep whilehypopnea is defined as the reduction in airflow. In sleep apnea/hypopnea disorders,apnea/hypopnea (A/H) events happen intermittently during an overnight of sleep.apnea/hypopnea Index (AHI) is estimated as the number of A/H events happen inone hour of sleep and is used as a metrics of the severity of sleep apnea/hypopneasyndromes. Based on the AHI, the severity of sleep apnea syndrome in adults isclassified as follows:3Breathing at an abnormally slow rate4An abnormally low concentration of oxygen in the blood3- None/Minimal: AHI < 5- Mild: AHI \u00E2\u0089\u00A5 5, but < 15- Moderate: AHI \u00E2\u0089\u00A5 15, but < 30- Severe: AHI \u00E2\u0089\u00A5 30Obstructive Sleep Apnea DisorderIn obstructive sleep apnea (OSA) disorder, A/H events are caused by the complete/-partial closure of upper airway during sleep. In these events, the airflow is com-pletely ceased or dramatically reduced in the presence of respiratory efforts [51].According to the definition provided by AASM, A/H events last for a minimumof 10 s. Most A/H events take 10 s to 30 s. However, some of them may last tomore than one minute. The frequent occurrence of obstructive A/H events mayreduce the blood oxygen saturation which leads to a brief or complete arousal fromsleep to resume respiration. A/H events may occur in different sleep stages butmore frequently happen in stage N1, stage N2, and REM sleep than in stage N3.Frequent arousals and sleep fragmentations may cause daytime symptoms like ex-tensive sleepiness and fatigue which affect the quality of life.In adults, the prevalence of obstructive sleep apnea associated with daytimesleepiness has been estimated at 3% to 7% for males and 2% to 5% for females.However, OSA without daytime sleepiness may occur in 24% of adult men and 9%of adult women. Obesity, enlarged adenotonsillar tissue and structural informalityof upper airway are the main risk factors of OSA in adults. In adult patients, OSAis a risk factor of development of systematic hypertension and diabetes type 2.The prevalence of obstructive sleep apnea in children has been estimated at 1%to 4%. In children younger than 13 years old, the disorder occurs equally amongboys and girls but among adolescents, the provenance is higher in boys. Obesityand the enlarged tonsils and adenoids are the main cause of obstructive sleep ap-nea in children. Excessive sleepiness happens more in older children and less inyounger ones. Left untreated, OSA in children may have serious consequencesincluding developmental, behavioral and learning issues including concentrationproblems, hyperactivity and, moodiness [64].AASM has different criterion for defining obstructive A/H events in adults and4Table 1.1: Rules for scoring obstructive apnea/hypopnea events in adults andchildren defined by American Academy of Sleep Medicine (AASM)Adults ChildrenObstructiveApnea-A drop in airflow by \u00E2\u0089\u00A5 90%of pre-event baseline-Drop lasts for \u00E2\u0089\u00A5 10 s-A drop in airflow by \u00E2\u0089\u00A5 90%of pre-event baseline-Drop lasts for \u00E2\u0089\u00A5 2 breathsObstructiveHypopnea-A drop in airflow by \u00E2\u0089\u00A5 30%of pre-event baseline-Drop lasts for \u00E2\u0089\u00A5 10 s-An oxygen desaturation \u00E2\u0089\u00A5 3%from pre-event baseline or-the event is associated with an arousal-A drop in airflow by \u00E2\u0089\u00A5 30%of pre-events-Drop last for at least 2 breaths-An oxygen desaturation \u00E2\u0089\u00A5 3%from pre-event baseline or-the event is associated with an arousalchildren summarized in Table 1.1.Central Sleep Apnea DisorderCentral sleep apnea/hypopnea events are caused by complete or partial reductionsin central neural outflow to the respiratory muscles during sleep that leads to com-plete or partial cessation of airflow for at least 10 seconds, respectively [10]. Incontrast to obstructive apneas, in which respiratory efforts are observable, no respi-ratory effort is generated during central apnea/hypopnea events due to the cessationof respiratory drive. Thus central apneas are distinguished from obstructive apneasby the absence of respiratory effort.In the general population, the prevalence of central sleep apnea is less than 1%.However, central sleep apnea/hypopnea disorder has been reported in 25-40% ofpatients with heart failure and in 10% of patients who have had a stroke.Polysomnography (Sleep Study)A sleep study or polysomnography (PSG) is currently known as the gold standardfor diagnosis sleep-related disorders, especially SBDs. In an overnight PSG, thephysiological activity of body that occur during sleep are monitored in order todiagnose a wide range of respiratory and non-respiratory disorders of sleep.5PSG involves the measurement of several physiologic recordings including theEEG, EOG, electrocardiogram (ECG), submental and leg EMG, body position,pulse oximetry, measurements of airflow, and measurements of thoracic and ab-dominal respiratory effort.EEG records neural activity from electrodes placed on the patient\u00E2\u0080\u0099s scalp. Asmentioned in the previous section (section 1.1), EEG is performed to identify thestate of wakefulness and sleep and also to determine the different sleep stages, inaddition to recording arousals from sleep that may or may not be associated withrespiratory events. Recording of EOG and submental EMG are also necessary fordistinguishing wakefulness and REM sleep from other sleep stages.Monitoring the cardiorespiratory activity is essential for the diagnosis of sleepbreathing disorders. One lead-ECG is recommended by AASM to detect cardiacrhythm and identification of nocturnal arrhythmias. Nasal and oral airflow is mea-sured using a thermistor and/or a nasal-cannula pressure transducer. The results ofone study showed that the use of a nasal pressure transducer in conjunction witha thermistor was more sensitive than the thermistor alone in detecting hypopneaevents in adults and children [63]. So it is recommended by AASM to use the pres-sure transducer and thermistor together for measuring airflow. Beside measuringnasal/oral airflow measurement, monitoring the respiratory effort is also essentialfor assessing SBDs, especially for discrimination between obstructive and centralsleep apnea. In standard PSG, thoracic and abdominal respiratory effort is mea-sured using the respiratory inductance plethysmography (RIP) belts fasten aroundchest and abdomen. Pulse oximetry is used to detect reductions in blood oxygensaturation as a result of A/H events.PSG is highly resource-intensive [13] and requires a specialized sleep labo-ratory, expensive equipment and an overnight stay in the facility, confining PSGmonitoring to centralized specialist facilities. For example, in British Columbia,all PSG studies in children are performed at the British Columbia Children\u00E2\u0080\u0099s Hos-pital (BCCH) in Vancouver. This greatly limits access, especially for those wholive in remote locations. The capacity to perform PSG at BCCH is limited to fewerthan 250 cases per year, resulting in a waitlist of six months. Beside the limitedaccess, the high cost (approximately $800 per night in direct health care costs atBCCH) of PSG has generated a great interest in alternative techniques to simplify6the standard procedure.1.3 Pulse OximetryA pulse oximeter is a photoelectric device which non-invasively detects the bloodvolume changes, or photoplethysmogram (PPG), by measuring the light reflectedor transmitted through the body tissue (e.g finger, ear, forehead or nose lobe).A conventional transmitted pulse oximeter sensor has two light-emitting diodes(LEDs) and a light detector mounted on the opposite side of the LEDs. The LEDsalternatively emit red and infra-red light through the body and the light detectorcaptures the amount of transmitted light (Figure 1.2). The light intensity decreasesas the red and infrared beams pass through the body (e.g. skin, bones, tissue, ar-terial and venous blood). According to the Beer-Lambert low, the light intensitydecreases exponentially with the concentration and length of the light path as ex-plained by:I = I0e\u00E2\u0088\u0092l\u00CE\u00B1 (1.1)where I and I0 represent the intensity of transmitted and incident lights, re-spectively, l is the path length light traveled and \u00CE\u00B1 is the absorption coefficient ofblood.Based on the Beer-Lambert\u00E2\u0080\u0099s law, the density of transmitted light decreasesduring systole when the peripheral arterial blood volume is at its maximum valueand increases during diastole when the blood is minimum at the arteries. The PPGsignal generated by the light detector, then, has a pulsatile waveform (AC) whosepeaks and troughs reflect light transmitted through the tissue when blood volumeis minimal and maximal, respectively (Figure 1.3). AC offsets by a large baselinecomponent (DC) mainly rises because of constant absorption of light travellingthrough constant components e.g. skin, bones, and tissues. A small variation ob-served in DC is mostly due to venous blood variation which changes the intensityof the light captured by the light detector.7Figure 1.2: A conventional transmitted pulse oximeter sensor has two light-emitting diodes (LEDs) and a light detector mounted on the oppositeside of the LEDs (inspired by [75])Figure 1.3: Light transmitted through the living tissue (inspired by [75])Estimation of SpO2The absorption coefficient of oxyhemoglobin (oxy Hb) and deoxyhemoglobin (de-oxy Hb) is different at different wavelengths. The oxy Hb absorbs more infra-redlight than red light while deoxy Hb absorbs more red light than infra-red light. Bycomparing the amount of light absorbed by oxy Hb and deoxy Hb at two differentwavelengths, the pulse oximetry calculates oxygen saturation (SPO2) as explainedby [75]:8S = 1\u00E2\u0088\u0092R (1.2)R is called the Ratio of Ratios and is calucated asR =ln(AC+DC/DC)|\u00CE\u00BB1ln(AC+DC/DC)|\u00CE\u00BB2 (1.3)where \u00CE\u00BB1 and \u00CE\u00BB2 are the wavelengths of the red (660nm typ.) and infrared(890nm typ.) light, respectively.Figure 1.4: Absorption spectra of hemoglobin (from [75])Phone OximeterTMTo increase the accessibility of pulse oximeter and to take advantage of the preva-lence of mobile phones, a clinical pulse oximeter sensor can be interfaced to amobile phone. The commercially available pulse oximeter sensors have a micro-controller module featuring low power supply requirements and a communicationunit which are compatible with smart phones.The researchers in the Electrical & Computer Engineering in Medicine group inthe University of British Columbia, Vancouver, Canada developed a mobile device,named Phone Oximeter TM, which interfaces a commercial micro controller-based9pulse oximeter module with a smart phone [39] (Figure 1.5).Figure 1.5: The Phone OximeterTM interfacing a microcontroller-based pulseoximeter module with a smartphone.The use of the smartphone as the pulse oximeter display and power source over-comes pertinent challenges of distributing the technology. The Phone OximeterTMimproves accessibility of pulse oximetry, enables the acquisition, monitoring andanalysis of vital signs and provides intuitive display of information to health careproviders [62]. Usability studies of the Phone OximeterTM prototype previouslyundertaken both in Canada and Uganda have shown overall usability scores of 82%and 78% respectively, indicating that a smartphone can be a functional oximeter in-terface [32].Phone Oximeter TM has been further developed to perform all processing onthe mobile device through the audio interface [62] (Figure 1.6).For the purpose of this study, Phone Oximeter TM has been used for collectingthe PPG recordings.1.4 MotivationIn children with sleep apnea/hypopnea syndrome, the frequent cessation of breath-ing during sleep results in oxygen desaturations (a low level of oxygen in theblood), frequent arousal from sleep to resume breathing, fragmented sleep cycles10(a) (b)Figure 1.6: (a) Sensor and (b) the user interface of iOS App of Kenek O2Pulse Oximeterand ultimately sleep deprivation. Untreated sleep apnea in children has been linkedto cognitive and behavioral deficits, growth disorders, metabolic disorders, sys-temic inflammation, and serious cardiovascular consequences. Thus, it is clear thatsleep apnea has serious developmental consequences for children, highlighting theimportance of prompt diagnosis and treatment.PSG is the commonly used technique for sleep apnea diagnosis. Using PSGdata, sleep technicians visually identify apnea/hypopnea events, associated oxy-hemoglobin desaturations and arousals to estimate the sleep apnea severity. Thesleep states, sleep quality, sleep quantity and the number of non-REM-related andREM-related A/H events are also measured.PSG is highly resource-intensive [13] and requires an overnight stay at a highlyequipped sleep laboratory with an overnight attending sleep technician. This com-plexity confines the PSG test to the centralized facilities. For example, in BritishColumbia, all PSG studies in children are performed at the British Columbia Chil-11dren\u00E2\u0080\u0099s Hospital (BCCH) in Vancouver. This greatly limits access, especially forthose who live in remote locations. The capacity to perform PSG at BCCH islimited to fewer than 250 cases per year, resulting in a waitlist of six months. Inrecently developed clinical practice guidelines for the diagnosis and managementof SDB in children and adolescents [52], the American Academy of Pediatricsconcluded that all children/adolescents should be screened for snoring and OSAsymptoms (defined in the guidelines [52]) and PSG should be performed in thosewith regular snoring and signs of OSA.The complexity and cost of PSG (approximately $800 per night in direct healthcare costs at BC Children Hospital) [53] and limited access of PSG have generateda great interest in alternative techniques to simplify the standard procedure.The ultimate goal of this study was to develop a simple mobile screening toolfor sleep and SBD in children using the Phone Oximeter TM . The characterizationof the SpO2 and PPG signals both obtained by the Phone Oximeter TM were usedto detect the A/H epochs, different sleep stages and respiratory rate during sleep.1.5 ObjectivesThe objectives of this study are defined as:- to investigate the relative impact of SBD on sympathetic and parasympatheticactivity in children through spectral analysis and detrended fluctuation analysis(DFA) of pulse rate variability (PRV) extracted from PPG.- to develop a novel method for extracting the instantaneous respiratory rate(IRR) from PPG.- to extract the different states of the overnight sleep based on the activity ofcardiorespiratory system using the pulse oximeter PPG.- to propose a model to use the SpO2 pattern characterization and PPG analysisto identify the epochs with A/H events using the Phone OximeterTM .1.6 ContributionWe have made three significant contributions to the field:-To estimate the instantaneous respiratory rate (IRR) during sleep, we pro-posed and developed a novel algorithm for extracting the (IRR) from the PPG. The12method extracts the three respiratory-induced variation signals from PPG and esti-mates the IRR from them using a time-frequency transform called synchrosqueez-ing transform (SST). A novel algorithm, called peak-conditioned fusion, is pro-posed to fuse the IRR estimates and produce the final estimate of IRR. The noveltymostly was in designing and developing the peak-conditioned fusion algorithm.The details are described in Chapter 3.-To measure the sleep staging and to be able to determine the REM-relatedand non-REM-related A/H epochs, we designed and develop a novel model foridentifying the cycles of REM and non-REM of the overnight sleep based on theactivity of cardiorespiratory system using the overnight PPG signals. We build amultivariate model with two binary classifiers to identify wakefulness from sleep(wake/sleep classifier) and REM from non-REM sleep (non-REM/REM classifier).The developed classifiers were assessed epoch-by-epoch for each subject individ-ually and provided a detailed epoch-by-epoch sleep analysis, similar to the hypno-gram provided by PSG. The novelty was to use the characterization of PPG foridentifying sleep from wakefulness and furthermore, detecting the REM and non-REM stages of sleep. The details are presented in Chapter 4.- To screen apnea/hypopnea syndrome, we combined the SpO2 pattern charac-terization and PPG analysis to design and develop a model with two binary mul-tivariante logistic classifiers to automatically reject the 30-s PPG epochs contam-inated with the artifact and later identify the epochs with the A/H events. Thedevelopded model was assessed epoche-by-epoch for each subject and provided adetailed epoch-by-epoch A/H monitoring, similar to the one provided by PSC. Thenovelty was to combine the characterization of PPG and SpO2 for training two dif-ferent models for detecting the A/H events and rejecting aftifact. The details werepresented in Chapter 5.1.7 Organization of the DissertationThis dissertation is organized in 6 chapters (Figure 1.7).Chapter 1 provides an introduction to sleep, sleep breathing disorders and thetechnology of the pulse oximetry.Chapter 2 discusses how the analysis of PPG can be used to assess the cardiac13Figure 1.7: Organization of the Dissertation.modulation in children in response to sleep breathing disorders in different sleepstages.Chapter 3 presents a novel approach for extracting instantaneous respiratoryrate from PPG using the synchrosqueezing transform (SST)Chapter 4 discusses the application of PGG analysis for identifying differentsleep stages and presents a novel method for extracting sleep structure using thePPG features.Chapter 5 discusses the design and development a stand-alone tool for moni-toring and screening sleep breathing disorders at home using the Phone OximeterTMChapter 6 concludes the dissertation and presents suggestions for future workin monitoring sleep and sleep breathing disorders.14Chapter 2Evaluation of CardiacModulation in Children inResponse to Apnea/Hypopnea2.1 IntroductionThe autonomic nervous system (ANS) and circulating hormones play a signifi-cant role in regulating cardiovascular function. Regulation of heart rate is drivenmainly by interaction between the sympathetic and parasympathetic branches ofthe ANS. To increase heart rate, the ANS increases sympathetic outflow to thesinoatrial (SA) node, and concurrently reduces parasympathetic tone. Depressionof parasympathetic activity is necessary for the sympathetic nerves to increase heartrate because parasympathetic activity reduces the action of sympathetic nerve ac-tivity [44]. Since the regulation of heart rate is mainly controlled by the ANS, heartrate variability (HRV) has received significant attention as a promising non-invasiveindicator of cardiac autonomic function.HRV is defined as the variation in the inter-beat intervals (RRIs) conventionallyobtained from an electrocardiogram (ECG). RRIs time series are typically non-stationary and exhibit short and long-range fluctuations that occur in irregular andcomplex patterns, even during rest [59], [60], [33]. Short-range fluctuations cor-15respond to fast changes of heartbeat intervals associated with breathing and theregulation of blood pressure, whereas long-range fluctuations correspond to slowchanges of heartbeat intervals and reflect the effort of the ANS to limit heart rate[60].Power spectral analysis of HRV has been extensively used to study the fre-quency distribution of heart rate. Measured in short segments of RRIs time series,the power in the frequency range of 0.15 to 0.4 Hz, referred to as the high fre-quency power (HF), is commonly utilized to quantify parasympathetic activity.The power of HRV in the frequency range of 0.04 to 0.15 Hz, referred to as the lowfrequency power (LF), can be related to both sympathetic and parasympathetic ac-tivity. The ratio of LF to HF (LF/HF ratio) is defined as an index that represents thesympathetic/parasympathetic balance; a higher LF/HF ratio implies a shift towardsympathetic activity [33].Power spectral analysis assumes that the studied signal is stationary, and mayproduce inaccurate results when applied to non-stationary signals. This makespower spectral analysis inappropriate for quantifying the long-range fluctuation ofheart rate. To overcome this limitation, Peng et al introduced the Detrended Fluctu-ation Analysis (DFA)[59]. DFA determines the short- and long-range correlationsin a time series, expressed as scaling exponents. Peng et al showed that it is possi-ble to distinguish healthy subjects from those with severe heart failure by lookingat the short and long-range correlations in heartbeat intervals [60]. Later, Penzelet al. investigated the short- and long-range correlation of heart rate intervals mea-sured by DFA in individuals with SDB in different sleep stages and found that DFAimproved sleep apnea severity rating compared to spectral analysis [61].Traditionally, HRV is measured from the RRIs of the ECG. However, it ispossible to use pulse rate variability (PRV) extracted from the photoplethysmogra-phy signal (PPG) as an alternative measurement of HRV. More recent studies haveshown that in stationary conditions PRV could be used as an estimate of HRV [18],[43]. During non-stationary conditions, Gil et al[29] reported that there was a pos-itive bias, due to pulse time transit variability, in the estimation of PRV, especiallyin respiratory band. They showed that these differences were sufficiently small toallow the use of PRV as an alternative measurement of HRV.In individuals with SDB, intermittent sleep fragmentation and disturbance in16normal respiration and oxygenation that accompany most apnea/hypopnea eventscause changes in cardiac autonomic regulation [36]. These changes are reflected byreduced parasympathetic activity and enhanced sympathetic activity that persistsduring wakefulness [36]. Previous studies based on HRV analysis have demon-strated cardiac autonomic modulation due to SDB, and have shown that both theLF power and the LF/HF ratio are more pronounced in subjects with SDB, whilethe HF power is reduced [57], [74]. Cardiac sympathetic and parasympatheticmodulation in response to apnea/hypopnea has been well studied in adults, but isless extensively studied in children.In this study, we investigated the relative impact of SDB on sympathetic andparasympathetic activity in children through spectral analysis and DFA of PRV. Weestimated PRV from the pulse-to-pulse intervals of the PPG signal. The PPG sig-nals were recorded from 160 children using the Phone OximeterTM in the standardsetting of overnight polysomnography (PSG).2.2 Materials and methods2.2.1 ParticipantsFollowing approval by the University of British Columbia Clinic Research EthicsBoard (H11-01769) and informed parental consent, 160 children were recruitedfor this study. The children were suspected of having SDB and had been referredto the British Columbia Children\u00E2\u0080\u0099s Hospital for overnight PSG. Children with acardiac arrhythmia or abnormal hemoglobin were excluded from the study. Therecordings of 14 subjects were removed from the dataset due to inadequate lengthof sleep (less than 3 hours). The children were divided into two groups using thePSG outcomes and diagnostic report of the respiratory specialist: subjects with anAHI greater than 5 apnea/hour (SDB group) and children with an AHI less than 5apnea/hour (non-SDB group) (table 2.1).2.2.2 Data collectionStandard PSG recordings were performed with the Embla Sandman S4500 (Em-bla Systems, ON, Canada) and included overnight measurements of ECG, elec-17Table 2.1: Demographics and AHI index of studied population expressed asmean \u00C2\u00B1 standard deviation.In this table: 1Rapid Eye Movement; 2Body Mass Index; 3Total SleepTime; 4Total Bed Time; \u00E2\u0088\u0097p < 0.001; \u00E2\u0088\u0097\u00E2\u0088\u0097p < 0.0001 comparing SDB andnon-SDB; \u00E2\u0080\u00A0p-value< 0.001 comparing AHI in REM and non-REM sleepstagesDataset SDB non-SDBNumber 56 (18, 38) 90 (41, 49)Age (y) 8.8 \u00C2\u00B1 4.6 9.3 \u00C2\u00B1 4AHI 19.7 \u00C2\u00B1 19.5\u00E2\u0088\u0097\u00E2\u0088\u0097 1.4 \u00C2\u00B1 1.1AHI in REM1 \u00E2\u0080\u00A0 34.8 \u00C2\u00B1 27.8\u00E2\u0088\u0097\u00E2\u0088\u0097 4.4 \u00C2\u00B1 5.1AHI in non-REM 15.8 \u00C2\u00B1 22.8\u00E2\u0088\u0097\u00E2\u0088\u0097 0.8 \u00C2\u00B1 1.0BMI2 (kg/m2) 23.2 \u00C2\u00B1 8.3\u00E2\u0088\u0097 19.6 \u00C2\u00B1 6.6Sleep efficiency (%) 75.1 \u00C2\u00B1 16.2 76.6 \u00C2\u00B1 15.3TST3 (min) 362.1 \u00C2\u00B1 82.6 368.0 \u00C2\u00B1 73.8TBT4 (min) 479.9 \u00C2\u00B1 40 481.4 \u00C2\u00B1 24.1non-REM (%) 78.7 \u00C2\u00B1 9.3 81.7 \u00C2\u00B1 7.6REM (%) 20.2 \u00C2\u00B1 8 18.2 \u00C2\u00B1 6.1Awakenings 21.2 \u00C2\u00B1 10.6 18.6 \u00C2\u00B1 9.3Respiratory arousals 13.6 \u00C2\u00B1 13.9\u00E2\u0088\u0097\u00E2\u0088\u0097 1.0 \u00C2\u00B1 0.9troencephalography (EEG), oxygen saturation (SpO2), PPG, chest and abdominalmovement, nasal and oral airflow, left and right electrooculography (EOG), elec-tromyography (EMG) and video capture. The PSG was later annotated by a sleeptechnician with sleep phases and events (apneas, hypopneas, and arousal).In addition to PSG, PPG, heart rate, and SpO2 were recorded simultaneouslywith the Phone OximeterTM . The SpO2 and PPG signals were sampled at 1 Hzand 62.5 Hz, respectively, with 32-bit resolution.2.2.3 Pre-processingAfter baseline removal and smoothing with a Savitzky-Golay FIR filter (order 3,frame size 11 samples), all PPG signals recorded using the Phone OximeterTMwere divided into one-minute segments with a 30-second shift. These one-minutesegments were used to assess autonomic cardiac modulation during the A/H events18for each subject with SDB (intra-individual event analyses). In addition, the PPGsignals were divided into five-minute segments with 30 seconds shift and usedto assess autonomic cardiac modulation in subjects with and without SDB (inter-groups analyses).Each segment was assigned a signal quality index between 0 and 100 based on across correlation method [40] and segments with low signal quality index (less than50) were excluded from further analysis even if a very small part of segment wascontaminated by artifact. In order to obtain the PPIs time series, a peak detectionalgorithm based on zero-crossing was used to locate the pulse peaks in the PPGsignal, and the intervals between successive peaks were computed. PPIs shorterthan 0.33 s and greater than 1.5 s were considered artifacts [61] and consequentlydeleted from the time series.2.2.4 Sleep and apnea analysisAll segments were scored as wakefulness, REm or non-REM based on the labels inthe PSG event log file. Segments with any sleep state transition containing multiplesleep state labels were removed from the data set.One-minute segments with any period of SDB, such as obstructive or centralsleep apnea were labelled as A/H. According to the AASM 2012 standard criteria[5], obstructive apneas in children are defined as complete cessation of airflow (onairflow cannula) in the presence of respiratory effort lasting for more than 10 s.When respiratory effort partially or totally ceased, apneas were scored as mixedapnea or central apnea, respectively. Hypopneas were defined as a 30% airflowreduction for the duration of two breaths (Table 1.1).2.2.5 Parameter extractionTime-domain parametersThree time domain parameters were extracted from the PPIs time series, includingthe mean of the PPIs (meanPP), the standard deviation of the PPIs (SDPP) and theroot mean square of difference of the successive PPIs (RMSSD).19Power spectral analysisPPIs were resampled into the equivalent, uniformly spaced time series (so calledPRV) at a sampling rate of 4 Hz using the Berger algorithm [8]. PRV was char-acterized in the spectral domain using power spectral density (PSD). To provide abetter frequency resolution a parametric power spectral estimation was performedthrough an autoregressive modeling with 1024 points and an order of 16. Thepower in each of the following frequency bands was computed by determining thearea under the PSD curve bounded by the band of interest: Very Low Frequency(VLF; 0.01-0.04 Hz), Low Frequency (LF; 0.04-0.15 Hz) and High Frequency (HF;0.15-0.4 Hz). Normalized LF (nLF) and normalized HF (nHF) powers were deter-mined by dividing LF and HF powers by the total spectral power of PRV between0.04 and 0.4 Hz, respectively. The ratio of low-to-high frequency power (LF/HFratio) was also computed.Detrended Fluctuation Analysis (DFA)To quantify the short and long-range fluctuation of heart rate, we applied DFA tothe PPIs time series. DFA detects the internal correlation of signal expressed byscaling properties. To calculate DFA, we followed a four-step procedure [60]:Step 1: An integrated version of the original PPIs time series was calculated asy(k) =k\u00E2\u0088\u0091i=1[PPI(i)\u00E2\u0088\u0092PPIavg] (2.1)where PPI(i) was the ith PPIs, PPIavg was the mean of PPIs and k = 1,...,N. N wasthe total number of pulses.Step 2: The time series y(k) was divided into equally spaced Nn = int(N/n)non-overlapping windows with length n (number of pulses in each window).Step 3: For each window, the local trend yn(k) was separately calculated by aquadratic least-squares fit. Then the variance was determined for each window by\u00CF\u00832n (v) =1nn\u00E2\u0088\u0091k=1[y((v\u00E2\u0088\u00921)\u00E2\u0088\u0097n+ k)\u00E2\u0088\u0092 yn(k)]2 (2.2)where v=1,...,Nn.20Step 4: Finally, to obtain F(n), the fluctuation function, the root-mean-squareof all variances was calculated byF(n) =\u00E2\u0088\u009A1NnNn\u00E2\u0088\u0091v=1\u00CF\u00832n (v) (2.3)In order to determine how F(n) depends on the time scale n, the process was re-peated for several time scales n. Typically, F(n) increases as a power law when nincreases,F(n)\u00E2\u0088\u00BC n\u00CE\u00B1 (2.4)In a double logarithmic plot, the scaling exponent \u00CE\u00B1 shows the slope of a line thatfits log(F(n)) to log(n) (Figure 2.1). An \u00CE\u00B1 = 0.5 corresponds to an uncorrelatedtime series. 0 < \u00CE\u00B1 < 0.5 is indicative of anti-correlation time series, which meansthat short and large intervals are more likely to alternate. 0.5 < \u00CE\u00B1 < 1 representscorrelation in the time series which means short intervals are more likely to befollowed by short intervals and vice versa [60].In short-range correlations, \u00CE\u00B1 differs from 0.5 for small ns but will approach0.5 for large ns. In long-range correlations \u00CE\u00B1 is greater than 0.5 and less than 1 forlarge ns.To determine the short and long-range correlation in PPIs sequences, we de-fined \u00CE\u00B1S and \u00CE\u00B1L respectively, as the slopes of log(F(n)) as a function of log(n) forthe range 10 < n < 40 and for the range 70 < n < 200 [61].2.2.6 Data analysisThe Lilliefors test showed that the extracted parameters were not normally dis-tributed. The Wilcoxon Signed Rank test was therefore performed to evaluate thedifferences between the the segments with and without A/H events. The WilcoxonSum Rank test was also used to assess the differenced between the parameters ofthe two groups with and without SDB. A probability of p < 0.05 was consideredsignificant and no multiple-comparison correction method was used.To distinguish children with and without SDB during the entire sleep, a logisticregression model was fitted to the data set. Least absolute Shrinkage and SelectionOperator (LASSO) method was used to select the significant features [35]. \u00CE\u00BB was21100 101 102 103100101102103Interval n (pulses)F(n) non\u00E2\u0088\u0092REMREMFigure 2.1: In double logarithmic plot, the fluctuation function of PPIs, F(n),is plotted as a function of n (the number of pulses) for a child withoutSDB during non-REM (blue squares) and REM sleep (red stars). Theslopes of the curves correspond to the fluctuation scaling exponent \u00CE\u00B1 .For n > 100, the fluctuation function of PPI during REM and non-REMare distinguishable.tuned by stratified 10-fold cross validation; significant features were selected basedon the chosen \u00CE\u00BB . The LASSO model predicted the probability of having SDB foreach subject. To classify subjects into SDB and non-SDB groups based on thepredicted probabilities, instead of using a default threshold of 0.5, we calculated arisk threshold based on the maximum weighted classification score [67].2.3 ResultsIn the following subsections, the estimation of different parameters during A/Hevents for the individual children with SDB (intra-individual event analyses) andalso in groups with and without SDB (inter-groups analysis) have been presented.222.3.1 Intra-individual event analysisFor the whole group, totalling 70856 one minute segments, 32574 were included inthe analysis, with 38282 excluded due to artifacts. Of a total of 5040 segments la-belled as apnea/hypopnea, 3267 were included in the analysis, with 1377 excludedfor artifacts and 326 excluded due to multiple sleep labels.Based on Wilcoxon Signed Rank test, spectral domain parameters differed sig-nificantly (p-value < 0.0001) in apnea/hypopnea events.For the duration of the entire sleep, the nLF increased in apnea/hypopnea eventsfor 96% of the children with SDB. Similarly, the LF/HF ratio increased in ap-nea/hypopnea events for 96% of the children with SDB, while nHF decreased in94% of children with SDB during apnea/hypopnea events(Figure 2.2).During non-REM sleep, for 95% of children with SDB, higher nLF, higherLF/HF ratio, and lower nHF were recognized in segments with apnea/hypopneaevents compared to segments without SDB (Figure 2.3).During REM sleep, for 73% of the children with SDB, the nLF and LF/HFratio increased in apnea/hypopnea events. In addition, for 68% of the children withSDB, nHF decreased in the apnea/hypopnea events (Figure 2.4).The VLF increased during apnea/hypopnea events for almost 90% of the chil-dren with SDB during non-REM sleep and REM sleep (Figure 2.3 and Figure 2.4).Time domain parameters differed in apnea/hypopnea events but the differenceswere not statistically significant.2.3.2 Inter-groups analysisThe meanPPIs were significantly shorter in children with SDB during whole sleep,non-REM and REM sleep. SDPP and RMSSD did not vary significantly betweenthe two groups (Table 2.2, Table 2.3 and Table 2.4).The VLF was higher in children with SDB compared to the group withoutSDB. The differences were greater during non-REM sleep. Compared to childrenwithout SDB, in the SDB group, the nLF and LF/HF ratio were significantly higherduring non-REM sleep, but did not differ significantly during REM sleep. The nHFwas lower in children with SDB relative to children without. This difference wasgreater during non-REM sleep compared to REM sleep (Figure 2.5, Table 2.2,23non\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91VLFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized LFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized HFnon\u00E2\u0088\u0092A/H A/H00.511.522.5LF/HF RatioFigure 2.2: Comparison of spectral parameters in segments with and withoutapnea/hypopnea events for children with SDB (AHI > 5) during theentire period of sleep. Blue (thin) and red (thick) lines show the meanincrease and decrease of parameters, respectively. The nLF parameterincreased in apnea/hypopnea events for 96% of the children with SDB(Blue lines). The LF/HF ratio increased in apnea/hypopnea events for96% of the children with SDB (Blue lines), while nHF decreased in94% of children with SDB during apnea/hypopnea events (Blue lines).The VLF parameter increased during apnea/hypopnea events for almost92% of the children with SDB (Blue lines).Table 2.3 and Table 2.4).In a double logarithmic representation, the function F(n) in the range of 10 < n< 200, was clearly distinct between the SDB group and the non-SDB group, duringnon-REM sleep (Figure 2.6). However, during REM sleep these two functionswere not clearly demarcated (Figure 2.7).Greater \u00CE\u00B1S and \u00CE\u00B1L values were observed for children with SDB compared tochildren without. However, \u00CE\u00B1L varied much more significantly than \u00CE\u00B1S and the24non\u00E2\u0088\u0092A/H A/H00.20.40.60.81VLFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized LFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized HFnon\u00E2\u0088\u0092A/H A/H00.511.522.5LF/HF RatioFigure 2.3: Comparison of spectral parameters in segments with and with-out apnea/hypopnea events for children with SDB (AHI > 5) duringthe non-REM sleep. Blue (thin) and red (thick) lines show the meanincrease and decrease of parameters, respectively. For 95% of childrenwith SDB, higher nLF, higher LF/HF ratio, and lower nHF were recog-nized in segments with apnea/hypopnea events compared to segmentswithout SDB. The VLF parameter increased during apnea/hypopneaevents for almost 90% of the children with SDB.differences were greater during non-REM sleep (Figure 2.8).By applying the LASSO method to the data set to classify children with andwithout SDB during the entire sleep, a model with three significant features (meanPPIs,VLF, and \u00CE\u00B1L) was selected. Based on a calculated risk threshold of 0.36, an AUCof 78% was obtained for this model, providing accuracy, sensitivity and specificityof 71%, 76% and 68%, respectively.25non\u00E2\u0088\u0092A/H A/H00.20.40.60.811.2VLFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized LFnon\u00E2\u0088\u0092A/H A/H00.10.20.30.40.50.60.70.80.91Normalized HFnon\u00E2\u0088\u0092A/H A/H00.511.522.5LF/HF RatioFigure 2.4: Comparison of spectral parameters in segments with and with-out apnea/hypopnea events for children with SDB (AHI > 5) during theREM sleep. Blue (thin) and red (thick) lines show the mean increaseand decrease of parameters respectively. During REM sleep, for 73%of the children with SDB, the nLF and LF/HF ratio increased in ap-nea/hypopnea events. In addition, for 68% of the children with SDB,nHF decreased in the apnea/hypopnea events. The VLF parameter in-creased during apnea/hypopnea events for almost 90% of the childrenwith SDB.2.4 DiscussionThe results of this study showed that the cardiac sympathetic indices of PRV werehigher during A/H events for more than 95% of children with SDB (AHI > 5).These indices were also higher in children with SDB compared to children without.In addition, heart rate was higher and the short- and long-range fluctuations ofheart rate were more strongly correlated in children with SDB. Also, we found thatcardiac sympathetic indices were modulated by sleep stages.Although many studies have been conducted in adults, few studies have in-26Table 2.2: Descriptive results (median) of estimated parameters for childrenwith and without SDB during the entire sleep periodmean 95% CInon-SDB SDB differences (Low, High) p-valuemeanPPIs 0.800 0.710 0.070 (0.012 , 0.124) 0.005SDPP 0.050 0.054 0.007 (-0.004, 0.017) 0.100RMSSD 0.052 0.052 0.004 (-0.008, 0.014) 0.29VLF 0.100 0.190 0.083 (0.024 , 0.145) 0.0001nLF 0.280 0.340 0.050 (0.005 , 0.098) 0.001nHF 0.710 0.650 0.050 (0.004 , 0.098) 0.001Ratio 0.443 0.560 0.130 (0.022 , 0.258) 0.010\u00CE\u00B1S 0.700 0.820 0.090 (0.015 , 0.164) 0.010\u00CE\u00B1L 0.600 0.680 0.078 (0.032 , 0.122) 0.0005Table 2.3: Descriptive results (median) of estimated parameters for childrenwith and without SDB during non-REM sleepmean 95 % CInon-SDB SDB differences (Low, High) p-valuemeanPPIs 0.819 0.715 0.072 (0.011 , 0.126) 0.005SDPP 0.046 0.050 0.007 (-0.003, 0.017) 0.10RMSSD 0.049 0.054 0.004 (-0.007, 0.015) 0.23VLF 0.089 0.174 0.067 (0.012 , 0.120) 0.005nLF 0.251 0.314 0.041 (0.000 , 0.091) 0.050nHF 0.749 0.685 0.041 (0.000 , 0.091) 0.050Ratio 0.394 0.542 0.100 (0.000 , 0.220) 0.040\u00CE\u00B1S 0.688 0.777 0.077 (-0.007, 0.153) 0.030\u00CE\u00B1L 0.539 0.621 0.065 (0.022 , 0.108) 0.005vestigated the effects of SDB on the autonomic cardiac regulation in children. Inparticular, few studies have examined autonomic function in children through theanalysis of the PPG obtained from a pulse oximeter, and none have used a mobiledevice for this purpose. In the rest of this section, we compare our findings withthe results of studies based on HRV.Gil et al showed that during non-stationary conditions there were some smalldifferences between HRV and PRV, mainly in the respiratory band, which were27Table 2.4: Descriptive results (median) of estimated parameters for childrenwith and without SDB during REM sleepmean 95% CInon-SDB SDB differences (Low, High) p-valuemeanPPIs 0.761 0.697 0.070 (0.017 , 0.120) 0.005SDPP 0.049 0.059 0.004 (-0.005, 0.015) 0.21RMSSD 0.047 0.044 0.002 (-0.008, 0.012) 0.35VLF 0.233 0.338 0.048 (-0.044, 0.144) 0.14nLF 0.408 0.430 0.026 (-0.035, 0.087) 0.2nHF 0.591 0.569 0.026 (-0.035, 0.087) 0.2Ratio 0.704 0.874 0.096 (-0.086, 0.301) 0.15\u00CE\u00B1S 0.884 0.96 0.044 (-0.053, 0.134) 0.18\u00CE\u00B1L 0.817 0.873 0.032 (-0.043, 0.105) 0.18related to the pulse transit time variability [29]. However, they concluded thatthese differences were sufficiently small to suggest the use of PRV as an alternativemeasurement of HRV.2.4.1 Intra-individual event analysisDuring non-REM sleep, the segments with apnea/hypopnea events were character-ized by higher values of the nLF and LF/HF ratio and lower values of nHF for 95%of children with SDB. This may show that sympathetic modulation was predomi-nant during apnea/hypopnea events while parasympathetic activity was diminished.During REM sleep, we found that for 73% of SDB children, the nLF andLF/HF ratios increased in apnea/hypopnea events and for 68% of children, thenHF power decreased in the apnea/hypopnea events. These results indicate thatthe predominance of sympathetic activity (increase in the nLF and LF/HF ratios)in apnea/hypopnea events is suppressed by cardiac sympathetic modulation duringREM sleep.The VLF was higher in apnea/hypopnea events for 90% of the children withSDB, during non-REM, consistent with an increase in the slow regulation of car-diac function [36]. However, longer signal segments (>1-minute) are required tofurther validate these results.28non\u00E2\u0088\u0092SDB SDB00.51**VLFnon\u00E2\u0088\u0092SDB SDB00.51**normalized LFnon\u00E2\u0088\u0092SDB SDB00.51**normalized HFnon\u00E2\u0088\u0092SDB SDB00.511.522.5**LF/HF Ratio(a)non\u00E2\u0088\u0092SDB SDB00.51**VLFnon\u00E2\u0088\u0092SDB SDB00.51*normalized LFnon\u00E2\u0088\u0092SDB SDB00.51*normalized HFnon\u00E2\u0088\u0092SDB SDB00.511.522.5*LF/HF Ratio(b)non\u00E2\u0088\u0092SDB SDB00.51VLFnon\u00E2\u0088\u0092SDB SDB00.51normalized LFnon\u00E2\u0088\u0092SDB SDB00.51normalized HFnon\u00E2\u0088\u0092SDB SDB00.511.522.5LF/HF Ratio(c)Figure 2.5: Frequency domain parameters in children with and without SDBduring (a) the entire sleep period, (b) non-REM sleep and (c) REMsleep. Significant differences between the SDB and non-SDB groupsare marked by one star (*) when p-value < 0.05 and by two stars (**)when p-value < 0.01. Quartile values are displayed as the bottom, mid-dle and top horizontal line of the boxes. Whiskers are used to representthe most extreme values within 1.5 times the interquartile range fromthe median. Outliers (data with values beyond the ends of the whiskers)are displayed as (+).Bahavaret et al employed HRV spectral analysis to assess autonomic cardiacregulation in children with SDB in overnight sleep studies [6]. They also foundthat epochs containing obstructive sleep apneas had higher values of the nLF andLF/HF ratios and lower nHF than the epochs without the respiratory events.29100 101 102 103101102103interval n (beats)f(n) SDBnon\u00E2\u0088\u0092SDBFigure 2.6: The fluctuation function F(n) during non-REM sleep for childrenwith SDB children (blue squares) and non-SDB children (red stars).2.4.2 Inter-groups analysisDuring both REM and non-REM sleep, the PPIs appeared shorter in children withSDB (decreased meanPPIs). Since the meanPPIs did not significantly vary in ap-nea/hypopnea events, we would argue that heart rate was generally higher in chil-dren with SDB compared to those without, which may indicate higher sympatheticmodulation in children with SDB. Khandoker et al who investigated PPIs duringsleep apnea in adults also reported a significant higher heart rate [43].During non-REM sleep, we found that the nLF and LF/HF ratios were signifi-cantly higher and nHF was lower in the SDB group, relative to the non-SDB group.The same trend was observed in children with SDB during REM sleep, althoughthese differences were not statistically significant. These findings showed an en-hanced sympathetic activity and a diminished parasympathetic activity in childrenwith SDB in response to sleep apnea. However, during REM sleep, this cardiacmodulation was also provoked by the sleep state. Furthermore, we discovered that30100 101 102 103100101102103interval n (beats)f(n) SDBnon\u00E2\u0088\u0092SDBFigure 2.7: The fluctuation function F(n) during REM sleep for children withSDB children (blue squares) and non-SDB children (red stars).the decrease in the nHF in children with SDB was more significant than the in-crease in the nLF. This may indicate that children with SDB exhibit a strongerdecrease of parasympathetic activity rather than an increase of sympathetic activ-ity, as confirmed by Chouchou et al [14].Baharavet et al also showed that the nLF and LF/HF ratios were higher forchildren with SDB during non-REM and REM sleep [6]. They reported statisticallysignificant differences in nHF and LF/HF ratios during non-REM sleep betweentwo groups, in agreement with our findings.Our findings from DFA analysis suggest that the short- and long-range fluctu-ation of heart rate are more strongly correlated in children with SDB compared tochildren without SDB. We found that in children with SDB, both \u00CE\u00B1S and \u00CE\u00B1L werelarger, relative to the children without SDB, during both non-REM and REM sleepstages. Since the short-range correlation is associated with the effects of breath-ing on heart rate, this large \u00CE\u00B1S value may indicate that the control of heart rate inthe range of respiratory related time scales (10 < n < 40) is much tighter in chil-31non\u00E2\u0088\u0092SDB SDB non\u00E2\u0088\u0092SDB SDB0.20.40.60.811.2 ** **\u00CE\u00B1S \u00CE\u00B1L(a)non\u00E2\u0088\u0092SDB SDB non\u00E2\u0088\u0092SDB SDB0.40.60.811.2 * **\u00CE\u00B1S \u00CE\u00B1L(b)non\u00E2\u0088\u0092SDB SDB non\u00E2\u0088\u0092SDB SDB0.40.60.811.2\u00CE\u00B1S \u00CE\u00B1L(c)Figure 2.8: \u00CE\u00B1S and \u00CE\u00B1L in children with and without SDB during (a) entiresleep period, (b) non-REM sleep and (c) REM sleep. Significant differ-ences between SDB and non-SDB group are represented by one star (*)when p-value < 0.05 and by two stars (**) when p-value < 0.01.32dren with SDB. Furthermore, as mentioned by Khoo et al [36], in subjects withSDB respiratory modulation is not limited to the high frequency band (0.15 - 0.4Hz). In SDB, respiratory modulation of heart rate takes the form of a large cyclicalvariation that correlates with episodic apnea or hypopnea and mostly elevates thecomponents of VLF band. These results, showing greater values of \u00CE\u00B1L in childrenwith SDB, are consistent with an elevated VLF band.Penzel et al investigated the short and long range correlations of heart rateintervals measured by DFA in adults during different sleep stages [61]. They found\u00CE\u00B1S = 1.00 and \u00CE\u00B1L = 0.67 for adults without SDB (age = 33.0 \u00C2\u00B1 6.4 years) duringthe whole sleep. These values are larger than our values calculated for childrenwithout SDB (age = 9.1\u00C2\u00B1 4.2 years). This suggests that the fluctuation in the RRIsof adults without SDB is more strongly correlated than the fluctuation in PPIs ofchildren without SDB.We analysed the different features of PRV in different sleep stages. We foundthat in non-REM sleep, the features of PRV varied significantly in apnea/hypopneaevents. However, during REM sleep, the same features extracted from segmentswith apnea/hypopnea events were not distinguishable from segments without ap-nea/hypopnea events. Nevertheless, the results obtained from the PRV analysis ap-plied to the whole sleep recording, showed that even without considering the stageof sleep, PRV features were significantly different in segments with apnea/hypop-nea events. This means that even when the sleep stage information is not available,it is possible to distinguish apnea/hypopnea events through PRV.To classify children with and without SDB based on only the PRV featuresacross the entire sleep, we achieved an accuracy of 71% using a fitted model withthe three selected features (meanPPIs, VLF, and \u00CE\u00B1L). This is comparable to theresults of a study by Penzel et al [61] which showed an accuracy of 72.9% classi-fying adults based on their apnea severity using eight spectral and DFA features ofHRV.In intra-individual event and inter-group analyses we characterized PRV using1- and 5-minute sliding windows respectively, to answer two different questions. Inthe intra-individual event analysis we compared the extracted temporal and spec-tral parameters between the segments with and without apnea/hypopnea events.We considered 1-minute segments to ensure that the segments are small enough33to contain only the apnea/hypopnea event(s) and/or the arousal(s) accompanyingthem. In the inter-individual analysis, to assess the cardiac modulation in SDB,we divided the children into two groups; those with and without SDB. Accordingto a study by Penzel et al [61], the DFA parameters extracted from segments witha duration of 5 minutes or more are more distinguishable between children withand without SDB in different sleep stages (Figure 2.1). So, we estimated PPIs andextracted parameters for each group using a 5-minute sliding window.2.4.3 Limitations and future workWe found that A/H events induced cardiac modulation; however, we did not inves-tigate whether this modulation was influenced by arousal, hypoxia or the durationof A/H events.In this study, we considered the AHI >= 5 as the criteria for SDB. However,there is no discrete definition of SDB based on AHI alone, but rather a continuumfrom normal to abnormal. We recognize that some studies consider an AHI >= 2as mild SDB. Therefore, we will further investigate the characterization of PRV formonitoring children with SDB based on different AHI thresholds (AHI >= 1, AHI>= 2).In this study, to characterize PRV in intra-individual event and inter-group anal-yses, we chose two sliding windows with different lengths, which may be consid-ered as a study limitation.2.4.4 Clinical relevanceThe findings of this study confirm that SDB affects the regulation of cardiac func-tion, suggesting that it would be possible to use the effects of SDB on cardiacmodulation to detect apnea/hypopnoea events in children. Furthermore, we havepreviously shown that the characterization of overnight SpO2 pattern measured bythe Phone OximeterTM successfully identifies children with significant SDB [25].Hence, combining the characterization of SpO2 and PRV, both recorded by PhoneOximeterTM holds promise as a low-cost approach to automatically assess SDB athome [26]. This can greatly increase the accessibility to sleep apnea screening andimprove the quality of life for the many children currently affected by SDB related34disorders.35Chapter 3Extracting InstantaneousRespiratory Rate fromPhotoplethysmogram3.1 IntroductionRespiratory rate (RR), along with other vital signs like heart rate and blood pres-sure, is monitored for primary or continuous assessment of patient wellness. Thereis significant evidence that an abnormal respiratory rate is an important predictorof serious illness. For example, in children aged 1-5 years old, an elevated RR (>40 breaths/min) is an important criterion for the diagnosis of pneumonia [76]. Fur-thermore, Fieselmann et al analyzed the measurements of vital signs during the 72hours prior to cardiac arrest and showed that a high respiratory rate (> 27 breath-s/min) was a significant predictor of cardiac arrest in hospitals [22]. In addition,Subbe et al showed that relative changes in respiratory rate are much more signif-icant than changes in heart rate or systolic blood pressure in unstable patients andtherefore the respiratory rate is more likely to be a better predictor for identifyingthe patient at risk [69].RR can be measured by a nurse counting the number of times the chest rises inone minute [47]. Continuous monitoring of RR, though, needs a monitoring device36and can be performed using capnography, transthoracic impedance pneumogra-phy, nasal/oral pressure transducers and abdominal/thoracic respiratory inductanceplethysmography belts, among others. However, recent studies have found thatneither the nurses nor the monitoring devices provide accurate and reliable mea-surements of RR [47]. Therefore, there is a clear need for a robust, automatic,reliable and non-invasive measure of RR for performing a spot-check and for con-tinuous monitoring.Analysis of the PPG recorded using a pulse oximeter could offer an alternativemethod for monitoring RR. The PPG waveform contains information about a widerange of physiological parameters such as heart rate (HR), heart rate variability(HRV), oxygen saturation (SpO2), vascular tone, blood pressure, cardiac outputand respiration [65]. However, most conventional pulse oximeters only provideHR and SpO2. In this study, we present a novel algorithm for robust estimationof instantaneous respiratory rate (IRR) from PPG with the aim of developing aportable solution based on pulse oximetry, suitable for both continuous monitoringand spot-check applications.3.1.1 BackgroundRespiration may induce variation in PPG in three different ways [54] (Figure 3.1):1) respiratory-induced intensity variation (RIIV): Changes in venous return dueto changes in intra-thoracic pressure throughout the respiratory cycle cause a base-line (DC) modulation of the PPG signal. During inspiration, decreases in intra-thoracic pressure result in a small decrease in central venous pressure increasingvenous return. The opposite occurs during expiration. As the venous bed at probingsite cyclically fills and drains, the baseline is modulated accordingly.2) respiratory-induced amplitude variation (RIAV): During inspiration, left ven-tricular stroke volume decreases due to changes in intra-thoracic pressure leadingto the decreased pulse amplitude. The opposite happens during expiration.3) respiratory-induced frequency variation (RIFV): Heart rate varies throughoutthe respiratory cycle; heart rate increases during inspiration and decreases duringexpiration. This phenomenon well-known as respiratory sinus arrhythmia (RSA)is mainly due to the autonomic regulation of heart rate during respiration.37Figure 3.1: From top, PPG with no modulation, Respiratory-Induced In-tensity Variation (RIIV), Respiratory-Induced Amplitude Variation(RIAV), and Respiratory-Induced Frequency Variation (RIFV)Respiration may induce variation in PPG differently among different individu-als in health and disease. For instance, RIFV, as an indicator of autonomic activity,may be affected by diseases and disorders (e.g. myocardial infarction, diabeticneuropathy or sleep breathing disorders [20]). RIAV and RIIV are also very sen-sitive to dehydration and hypovolemia. In addition, respiratory-induced variationsare different for women and men (For men, when the respiratory rate was not morethan 10 breaths/min, the frequency variation has the strongest correlation to the res-piratory signal; whereas up to or above 15 breaths/min, in the sitting position, theintensity variation has the strongest correlation to the respiratory signal and in thesupine position, amplitude variation has the strongest. For women, the frequencyvariation correlates with respiration more strongly than the other variations, nearlyindependent of the respiratory rate or posture)[34]. As such, estimation of IRRby combining the information from three respiratory-induce waveform variations,improves the algorithm performance and increases the robustness of results [41].Many algorithms have been proposed to estimate RR from PPG. Auto-regression38[71], Fourier transform analysis [41], correntropy spectral density [26], digital fil-ters [56] and empirical mode decomposition [24] were successfully used, amongothers. These algorithms have mostly focused on estimating average RR from aPPG segment. For example, [41] and [26] estimated RR every second using 16,32, 64-second segments of PPG data.Few algorithms, however, have proposed to estimate RR instantaneously, mostlyperformed by time-frequency approaches based on a continuous wavelet [3] , [16],variable frequency complex demodulation methods (VFCDM) [42] and short-timeFourier analysis (STFT) [66].In this study, we have proposed a novel method for extracting IRR from PPG.The method is performed in three main steps: extraction of RIIV, RIAV and RIFVsignals from PPG, estimation of IRR from each extracted respiratory-induced vari-ation signals and fusion of IRR estimates. A time-frequency transform calledsynchrosqueezing transform (SST) [17] is used to extract RIIV, RIAV and RIFVfrom PPG. Later, a second SST is applied to estimate IIR from respiratory-inducedvariation in signals [2]. To fuse IRR estimates corresponding to each respiratory-induced variation signal, a novel method, called peak-conditioned fusion algorithmis proposed.3.2 Algorithm Description3.2.1 Instantaneous Frequency (IF)The instantaneous frequency is the frequency at a given time. Consider a multi-component signal f that can be modelled asf (t) =K\u00E2\u0088\u0091k=1fk(t) =K\u00E2\u0088\u0091k=1Ak(t)cos(2pi\u00CF\u0086k(t)) (3.1)where Ak(t) and \u00CF\u0086k(t) are the time-varying amplitude and phase of kth fre-quency component, respectively.The instantaneous frequency (IF) is defined as the derivative of the phase func-39tion with respect to time asIFf = {\u00CF\u0086 \u00E2\u0080\u00B2k(t)}1\u00E2\u0089\u00A4k\u00E2\u0089\u00A4K (3.2)3.2.2 Synchrosqueezing Transform (SST)The SST was first introduced by Daubechies et al. [17] in 1996 and then imple-mented by Thakur et al. [70]. SST is a combination of wavelet analysis and areallocation method which sharpens a time-frequency representation by allocatingits points to another locations in the time-frequency plane. SST can provide anaccurate estimation of IF.As defined in [17], SST involves three steps:Step 1: Estimation of the continuous wavelet transform (CWT)The CWT of f is calculated asWf (a,b) =\u00E2\u0088\u00ABf (t)a\u00E2\u0088\u00921/2\u00CF\u0088(t\u00E2\u0088\u0092ba)dt (3.3)where \u00CF\u0088 is a wavelet with \u00CF\u0088\u00CB\u0086(\u00CE\u00BE ) = 0 for \u00CE\u00BE \u00E2\u0089\u00A4 0 and a and b are scale andlocation variables, respectively. \u00CF\u0088(\u00CE\u00BE ) is the complex conjugate of \u00CF\u0088(\u00CE\u00BE ) and \u00CF\u0088\u00CB\u0086(\u00CE\u00BE )is the Fourier transform of \u00CF\u0088(\u00CE\u00BE ) estimated as\u00CF\u0088\u00CB\u0086(\u00CE\u00BE ) =\u00E2\u0088\u00AB\u00CF\u0088(\u00CE\u00BE )e\u00E2\u0088\u0092i(2pi\u00CE\u00BE )tdt (3.4)Step 2: Estimation of the instantaneous frequencyIf \u00CF\u0088\u00CB\u0086(\u00CE\u00BE ) is concentrated around \u00CE\u00BE =\u00CF\u00890, then Wf (a,b) will be spread out aroundthe horizontal line a= \u00CF\u00890/\u00CF\u0089 on the time-scale presentation for a given frequency of\u00CF\u0089 . However, Daubechies et al. [17] showed that the oscillation of Wf (a,b) aroundb tends to the original frequency \u00CF\u0089 , irrespective of the value of a. Therefore, forany (a,b) where Wf (a,b) 6= 0, the instantaneous frequency \u00CF\u0089 f (a,b) for signal fcan be defined as\u00CF\u0089 f (a,b) =\u00E2\u0088\u0092 i2pi ((Wf (a,b))\u00E2\u0088\u00921 \u00E2\u0088\u0082\u00E2\u0088\u0082bWf (a,b)) (3.5)Step 3: Transfer to the time-frequency plane40In this step, each point on the time-scale plane is allocated to a point on thetime-frequency plane using the map (a,b)\u00E2\u0086\u0092 (\u00CF\u0089 f (a,b),b). The frequency variable\u00CF\u0089 and the scale variable a are both binned: Wf (a,b) is computed only at discretevalues ak, with ak\u00E2\u0088\u0092ak\u00E2\u0088\u00921 = (\u00E2\u0088\u0086a)k and its SST, Tf (\u00CF\u0089,b) is estimated only at the cen-ters \u00CF\u0089l of the successive bins [\u00CF\u0089l\u00E2\u0088\u0092 12 ,\u00CF\u0089l + 12 ], with \u00CF\u0089l\u00E2\u0088\u0092\u00CF\u0089l\u00E2\u0088\u00921 = \u00E2\u0088\u0086\u00CF\u0089 , by summingdifferent points:Tf (\u00CF\u0089l,b) = (\u00E2\u0088\u0086\u00CF\u0089)\u00E2\u0088\u00921 \u00E2\u0088\u0091ak:|\u00CF\u0089(ak,b)\u00E2\u0088\u0092\u00CF\u0089l |\u00E2\u0089\u00A4 \u00E2\u0088\u0086\u00CF\u00892Wf (ak,b)a\u00E2\u0088\u009232k (\u00E2\u0088\u0086a)k. (3.6)3.3 Material and Methods3.3.1 Data setsCapnobase data setThe Capnobase contains test and calibration data sets [38]. The test data set con-tains forty-two 8-min segments of recordings obtained from 29 pediatric and 13adults receiving general anesthesia at the British Columbia Childrens Hospital andSt. Pauls Hospital, Vancouver, BC, respectively. Calibration data set contains onehundred twenty-four 2-min segments of recordings used for tuning the parametersof the proposed algorithm.In both data sets , the recordings included ECG, capnometry, and PPG (sampledat 300 Hz, 300 Hz and 100 Hz, respectively) obtained with S/5 collect software(Datex-Ohmeda, Finland). The capnography waveform was used as the referencegold standard recording for RR. A research assistant manually labelled each breathin the capnogram. The beginning and end of all artifacts in the PPG waveformswere also manually labelled. Both datasets can be downloaded from the on-linedatabase, CapnoBase.org.Sleep data setThe Sleep database contains forty-three 20-min segments of recording from 43children referred to the British Columbia Children\u00E2\u0080\u0099s Hospital for overnight stan-41dard polysomnography (PSG). The children had been recruited following approvalby the University of British Columbia Clinic Research Ethics Board (H11-01769)and informed parental consent. Children with a cardiac arrhythmia or abnormalhemoglobin were excluded from the study.Standard PSG recordings included overnight measurements of ECG, electroen-cephalography (EEG), oxygen saturation (SpO2), PPG, chest and abdominal move-ment, nasal and oral airflow, left and right electrooculography (EOG), electromyo-graphy (EMG) and video capture. The PSG recordings were performed with theEmbla Sandman S4500 (Embla Systems, ON, Canada).In addition to PSG, the PPG was recorded simultaneously using the PhoneOximeterTM sampled at 62.5 Hz with 32-bit resolution.The nasal/oral airflow waveform was used as the reference gold standard record-ing for RR. Two expert manually labelled each breath in nasal/oral airflow wave-form. The beginning and end of all artifacts in the oral/nasal waveforms were alsomanually labelled.3.3.2 Estimation of IRR from PPGTo perform IRR estimation, after a preprocessing stage, a first SST was appliedto PPG to extract RIIV, RIAV and RIFV. Later, a second SST was performed toestimate IIR from the respiratory-induced variation signals. The peak-conditionedfusion algorithm was then used to fuse simultaneous IRR estimates. This pro-cedure, inspired by the method known as secondary wavelet feature decoupling(SWFD) [2], involves the following steps (Figure 3.2):1) The first SST is applied to the PPG signal.2) In the STT surface plot, two components are identified: a strong cardiaccomponent in the cardiac band (0.5-3 Hz, 30-180 beats/minute) and a respiratorycomponent in the respiratory band (0.14-0.9 Hz, 8-54 breaths/minute) (Figure 3.3).In this study, reference ranges of cardiac and respiratory bands were extractedfrom a review of observational studies that used HR from 143,346 children andRR data from 3,881 children (from 6 months to 18 years old) [22]. Based on 99thand 1st centiles for children and young adults, the HR could range from 30 to 180beats/min (0.50 to 3 Hz, respectively) and RR from 8 to 54 breaths/min (0.14 to42Figure 3.2: To extract IRR from PPG, the first SST was applied to PPG toextract RIIV, RIAV and RIFV. Later, the second SST was performedto estimate IIR from a respiratory-induced variation signals. The peak-conditioned fusion algorithm was then used to fuse simultaneous IRRestimates0.9 Hz, respectively). The range in adults is much more restricted, thus it would beincluded in this range.3) The respiratory component in the SST surface plot shows RIIV and its ridgein the frequency-time plane represents RIIV-derived IRR (IRRriiv) (Figure 3.3).4) The ridge of the cardiac component is followed either in the amplitude-time plane to get RIAV or in the frequency-time plane to get RIFV. This is doneby projecting the cardiac ridge points onto the amplitude-time or frequency-timeplanes, respectively.5) The second SST applied to RIAV results in a dominant single componentin the respiratory band (0.14-0.9 Hz, 8-54 breaths/minute) whose ridge representsRIAV-derived IRR (IRRriav)6) A second SST is applied to the RIFV signal as well to get a dominant sin-gle component in the respiratory band whose ridge represents RIFV-derived IRR(IRRri f v).7) Estimation of final IRR (IRRppg) is performed using a proposed peak fre-quency tracking method (so-called peak-conditioning fusion) which combines theinstantaneous frequency information from (IRRriiv), (IRRriav) and (IRRri f v).43Figure 3.3: In the STT surface of PPG, two components are identified: astrong cardiac component in the cardiac band (0.5-3 Hz, 30-180 beat-s/minute) and a respiratory component in the respiratory band (0.14-0.9Hz, 8-54 breaths/minute)PreprocessingThe PPG signals were lowpass filtered by a lowpass Chebyshev Type I IIR filter oforder 8 and down sampled to 10 Hz.Estimation of IRRriivConsider a PPG signal as a vector ppg \u00E2\u0088\u0088 Rn, n = 2L+1 where L is a nonnegativeinteger. The CWT of ppg, Wppg, was calculated using the Morlet wavelet, \u00CF\u0088 , whereits Fourier transform was concentrated around 1.25 Hz. The Wppg was sampled atthe location (a j,b), where a j = 2 j/nv , j = 1, ...,Lnv, nv = 32 and b = 1, ...,n. Theresult is a Lnv\u00C3\u0097n matrix denoted W\u00CB\u009Cppg.When W\u00CB\u009Cppg > 0, \u00CF\u0089\u00CB\u009Cppg was implemented as follow44\u00CF\u0089\u00CB\u009Cppg =\u00E2\u0088\u0092 i2piDbW\u00CB\u009Cppg(a j,b)W\u00CB\u009Cppg(a j,b)\u00E2\u0088\u00921 (3.7)where DbW\u00CB\u009Cppg was the finite differences of W\u00CB\u009Cppg with respect to b.Then frequency variable, \u00CF\u0089 , was binned into frequency division \u00CF\u0089l = 2l4\u00CF\u0089\u00CF\u0089 ,l = 0, ...,Lnv\u00E2\u0088\u0092 1, where 4\u00CF\u0089 = 1Lnv\u00E2\u0088\u00921 log2(n2), \u00CF\u0089 = 1n4t and \u00CF\u0089\u00C2\u00AF = 124t . \u00CF\u0089\u00C2\u00AF and \u00CF\u0089 ,were maximum and minimum frequencies respectively and were chosen based onNyquist sampling theorem.The SST of PPG was calculated asTppg(\u00CF\u0089l,b) = \u00E2\u0088\u0091a j:|\u00CF\u0089(a j,b)\u00E2\u0088\u0092\u00CF\u0089l |\u00E2\u0089\u00A4 \u00E2\u0088\u0086\u00CF\u00892log2LnvW\u00CB\u009Cppg(a j,b)a\u00E2\u0088\u009212j . (3.8)Tppg over time shows both cardiac and respiratory bands (Figure 3.3).A ridge fitting the dominant area of Tppg in the respiratory band (0.14 Hz - 1Hz) represented IRRriiv and was extracted by tracking the local maximum valuesin this region.Estimation of IRRriavConsider RIAV as a vector riav\u00E2\u0088\u0088Rn, where n is the length of ppg. In the amplitude-time plane of Tppg, riav estimated as a ridge fitting the dominant area of Tppg in thecardiac band (0.5 Hz - 3 Hz, 30 - 180 beats/minute). The ridge extracted by findingthe local maximum values which minimize the following cost function [1]:Cost =n\u00E2\u0088\u0091b=1[\u00E2\u0088\u0092|Tppg(riav(b),b)|2+ |riav(b)\u00E2\u0088\u0092 riav(b\u00E2\u0088\u00921)|2] (3.9)The SST of riav, Triav was calculated using the same implementation describedin the previous section, .A ridge fitting the dominant area of Triav in the respiratory band (0.14 Hz - 1Hz) represented the RIAV-derived IRR (IRRriav) and can be extracted by trackingthe local maximum values in this region.45Estimation of IRRri f vConsider RIFV as a vector ri f v\u00E2\u0088\u0088Rn, where n is the length of ppg. In the frequency-time plane of Tppg, rifv estimated as a ridge fitting the dominant area of Tppg in thecardiac band (0.5 Hz - 3 Hz, 30 - 180 beats/minute). The ridge extracted by findingthe local maximum values which minimize the following cost function [1]:Cost =n\u00E2\u0088\u0091b=1[\u00E2\u0088\u0092|Tppg(ri f v(b),b)|2+ |ri f v(b)\u00E2\u0088\u0092 ri f v(b\u00E2\u0088\u00921)|2] (3.10)The SST of riav, Triav was calculated using the same implementation describedin the section 3.3.2.A ridge fitting the dominant area of Triav in the respiratory band (0.14 Hz - 1Hz) represented the RIFV-derived IRR (IRRrifv) and can be extracted by trackingthe local maximum values in this region.Peak-Conditioned FusionThe peak-conditioned fusion method, inspired by [45], was proposed to combinethe IRR estimates from three respiratory-induced variations to provide the finalIRRppg.The calculated Tppg, Triav and Trifv are two-dimensional matrices \u00E2\u0088\u0088 RLnvn, n =2L+1 where L is a nonnegative integer and nv = 32. Each column of Tppg, Triav andTrifv matrices shows the frequency distribution of PPG, RIAV and RIFV signals ateach time instance, respectively. To reduce the variance, each matrix is averagedin time dimension using a moving window of length Tm = 16s every ts = 5 s. Theaveraged matrix is denoted as T\u00CB\u0086k, where k refers to ppg, riav or rifv (Figure 3.4).At instant b, the location of the largest peak in respiratory band of each T\u00CB\u0086k(:,b)column (for k = ppg, riav or rifv) is detected and denoted as irIk(b). Then, a refer-ence frequency interval, \u00E2\u0084\u00A6k(b), was defined as\u00E2\u0084\u00A6k(b) = [ f (b\u00E2\u0088\u00921)\u00E2\u0088\u0092\u00CE\u00B4 , f (b\u00E2\u0088\u00921)+2\u00CE\u00B4 ] (3.11)where (b\u00E2\u0088\u00921) was a respiratory rate reference estimated from the b\u00E2\u0088\u00921 previ-ous step.All peaks larger than 85% of irIk(b) inside \u00E2\u0084\u00A6(b) were detected and irIIk (b) was46chosen as the nearest to f(b\u00E2\u0088\u00921). By reaching to this point, irIIriiv(b), irIIriav(b) andirIIrifv(b) were available simultaneously.The final respiratory peak at instant b, IIRppg((b)), was then chosen amongirIIriiv(b), irIIriav(b) and irIIrifv(b) estimates with the largest Pk. Pk is a measure of thepeakness and was defined as the ratio of power contained in an interval centredaround the largest peak to the power of \u00E2\u0084\u00A6k(b). P mathematically calculated aspk(b) =\u00E2\u0088\u0091min{i f IIk (b)+0.6\u00CE\u00B4 , f (b)+2\u00CE\u00B4}max{i f IIk (b)\u00E2\u0088\u00920.6\u00CE\u00B4 , f (b)\u00E2\u0088\u0092\u00CE\u00B4}\u00CB\u0086Tk(:,b)\u00E2\u0088\u0091 f (b)+2\u00CE\u00B4f (b)\u00E2\u0088\u0092\u00CE\u00B4 T\u00CB\u0086k(:,b)(3.12)Estimation of respiratory rate as the largest peak in the respiratory band wouldincrease the risk of choosing the location of false peaks. To decrease this risk, thesearch for the largest peak was limited to the reference frequency interval, \u00E2\u0084\u00A6k(b)[45]. This is an asymmetric interval of 3\u00CE\u00B4 centred around a reference frequency.At each step the respiratory rate reference was updated usingf (b+1) = \u00CE\u00B2 \u00E2\u0088\u0097 f (b)+(1\u00E2\u0088\u0092\u00CE\u00B2 )\u00E2\u0088\u0097 IRRppg(b) (3.13)where f (b) = arg max(T\u00CB\u0086k(:,1)) in the frequency band of [0.2Hz,0.7Hz].Value of \u00CE\u00B4 was set as 0.1 and the value of a was tuned as 0.6 over the calibrationdata set.3.3.3 Algorithm EvaluationTo evaluate the performance of SST-based algorithms, agreement between refer-ence IRR and estimated IRR (using peak-conditioned fusion, simple fusion, singlerespiratory-induce variation) was assessed using the limits of agreement (LOA)technique. The bias and 95% LOA were estimated using the Bland-Altman plot.Since for each subject multiple measurement were observed, the Bland-Altmanmethod for multiple observations per individual [79] was used instead of the stan-dard Bland-Altman method. The bias was calculated as mean of IRRest - IRRre fand the 95% LOAs as mean bias 1.95 standard deviations. Two standard devia-tions (2SD) were also estimated in the purpose of ranking the proposed algorithmin this study based on the statistical analysis reported by [11].47Figure 3.4: The peak-conditioned fusion method combined the IRR estimatesfrom three respiratory-induced variations to provide the final IRRThe coverage probability (CP2) was also reported as the probability of mea-surement error falling within pre-defined bounds, set as 2 breaths per minute (bpm)in this study [7].3.4 Results3.4.1 Capnobase data baseIRR extracted from the capnography waveform (IRRCO2) was used as the referencegold standard. The distribution of the respiratory rates contained 3542 data pointsestimated every 5 second from IRRCO2 for the 16 second moving windows over thewhole dataset (Figure 3.5). The respiratory rates ranged from the lowest value of3.6521 bpm to the highest value of 44.22 bpm. The mean rate was 15.02 bpm with48Figure 3.5: Distribution of respiratory rates extracted from the capnographywaveform (IRRCO2) in the capnobase data set. The respiratory ratesranged from the lowest value of 3.6521 bpm to the highest value of44.22 bpm. The mean rate was 15.02 bpm with standard deviation of7.66 bpm.standard deviation of 7.66 bpm. About 7.7% of the data points were excluded fromthe further analysis due to to poor signal quality of the capnography signals.For each algorithm, the measures of agreement between the estimated IRRfrom PPG (IRRest) and IRRCO2 were estimated (Table 3.1). For peak selectionalgorithm, bias was estimated as 0.28 bpm with the 95% LOAs from -3.62 to 4.17(Figure 3.7). The value of 2SD was estimated as 3.97 bpm.The values of 2SD of the other algorithms ranged from 8.32 bpm to 16.00 bpm.49Figure 3.6: Distribution of the respiratory rates extracted from the nasal/oralairflow waveform (IRRnas) in the sleep data set. The respiratory ratesranged from the lowest value of 9.561 bpm to the highest value of 50.85bpm. The mean rate was 18.64 bpm with standard deviation of 5.66bpm.3.4.2 Sleep databaseIRR extracted from the nasal/oral airflow waveform (IRRnas) was used as the ref-erence gold standard in the sleep dataset. The distribution of the respiratory ratescontained 10553 data points estimated every 5 second from IRRnas over the 16 sec-ond moving window for all subjects. The respiratory rates ranged from the lowestvalue of 9.561 bpm to the highest value of 50.85 bpm. The mean rate was 18.64bpm with standard deviation of 5.66 bpm. About 0.66% of the data points wereexcluded from the further analysis due to to poor signal quality of the nasal/oralairflow signals (Figure 3.6).The measures of agreement between the estimated IRR from PPG (IRRest) and50Table 3.1: The performance of different method for estimation IRR from PPGDifferent IRR Proportionof windows withestimation Method 2SD Bias 95% LOA CP2 IRR estimate (%)RIIV 8.80 0.35 -8.29 to 8.98 88 100RIAV 16.00 1.27 -14.47 to 16.89 60 100Capnobase RIFV 9.22 0.04 -9.00 to 9.10 74 100dataset Simple Fusion 8.32 0.55 -7.62 to 8.69 63 100Peak-Conditioned Fusion 3.97 0.28 -3.62 to 4.17 89 100RIIV 11.00 0.66 -10.11 to 11.42 80 100RIAV 21.34 5.56 -15.36 to 26.49 31 100Sleep RIFV 8.44 -0.11 -8.40 to 8.16 79 100dataset Simple Fusion 9.51 2.03 -7.29 to 11.35 41 100Peak-Conditioned Fusion 5.90 0.04 -5.74 to 5.82 85 100Figure 3.7: Bland-Altman plot for comparison of IRRCO2 to IRRre f for allsubjects. The bias and 95% LOA are shown as solid lines. The bias was0.28 and the limits of agreement -3.62 to 4.1751Figure 3.8: Bland-Altman plot for comparison of IRRnas to IRRre f for all sub-jects. The bias and 95% LOA are shown as solid lines. The bias was0.04 and the limits of agreement -5.74 to 5.82IRRnas were estimated for each algorithm (Table 3.1). For peak selection algorithm,bias was estimated as 0.04 bpm with the 95% LOAs from -5.74 to 5.82 (Figure 3.8).The value of 2SD was estimated as 5.90 bpm.The values of 2SD of the other algorithms ranged from 8.32 bpm to 16.00 bpm.3.5 Discussion and ConclusionIn this study, we presented an algorithm to extract IRR from PPG. We extractedRIIV, RIAV and RIFV from PPG using SST, a sharpening time-frequency methodwhich provides instantaneous frequency rate. The peak-conditioned fusion wasproposed to combine the extracted information from three respiratory induced vari-ations waveforms to estimate respiratory rate at each instance. We validated theimplemented method with capnography and nasal/oral airflow as the reference RR.Compared to simple fusion and single respiratory-induced variation estimations,peak-conditioned fusion shows better performance (Table 3.1). It provided a bias52of 0.28 bpm with the 95% LOAs ranging from -3.62 to 4.17, validated againstcapnography (in the Capnobase dataset) (Figure 3.7) and a bias of 0.04 bpm withthe 95% LOAs ranging from -5.74 to 5.82, validated against nasal/oral airflow (inthe Sleep dataset) (Figure 3.8).In this study, the proposed method estimated IRR from three sources of respiratory-induced variation and fused the estimated rates to measure the final IRR. Our find-ings showed that fusion of estimation rates would increase the accuracy and ro-bustness of RR estimation. Even the simple fusion compared to single respiratory-induced variation estimations showed higher rank (narrower 2SD and greater CP2).It is consistent with the findings of [34] that respiratory activity may induce vari-ation in PPG differently in different individuals. As discussed by [41], ventilatoryconditions (spontaneous or mechanical ventilation) can change the behaviour ofrespiratory induced variations.In this study, we applied the proposed algorithm to two different data sets toinclude a broad range of subjects into the study. The Capnobase data set includeschildren adults, under controlled ventilation or spontaneously breathing over awide RR range. The subjects were under general anesthesia and were continuouslymonitored. The sleep data set includes children from 1-month to 17 years old spon-taneously breathing during two hours of overnight sleeping in a sleep lab. Duringrecording, respiratory rates might change significantly while sleep progressed dur-ing different stages of light sleep, deep sleep or REM sleep. Some of the childrenmay have experienced periods of breathing cessation, or obstructive sleep apnea,as well. Relate the finding to those of similar studiesA recent study [11] represented a very comprehensive assessment of RR esti-mation using PPG. A wide range of available techniques for estimation of respiratory-induced variations from PPG, estimation of RR from respiratory-induced varia-tions, and fusion of RR estimates were identified and then more than 300 algo-rithms were implemented by assembling all possible combinations of availabletechniques. The algorithms were ranked based on 2SD. The first ten top rankedalgorithms had the 2SD values ranging from to 6.2 to 7.9. Compared to the tentop ranked algorithms, our proposed method showed the best performance with the2SD values of 3.9 and 5.9 for Capnobase and Sleep datasets, respectively. In ad-dition, for the top ranked algorithms, the value of CP2 was reported as 71.5 while53we obtained a CP2 of 88 applying our proposed algorithm.It is important to note that all top ranked algorithm reported in [11] estimatedRR using 32-second windows while our method can estimate RR instantaneously.It suggests that our algorithm shows better performance compared to methods thatextract IRR based on time-frequency analysis [3], [34].In the [11], the methods for extracting RR from ECG were assessed as well.The findings of that study showed that algorithms performed better when usingECG than PPG. The best algorithm had 95% LOAs of 4.7 to 4.7 bpm and a bias of0.0 bpm when using the ECG.In the [11], the performance of thoracic Impedance Pneumography(IP) wereassessed as well providing a bias of 0.2 bpm with 95% LOAs of 5.6 to 5.2 bpm.Thoracic IP is a commonly-used technique for continuous monitoring of RR thatmeasures changes in the electrical impedance of the persons chest during respira-tion. Our results showed that the performance of our algorithm is comparable withthe performance of thoracic IP.Several studies based on the continuous wavelet transform (CWT) [3], [16],the short-time Fourier transform (STFT) [66], and empirical mode decomposition(EMD) [26] have been proposed to detect RR from PPG. The results of a studyconducted by Thakur et al [70] to compare SST to CWT, STFT and EMD showedthe superior precision of SST at identifying components of complicated oscillatorysignals. Moreover, the study showed that time-varying instantaneous frequenciescould be clearly distinguished in the SST while there is much more smearing anddistortion in the CWT and STFT.This study introduces a new method to estimate IRR from pulse oximetry. Thiswould expand the functionality of a conventional pulse oximetry beyond the mea-surement of HR and SpO2 to measure the respiratory rate continuously and in-stantly in the clinical setting and at home. Importantly, these are all achievablewith a simple, cheap, single-sensor solution.54Chapter 4Extracting the PediatricHypnogram fromPhotoplethysmogram4.1 IntroductionAs was mentioned in chapter 1 , sleep is divided into REM and non-REM sleep. Aregular overnight sleep occurs in cycles of non-REM and REM, usually four or fivesuch cycles per night. The hypnogram is a graph which depicts the basic structureof an overnight sleep (Figure 1.4)).The brain activity, eye movements and muscle tensions change during non-REM and REM stages. Also, sleep staging induces variation in heart rate, bloodpressure, respiration and vascular tone, mainly regulated by sympathetic and parasym-pathetic branches of the autonomic nervous system. The activity of the sympatheticnervous system decreases during non-REM sleep compared to wakefulness whichresults in a reduction in heart rate, blood pressure, respiratory rate and vasculartone. However, there might be some brief increase in heart rate and blood pres-sure due to respiratory events, arousals or body movements. Compared to thenon-REM, during REM sleep, there is a rise in the activity of sympathetic nervoussystem which leads to faster changes in heart rate, blood pressure, and respiratory55rate.As mentioned before, PSG is the gold standard for assessing sleep. In PSG,the recordings of brain activity (EEG), eye movement (EOG) and muscle activity(EMG) during sleep are used for sleep scoring. PSG requires an overnight stay ofpatients in the sleep laboratory with specialized equipment and all night attendingsleep technicians. The high cost and complicated procedure confine the PSG test tospecialized sleep centres, and it can rarely be used at any ambulatory environmentswhen the several days of monitoring of sleep behavior and circadian rhythm areneeded. Besides, the complex set-up and overnight stay in the hospital may affectsleep structure, resulting in inaccurate outcomes. As such, a less complex and lessexpensive ambulatory solution has been explored extensively.In recent years, activity of the cardiorespiratory system has been monitored forsleep staging. HR, HRV and respiration have recently been used as the reliabletools for identifying sleep and wake in adults [9], [61], [46], [37]. Penzel et alinvestigated different linear and non-linear features of HRV in subjects with andwithout sleep apnea [61] in various sleep stages. Lisenby et al classied REM andnon-REM states by analyzing heart rate in time and frequency domain [46]. Karlenet al used spectral analysis of ECG and respiratory signals recorded by a wearablesensor to classify sleep from wakefulness [37]. These studies showed that sleepclassification by monitoring the variation of heart and respiratory rate could attainresults similar to sleep scoring achieved by the technicians using PSG recordings.The purpose of this study is to extract the cycles of non-REM and REM of theovernight sleep based on the activity of cardiorespiratory system using the pulseoximeter PPG. We extracted the relevant features associated with PRV, RR, vascu-lar tone and movement from the PPG signal to build a multivariate model with aminimum set of features to identify wakefulness from REM and non-REM sleep.The PPG signals were recorded by Phone OximeterTM.564.2 Background4.2.1 Pulse Rate VariabilityPRV shows the variation of heart rate extracted from the pulse-to-pulse time inter-vals of PPG. Heart rate is mainly regulated by the inputs from the sympathetic andparasympathetic nervous systems. As such, power spectrum analysis of HRV hasbeen extremely used to verify the activity of the autonomic nervous system. In ourprevious studies, presented in chapter 2, we have assessed PRV as an estimate ofHRV during wakefulness, non-REM, and REM sleep. The results showed that thetemporal and spectral features of PRV were significantly different in wakefulness,non-REM and REM sleep [20].4.2.2 Vascular toneThe arterial vessels experience a level of contraction that determines their diame-ter, and therefore their tone [72]. The vascular tone can influence the morphologyof the PPG signal remarkably, involving the amplitude and area of each pulse; theamplitude of the PPG pulse is directly proportional to the vascular tone [65]. Dur-ing vasoconstriction, the pulse amplitude decreases, while during vasodilatation,the amplitude increases. Some studies show that non-REM sleep is associatedwith a decrease in sympathetic vascular tone and as a result a peripheral vasodila-tion while this condition is reversed in REM sleep. In this study, to identify REMfrom non-REM sleep, we measured the amplitude, width, and other characteristicsof pulse shape as the features of vascular tone induced variation during differentsleep stages.4.2.3 Respiratory rateAs was mentioned in chapter 3, respiration may modulate the PPG in three dif-ferent ways: 1) Respiratory-Induced Intensive Variation (RIIV), 2) RespiratoryInduced Amplitude Variation (RIAV), and 3) Respiratory Induced Frequency Vari-ation (RIFV).Since the respiratory rate changes during non-REM and REM, in this study,we estimated RIIV, RIAV and RIFV from the PPG signal and then estimated the57respiratory rates from respiratory-induced variation signals and used them as thefeatures for sleep staging.4.2.4 MovementIn actigraphy, the gross motor activity involved in movement and coordination ofthe arms, legs, and other large body parts is monitored to determine the sleep pat-terns. The PPG signal is usually corrupted very easily by motion artifacts due tomovement during data recording. In this study, in the absence of body accelera-tion measures, we used the motion artifacts in the PPG signal as the signs of bodymovement and being restless. We estimated the degree of signal corruption to lo-cate the motion artifact in the PPG signal using a signal quality index measure.Also, in most cases of movement, the cardiac synchronous pulsatile componentof arterial blood is corrupted by the random fluctuation of arterial blood, whichinduces changes in the morphology of PPG. Therefore, we also computed somerandomness measures of the PPG such as skewness and kurtosis, as important fea-tures of the PPG signal contaminated by motion artifact.4.3 Materials and Methods4.3.1 PPG PreprocessingThe same data set deployed for evaluation of cardiac modulation in children inresponse to apnea/hypopnea was used for this study (2.2.1). After baseline removaland smoothing with a Savitzky-Golay FIR filter (order 3, frame size 11 samples),all PPG signals recorded using the Phone Oximeter were divided into 30-secondepochs. A peak detection algorithm based on zero-crossing was used to locate thepulse peaks in the PPG signal segments. The accuracy of the peak detector wasestimated at approximately 99.2%. No attempt was made to distinguish normalpulses from others.4.3.2 Sleep LabellingAll epochs were scored as wakefulness, REM or non-REM, based on the labels inthe PSG event log file. The REM and non-REM epochs were scored as sleep as58well.4.3.3 Feature extractionFor each 30-second epoch of the PPG signal, the following features have beenextracted (Table 4.1):PRV FeaturesThe pulse-to-pulse intervals time series (PPIs) were computed as the intervals be-tween successive peaks. In the time domain, three parameters were extracted fromthe PPIs time series, including the mean of the PPIs (meanPP), the standard devia-tion of the PPIs (SDPP) and the root mean square of the difference of the successivePPIs (RMSSD).The PPIs were resampled into the equivalent, uniformly spaced time series (so-called PRV) at a sampling rate of 4 Hz using the Berger algorithm [8]. Then thepower spectral density of PRV was estimated using a parametric autoregressivemodel with 1024 points and an order of 7. The power in each of the followingfrequency bands was computed by determining the area under the power spectraldensity curve bounded by the band of interest: Very Low Frequency (VLF; 0.01-0.04 Hz), Low Frequency (LF; 0.04-0.15 Hz) and High Frequency (HF; 0.15-1Hz). Normalized LF (nLF) and normalized HF (nHF) powers were determined bydividing LF and HF powers by the total spectral power of PRV between 0.04 and0.4 Hz, respectively. The ratio of low-to-high frequency power (LF/HF ratio) wasalso computed.Vascular tone featuresSeveral morphology features were extracted from each PPG pulse to characterizethe vascular tone during different sleep stages.-meanAmp and stdAmp: the amplitude of each pulse was measured as thedifference between the maximum of a pulse (peak) and the previous minimum(trough). meanAmp and stdAmp were calculated as the average and standard de-viation of the amplitude of all pulses within the epoch, respectively.-meanWidthhal f and stdWidthhal f : the widthhal f of each pulse was calculated59as the width at 50% of pulse height; later, meanWidthhal f and stdWidthhal f werecalculated as the average and standard deviation of widthhal f of all pulses withinthe epoch, respectively.-meanWidth and stdWidth: the widthpulse of each pulse was calculated asthe width at 10% of pulse height; later, meanWidth and stdWidth were calculatedas the average and standard deviation of widthpulse of all pulses within the epoch,respectively.-meanTimerising and stdTimerising: the mean and standard deviation of Timerising(the time for a pulse takes to reach its peak) of all pulses within the epoch werecomputed.-meanTime f alling and stdTime f alling: the mean and standard deviation of Time f alling(the time for a pulse takes to reach its trough) of all pulses within the epoch werecomputed.-meanSlope and stdSlope: the mean and standard deviation of rising slope ofall pulses within the epoch-pwv: For each epoch, pulse wave variability (pwv) was estimated as:pwv =max(amp)\u00E2\u0088\u0092min(amp)(max(amp)+min(amp))/2(4.1)Respiratory rateFirst, three respiratory-induced variations (RIAV, RIIV and RIFV) were estimatedfrom each 30-second epoch of PPG and then the respiratory rates were estimated asthe maximum value peak frequencies in respiratory bands of the power spectrumof RIAV, RIIV and RIFV (0.15-1 Hz).-Respiratory Rate from RIAV (RRriav): To extract RIAV from the PPG sig-nal, the pulse amplitude time series were resampled into the equivalent, uniformlyspaced time series at a sampling rate of 4 Hz using the linear interpolation method.The power spectral density of RIAV was computed using a parametric autoregres-sive model with 1024 points and an order of 7. RRriav was estimated as the max-imum value peak frequency in the respiratory band of the RIAV power spectrum(0.15-1 Hz)-HFriav: the power within the respiratory band (0.15-1 Hz) of the RIAV power60spectrum.-Respiratory Rate from RIIV (RRriiv): for each epoch, first the intensity timeseries were estimated as the trend which connects the peaks of consequent pulses.The intensity time series were resampled into the equivalent, uniformly spacedtime series at a sampling rate of 4 Hz to get RIIV. The power spectral densityof RIIV was computed using a parametric autoregressive model with 1024 pointsand an order of 7. RRriiv was estimated as maximum value peak frequency in therespiratory band of the RIIV power spectrum (0.15-1 Hz).- HFriiv: the power within the respiratory band (0.15-1 Hz) of the RIIV powerspectrum.-Respiratory Rate from RIFV (RRri f v): RRri f v was estimated as maximumvalue peak frequency in the respiratory band of the RIFV power spectrum (0.15-1Hz).-HFri f v: the power within the respiratory band (0.15-1 Hz) of the RIFV powerspectrum.-RRmean: the mean of RRriav, RRriiv and RRri f vMovement Features-artifactepoch: for each pulse of PPG, a signal quality index (SQI) was estimatedusing the cross-correlation of consecutive pulses [40], ranging from 0 to 100 (fromlow to high quality). Later, artifactepoch feature was assigned to each epoch ac-cording the following rules:artifactepoch = 0, if all pulses of the epoch have an SQI higher than 80.artifactepoch = 1, if less that four pulses of the epoch have an SQI lower than 80(less than four pulses of the epoch contaminated with artifact).artifactepoch = 2, if more than four pulses of the epoch have an SQI lower than80 (more than four pulses of the epoch contaminated with artifact).-bRatioepoch: in each epoch, PPIs shorter than 0.33 s and greater than 1.5 swere considered artifacts and labelled as the abnormal intervals. bRatioepoch wasestimated as the ratio of the number of normal intervals over the number of allintervals.-skewepoch: a measure of the symmetry of each PPG epoch (or the lack of it)61around the mean, defined as:skewepoch = \u00C2\u00B53/\u00CF\u00833/2 (4.2)where \u00C2\u00B53 and \u00CF\u0083 are the third central moment and the standard deviation of eachPPG epoch.-kurtosisepoch: a measure of the peakedness (or flatness) of each PPG epochdistribution, relative to the normal distribution, defined by:kurtosisepoch = \u00C2\u00B54/\u00CF\u00834\u00E2\u0088\u00923 (4.3)where \u00C2\u00B54 and \u00CF\u0083 are the forth central moment and the standard deviation of eachPPG epoch.4.4 Statistical Learning4.4.1 IntroductionLeast absolute shrinkage and selection operator (LASSO)Linear regression is a method for modelling the relationship between a responsevariable Y and one or more predictor variable(s), X . Linear regression assumesthat there is approximately a linear relationship between X and Y, mathematically,modelled asY = \u00CE\u00B20+\u00CE\u00B21X1+\u00CE\u00B22X2+ \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7+\u00CE\u00B2pXp+ \u00CE\u00B5, (4.4)where X j shows the jth predictor, p represents the number of predictors and \u00CE\u00B2 jis a constant quantifying the association between the predictor X j and the responseY [35].\u00CE\u00B2 values, known as the model coefficients, are estimated using least squarefitting over the training data set. Consider y\u00CB\u0086i = \u00CE\u00B2\u00CB\u00860 + \u00CE\u00B2\u00CB\u00861x1 + \u00CE\u00B2\u00CB\u00862x2 + \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7+ \u00CE\u00B2\u00CB\u0086pxp bethe prediction for Y based on the ith value of X. Then ei = yi\u00E2\u0088\u0092 y\u00CB\u0086i represents the ithresidual which is the difference between the ith observed response value and the ithresponse value that is predicted by the linear model. The residual sum of squares62Table 4.1: Description of the features extracted from PPGFeature DescriptionPulse Rate VariabilitymeanPP The mean of the PPIsSDPP The standard deviation of the PPIsRMSSD The root mean square of the difference of the successive PPIspowprv Total spectral power of PRVVLF Power of PRV in very low frequency band (0.01-0.04 Hz)nLF Normalized power of PRV in low frequency (0.04-0.15 Hz)nHF Normalized power of PRV in high frequency (0.15-1 Hz)LF/HF The ratio of low-to-high frequency power (nLF/nHF ratio)Vascular TonemeanAmp The average of the amplitude of all pulses within the epochstdAmp The standard deviation of the amplitude of all pulses within the epochmeanWidthhal f The average of the width at 50% of height of all pulses within the epochstdWidthhal f The standard deviation of the width at 50% of height of all pulses within the epochmeanWidth The average of the width at 10% of height of all pulses within the epochstdWidth The standard deviation of the width at 10% of height of all pulses within the epochmeanTimerising The mean of Timerising (the time for a pulse takes to reach its peak)stdTimerising The standard deviation of Timerising (the time for a pulse takes to reach its peak)meanTime f alling The mean of Timefalling (the time for a pulse takes to reach its troughstdTime f alling The standard deviation of Timefalling (the time for a pulse takes to reach its troughmeanSlope The mean of the rising slope of all pulses within the epochstdSlope The standard deviation of the rising slope of all pulses within the epochPWV Pulse Wave VariabilityRespiratory RateRRriav Respiratory rate obtained from respiratory-induced amplitude variation (RIAV)RRriiv Respiratory rate obtained from respiratory-induced intensity variation (RIIV)RRri f v Respiratory rate obtained from respiratory-induced frequency variation (RIFV)RRmean The mean of RRriav, RRriiv and RRri f vHFriiv The power within the respiratory band (0.15-1 Hz) of the RIIV power spectrumHFriav The power within the respiratory band (0.15-1 Hz) of the RIAV power spectrumMovementartifactepoch artifactepoch = 0, if all pulses of the epoch have an SQI higher than 80artifactepoch = 1, if less than four pulses of the epoch have an SQI lower than 80artifactepoch = 2, if more than four pulses of the epoch have an SQI lower than 80bRatioepoch The ratio of the number of normal intervals over the numberof all intervals within each epochskewepoch The measure of the symmetry of each PPG epoch (or the lack of it) around the meankurtosisepoch The measure of the peakedness of each PPG epoch relative to the normal distribution63(RSS) is define asRSS =n\u00E2\u0088\u0091i=1e2i =n\u00E2\u0088\u0091i=1(yi\u00E2\u0088\u0092 y\u00CB\u0086i)2 (4.5)orRSS =n\u00E2\u0088\u0091i=1(yi\u00E2\u0088\u0092\u00CE\u00B20\u00E2\u0088\u0092\u00CE\u00B21xi1\u00E2\u0088\u0092\u00CE\u00B22xi2\u00E2\u0088\u0092\u00C2\u00B7\u00C2\u00B7 \u00C2\u00B7\u00E2\u0088\u0092\u00CE\u00B2pxip)2 =n\u00E2\u0088\u0091i=1(yi\u00E2\u0088\u0092\u00CE\u00B20\u00E2\u0088\u0092p\u00E2\u0088\u0091j=1\u00CE\u00B2 jxi j)2 (4.6)where p is the number of the predictors and n is the number of the labelledsamples used for training the model.The least squares approach chooses \u00CE\u00B20,\u00CE\u00B21, . . . ,\u00CE\u00B2p to minimize RSS.Often it happens that in a regression model, some of the p predictors are ir-relevant. It means that they are not associated with the response. Including suchpredictors leads to unnecessary complexity in the resulting model. To obtain amodel that is more easily interpreted, it is required to exclude the irrelevant pre-dictors from the final model. In shrinkage (or regularization) approach a modelis fitted involving all p predictors using least squares but later the coefficient ofirrelevant predictors are estimated as zero.Least Absolute Shrinkage and Selection Operator (LASSO) is a shrinkagemethod that estimates the coefficients, \u00CE\u00B2 , by minimizingn\u00E2\u0088\u0091i=1(yi\u00E2\u0088\u0092\u00CE\u00B20\u00E2\u0088\u0092p\u00E2\u0088\u0091j=1\u00CE\u00B2 jxi j)2+\u00CE\u00BBp\u00E2\u0088\u0091j=1|\u00CE\u00B2 j|= RSS+\u00CE\u00BBp\u00E2\u0088\u0091j=1|\u00CE\u00B2 j| (4.7)where \u00CE\u00BB \u00E2\u0089\u00A5 0 is a tuning parameter needed to be estimated separately [35].LASSO uses the least squares fit to estimates the coefficients to get smaller RSS.The second term, \u00CE\u00BB \u00E2\u0088\u0091pj=1 |\u00CE\u00B2 j|, known as LASSO penalty, is small when some of\u00CE\u00B2 js are zero. When \u00CE\u00BB = 0, the penalty term has no effect, and LASSO regressionwill produce the least squares estimates. When \u00CE\u00BB > 0 the impact of the penaltygrows, and some of the coefficients will be estimated as zero to reduce the effectof penalty term on RSS.For each value of \u00CE\u00BB , LASSO regression will produce a different set of coeffi-cients. So it is essential to tune \u00CE\u00BB sufficiently.64Logistic regressionLogistic regression is a specific type of regression where the response variable, Y , isa categorical variable falling into one of two classes, 1 or 0, for instance. Logisticregression models the probability that Y belongs to a particular class, using thelogistic function:p(X) =e\u00CE\u00B20+\u00CE\u00B21X1+\u00CE\u00B22X2+\u00C2\u00B7\u00C2\u00B7\u00C2\u00B7+\u00CE\u00B2pXp1+ e\u00CE\u00B20+\u00CE\u00B21X1+\u00CE\u00B22X2+\u00C2\u00B7\u00C2\u00B7\u00C2\u00B7+\u00CE\u00B2pXp, (4.8)where p(X) = Pr(Y = 1|X).Generally, a logistic regression model is fitted using a method called Maximumlikelihood based on the available training data. In this method, \u00CE\u00B2 coefficients areestimated such that the predicted probability p(xi) of xi corresponds as closely aspossible to the observed yi. Maximum likelihood is mathematically formalized as:l(\u00CE\u00B20,\u00CE\u00B21,\u00CE\u00B22, . . . ,\u00CE\u00B2p) = \u00E2\u0088\u008Fi:yi=1p(xi) \u00E2\u0088\u008Fi\u00E2\u0080\u00B2:yi\u00E2\u0080\u00B2=0(1\u00E2\u0088\u0092 p(xi\u00E2\u0080\u00B2)) (4.9)The estimates of \u00CE\u00B20,\u00CE\u00B21,\u00CE\u00B22, . . . ,\u00CE\u00B2p are chosen to maximize this likelihood func-tion.Decision threshold estimationAs mentioned in the previous section, Logistic regression estimated the probabilitythat Y belongs to a specific class. Later, the estimated probability is tested againsta decision threshold, \u00CF\u0084 , to assign Y into one of two classes, 0 or 1 (negative orpositive). The sensitivity, specificity, and accuracy of the model depend on thethreshold \u00CF\u0084 .For a given decision threshold, the performance of a classifier can be summa-rized by a 2 \u00C3\u0097 2 confusion matrix (Table 4.2). For each decision threshold, thesensitivity, specificity, and accuracy are estimated asSN(\u00CF\u0084) =T P(\u00CF\u0084)n1(4.10)SP(\u00CF\u0084) =T N(\u00CF\u0084)n0(4.11)65Table 4.2: The performance of a binary classifier is summarized by a 2 \u00C3\u0097 2confusion matrix for a given decision threshold \u00CF\u0084 . TN: number of truenegative, FP: number of false positive, FN: number of false negative, TP:number of true positive, PN: number of predicted negative, PP: numberof predicted positivePredicted Predicted total0 1True 0 TN(\u00CF\u0084) FP(\u00CF\u0084) n0True 1 FN(\u00CF\u0084) TP(\u00CF\u0084) n1Total PN(\u00CF\u0084) PP(\u00CF\u0084) nACC(\u00CF\u0084) =T P(\u00CF\u0084)+T N(\u00CF\u0084)n0+n1(4.12)where n0 denotes the number of 0 samples (negative samples) and n1 denotesthe number of 1 samples (positive samples). TP(\u00CF\u0084) and TN(\u00CF\u0084) are the numbers ofcorrect predictions for the 1 and 0 samples, respectively.The default value of the decision threshold is 0.5. When the class sample sizes(n0 and n1) are almost equal, a classifier using the default threshold provides anunbiased estimate of the sensitivity, specificity, and accuracy. But, when the classsizes are different, a classifier using the default threshold may provide an unac-ceptably low sensitivity (or specificity). So it is essential to estimate the decisionthreshold for each classifier properly.4.4.2 Multivariate model development and validationModel developmentSubjects were randomly divided into training and test sets. The epochs correspond-ing to the subjects in the training set were used to train the classifiers, and theepochs corresponding to each subject in the test set were fed to the trained modelsto the validate the performance of classifiers.In the training phase, to classify each epoch into one of the three classes ofawake, REM and non-REM (known as multi-class classification problem), a hier-archical binary classifier with two nodes was developed (Figure 4.1). Each node66Figure 4.1: The multi-class classifier has two binary classifiers: sleep/wakeclassifier and non-REM/REM classifier. The epochs corresponding tothe subjects in the training set were used to train these two classifiers,and the epochs corresponding to each subject in the test set were fed tothe trained models.corresponding to a binary multivariate logistic regression classifier trained usingthe training set: 1) the wake/sleep classifier to determine whether an epoch wouldbe scored as sleep or wake, and 2) the non-REM/REM classifier to determine thesleep epoch whether would be scored as REM or non-REM.LASSO was employed to select the relevant features and to develop the finalwake/sleep and non-REM/REM classifiers (using the glmnet R package). The tun-ing parameter was adjusted through a stratied 10-fold cross validation. For eachepoch, the final models estimated the probability of belonging to a certain class.Decision threshold determinationThe decision thresholds were separately chosen for two classifiers to maximize aweighted classification score defined as (TP(\u00CF\u0084) + TN(\u00CF\u0084)). The weighted classifica-tion score was computed for various previously established ratios of false positivecases to false negative cases (3:1, 5:1 and 10:1).Model classification performanceIn the validation phase, all the epochs of a subject from the test dataset were firstfed to the wake/sleep classifier. Later, the epochs scored as sleep by wake/sleep67classifier were fed to the REM/non-REM classifier to distinguish between REMand non-REM epochs (Figure 4.1). By combining the results from sleep/wake andREM/non-REM classifiers, each epoch of individual subject in the testing datasetwas scored as wake, REM and non-REM. The scored epochs were aligned togetherto predict a hypnogram for each subject.To validate the performance of wake/sleep and REM/non-REM classifiers, theaccuracy, sensitivity and specificity measures were calculated. In addition, thepredicted hypnogram for each subject was compared with the hypnogram extractedfrom PSG event log file and an individual accuracy measure was calculated foreach subject as the percentage of true classifications of wake, REM and non-REMof total epochs according to:accuracyindividual =true wake+ true REM+ true non\u00E2\u0088\u0092REMtotalwake+ totalsleep(4.13)The general performance of the model was then assessed using the distributionof the accuracy individual of all subjects in the testing dataset through the mean and95% confidence intervals (CI) of the median. These estimations were performedusing the bootstrap method; 100 bootstrap samples were generated using the orig-inal accuracy, sensitivity, and specificity data through sampling with replacing.4.5 ResultsThe data set of 146 subjects was randomly divided into the training and test datasets with 46 and 100 subjects, respectively.4.5.1 Wake/sleep classifierThe wake/sleep classifier was trained using the training set including 38,098 epochsscored as sleep (27,885) and wake (10,213) based on the PSG event log file.For each epoch, 31 features were extracted from the PPG signal (Table 4.1);among them, 15 features were selected as significant by LASSO method based on\u00CE\u00BB = 9.408e-05 (Table 4.3). To choose the best \u00CE\u00BB , the cross-validation error foreach value of \u00CE\u00BB was estimated.We then selected the \u00CE\u00BB value for which the cross-68Table 4.3: Estimated coefficient and error for 15 features selected withLASSO as the significant features for wake/sleep modelEstimated EstimatedModel Feature Coefficient Error p-valuewake/sleepModel(Intercept) -6.70 1.28 1.80e-07meanTime f alling 0.15 0.03 8.96e-06stdWidthhal f -0.10 0.01 7.34e-12pwv -0.41 0.10 4.82e-05HFriav -0.64 0.11 1.03e-08nHF 3.90 0.22 < 2e-16meanWidthhal f 0.16 0.01 < 2e-16RMSSD -6.87 1.71 5.68e-05stdRRI 10.01 2.50 6.10e-05meanTimerising 0.07 0.03 0.01meanRRI -3.36 1.63 0.04skewepoch 0.11 0.04 0.003RRri f v -0.04 0.004 < 2e-16HFriiv 0.93 0.12 5.39e-14RRriiv -0.03 0.005 2.33e-12stdTime f alling -0.06 0.016 0.000266validation error was smallest.This model presented an AUC of 0.85 wih the 95% confidence interval from0.84 to 0.87 (Figure 5.2a).The decision threshold was estimated as \u00CF\u0084 = 0.725. The accuracy, sensitivityand specificity values were estimated 0.82, 0.85 and 0.79, respectively, in trainingdata set for estimated \u00CF\u0084 .The accuracy, sensitivity and specificity values were estimated 0.77, 0.77 and0.79, respectively, in test data set for \u00CF\u0084 = 0.725.4.5.2 non-REM/REM classifierThe training data set contained 27,885 epochs scored as sleep. Among theseepochs, 22,590 and 5,295 entries were scored as non-REM and non-REM, respec-tively, based on the PSG event log file.Each epoch contains 31 features extracted from the PPG signal (Table 4.1).69(a)(b)Figure 4.2: The area under the curve (AUC) of the receiver operating char-acteristic (ROC) curve of a) the wake/sleep classifier and b) the non-REM/REM classifierThe REM/non-REM classifier selected 16 significant features (Table 4.4).This model presented an AUC of 0.77 wih the 95% confidence interval from70Table 4.4: Estimated coefficient and error for 16 features selected withLASSO as the significant features for non-REM/REM modelEstimated EstimatedModel Feature Coefficient Error p-valuenon-REM/REMModel(Intercept) 5.04 0.8 1.77e-09meanTime f alling -0.03 0.004 1.26e-12HFriiv -0.31 0.006 3.18e-06RRmean -0.02 0.008 0.000254meanWidthhal f 0.02 0.005 1.33e-06stdTime f alling -0.02 0.01 0.049398HFriiv -1.70 0.11 < 2e-16bRatioepoch -4.50 0.81 3.50e-08meanAmp -11.20 3.58 0.000948meanSlope 94.26 30.10 0.002171pwv 0.78 0.06 < 2e-16kurtosisepoch -0.02 0.005 0.000168RRriiv 0.02 0.005 8.58e-05skewepoch 0.1 0.03 0.0004HFriav 0.22 0.06 0.0013stdWidthhal f 0.05 0.01 9.17e-06stdWidth -0.12 0.011 < 2e-160.74 to 0.79 (Figure 5.2b).The decision threshold was estimated as \u00CF\u0084 = 0.19. The accuracy, sensitivityand specificity values were estimated 0.72, 0.70 and 0.73, respectively, in trainingdata set for estimated \u00CF\u0084 .The accuracy, sensitivity, and specificity values were estimated 0.73, 0.71 and0.73, respectively, in test data set for \u00CF\u0084 = 0.19.4.6 Discussion and ConclusionThe results of this study show that extracting the pediatric hypnogram, similar tothe one provided by PSG, based on the characterization of cardiovascular activityperformed using the overnight Phone Oximeter PPG will be practical, achievableand reliable.The most discriminant features for sleep staging were automatically selected71by a shrinking method, LASSO, which forces the coefficient estimates to be ex-actly equal to zero. During training, two different sets were selected with 15 and16 discriminant features out of 33 features, for wake/sleep and non-REM/REMclassifiers, respectively (Table 4.3 and Table 4.4). These two models were validatedseparately: the classification of sleep from wake showed the mean accuracy of 73%while the non-REM/REM model reached the mean accuracy of 69%. Later, thesetwo classifiers were combined together as a hierarchy model to classify epochs intothree classes of wake, REM and non-REM. In the validation phase, all the epochswere first fed to the wake/sleep classifier and then, the epochs scored as sleep werefed to the REM/non-REM classifier. It implies that the misclassified epochs wouldtransfer from wake/sleep classifier to the non-REM/REM classifier, which woulddegrade the overall performance of the model.About 38 of children participated in this study were diagnosed with SBDs withthe AHI more than five. In our previous study performed on the same dataset, weshowed that SBD modulates the sympathetic cardiac activity in both REM and non-REM sleep. However, our results indicate that the predominance of sympatheticactivity in A/H events is suppressed by cardiac sympathetic modulation duringREM sleep. Besides, the children with SBD are more prone to frequent arousals,most of the time associated with movement, which affects the cardiac regulation ofthe autonomic nervous system. These all cause sleep staging more challenging inchildren with SBD.Our results, obtained with the Phone OximeterTM, are comparable with previ-ous studies with more sophisticated approaches or devices for sleep staging basedon monitoring the activity of the cardiorespiratory system. In a recent study, Ucaret al [73] extracted 86 features from PPG recorded from 10 adult patients and usedk-nearest neighbors classification and support vector machines to identify sleepfrom wakefulness. The accuracy, sensitivity, and specificity of trained model werereported as 73.36 %, 0.81% and 0.77%, respectively.Yilmaz et al [77] successfully extracted the hypnogram for 17 adult individualswith SBD using the features extracted from ECG. The total accuracy of 73% wasreported for the one-vs-rest approach whose classifiers trained by the support vec-tor machines. However, in this study, a separate model has been trained for eachsubject, which reduces the feasibility of this approach in real clinical applications.724.6.1 Limitation of study and future workThe most challenging part of classification was choosing the decision thresholdsfor wake/sleep and non-REM/REM classifiers. The decision thresholds were cal-culated to maximize the accuracy of classifiers based on the percentage of theepochs scored as wake, REM and non-REM in the training dataset. During val-idation, the same decision threshold used for all subjects. Since each subject hasa unique pattern of sleep with the different shares of the wake, REM and non-REM stages, using the same threshold for all subjects decreases the performanceof classification. Therefore, we will further investigate the possibility of estimatinga separate risk threshold for each subject based on the quality and patterns of theirsleep by measuring sleep latency, sleep duration, habitual sleep efficiency, sleepdisturbances through self-assessment questionnaires.73Chapter 5Development of a MonitoringTool for Sleep DisorderedBreathing in Children Using thePhone Oximeter5.1 IntroductionAs mentioned in chapter 1, the high prevalence of A/H syndrome among chil-dren and adolescents and the compexity and hight cost of PSG have generateda great interest in alternative techniques to simplify the standard procedure. Al-ready part of the standard PSG, pulse oximetry is a simple non-invasive methodof measuring SpO2 and recording PPG. Numerous groups have studied the use ofovernight oximetry as a potential standalone method to diagnose SDB. Nixon etal. developed a severity scoring system using overnight oximetry and validated thescore as a tool to prioritize adenotonsillectomy surgeries [31], [58]. A\u00C2\u00B4lvarez et al.demonstrated that the characterization of overnight oximetry provided significantinformation to identify adults [49], [50] with significant OSA. Both studies focusedon SpO2 alone; however, there are some SDB events that occur in the absence ofSpO2 desaturation [78]. It has been reported that SDB affects the normal variation74of heart rate [55], [14], [20] suggesting that combining SpO2 and HRV analysismight provide a more robust SDB detector. Based on this concept, Heneghan et al.proposed a portable, automated OSA assessment tool with a Holter-Oximeter [30],[12].In our previous research, we showed that the characterization of overnightSpO2 pattern, measured by the Phone OximeterTM , successfully identifies chil-dren with significant SDB [25]. We also investigated the influence of SpO2 resolu-tion (0.1%, 1%) on the SpO2 pattern characterization and demonstrated that it hada great influence in regularity measurements and therefore should be consideredwhen studying SDB [27]. In addition, we calculated PRV from the Phone Oxime-ter\u00E2\u0080\u0099s PPG and compared it with HRV computed from simultaneous electrocardio-gram (ECG) [18], [20]. In the time domain, PRV provided accurate estimates ofHRV, while some differences were found in the frequency domain. Gil et al. alsoshowed that during non-stationary conditions there are some small differences be-tween HRV and PRV, mainly in the respiratory band, which were related to thepulse transit time variability [29]. However, they also concluded that these differ-ences are sufficiently small to suggest the use of PRV as an alternative measure ofHRV. We also conducted an additional investigation of the effects of SDB on PRVduring different sleep stages and concluded that the modulation of PRV might behelpful in improving the assessment of SDB in children [19].In our recent study, therefore, we combined the SpO2 pattern characteriza-tion and PRV analysis to identify the epochs with A/H events using the PhoneOximeterTM [28]. We recorded overnight SpO2 and PPG using the Phone OximeterTM, simultaneously with standard PSG from 160 children at the British ColumbiaChildrens hospital. The sleep technician manually scored all apnea/hypoapneaevents during the PSG study. Based on these scores we labeled each epoch asA/H epochs or non-A/H epochs. We randomly divided the subjects into trainingdata, used to develop the model applying the LASSO method, and the test data,used to validate the model. The developed model was assessed epoch-by-epochfor each subject. The model provided a median accuracy of 74%, sensitivity of75%, and specificity of 73% when using a risk threshold similar to the percentageof A/H epochs.However, we realized that more than 32% of epochs from the original database75had been excluded from the further analysis due to the poor quality of SpO2 orPPG. Among the total number of 134389 epochs labelled by a sleep technician,more than 30% had the low-quality PPG while only less than 2% had the low-quality SpO2. The purpose of study, presented in this chapter, is then to reducethe number of the excluded epochs. To reach this goal, we propose a method foridentifying the A/H epochs based on two trained models: one model is trained toidentify A/H epochs using the combined characteristic of SpO2 and PPG whereboth PPG and SpO2 epochs have high quality and the second model uses the SpO2characteristics for epochs with the low-quality PPG but the high-quality SpO2. Theresults of these two models in predicting the A/H epochs would be combined to getthe final prediction.5.2 Materials and Methods5.2.1 Apnea/Hypopnea LabellingThe same data set described in Chapter 2 (2.2.1) was used for this study. A sleeptechnician visually scored the PSG in 30-second epochs according to AASM 2007standard criteria [5]. Hypnograms were differentiated into rapid eye movement(REM) and non-REM sleep. According to the standard criteria, obstructive apneaswere defined as complete cessation of airflow in the presence of respiratory effortlasting seconds. Hypopneas were defined as a airflow reduction relative to the 2preceding breaths. Blood oxygen desaturations were defined as a decrease in arte-rial oxygen saturation. When respiratory effort partially or totally ceased, apneaswere scored as mixed or central sleep apnea, respectively. The number of A/Hevents was counted hourly to compute the average apneas/hypopnea index (AHI),which was specified also for REM and non-REM (NREM) sleep stages. The totalbed time (TBT), total sleep time (TST) and the percentage of time spent in thedifferent sleep stages were also analyzed (Table 2.1).The Phone OximeterTM recordings (SpO2 and PPG signals) were segmentedinto epochs of 30-seconds duration. All epochs were labelled as the A/H or non-A/H epochs using the scores performed by the sleep technician based on the PSGstudy.765.2.2 PPG Features ExtractionFor each 30-second epoch of the PPG signals, the following features were extracted(Table 5.1):Signal Quality Index of PPGA simple peak detection algorithm based on zero-crossing was applied the PPGsignals to locate the pulse peaks. The peak locations were used to segment thePPG into the pulses. An algorithm iteratively calculated a signal quality index(SQI) ranging from 0 to 100 for each pulse. Cross-correlation of consecutive pulsesegments is used to estimate signal quality. In the presence of artifacts and irregularsignal morphology, the algorithm outputs a low SQI number.If all pulses of an epoch have an SQI higher than 80, the feature artppg was setas 0. If less than four pulses of the epoch have an SQI lower than 80, the artppgwas set as 1 and if more than four pulses of the epoch have an SQI lower than 80,the artppg was set as 2 (Table 5.1).PRV FeaturesTo analyze PRV, in each epoch, the pulse-to-pulse intervals time series (PPIs) werecomputed as the intervals between successive peaks. In the time domain, threefeatures were extracted from the PPIs time series:-meanPP: the mean of the PPIs-SDPP: the standard deviation of the PPIs,-RMSSD: the root mean square of the difference of the successive PPIsto get PRV, the PPIs were resampled into the equivalent, uniformly spacedtime series at a sampling rate of 4 Hz using the Berger algorithm [8]. Then thepower spectral density of PRV was estimated using a parametric autoregressivemodel with 1024 points and an order of 7. The power in each frequency band wascomputed by calculating the area under the PSD curve bounded by the band ofinterest and the following features were extracted:-Pow: the total spectral power of the PRV-VLF: power in the very low frequency (0.01-0.04 Hz)-nLF: the LF power was estimated as the power in the low frequency band77(0.04-0.15 Hz). Later, normalized LF (nLF) was calculated by dividing LF powerby the total spectral power of PRV between 0.04 and 0.4 Hz.-nHF: the HF power was estimated as the power in the high frequency band(0.15-1 Hz). Later, normalized HF (nHF) was calculated by dividing HF power bythe total spectral power of PRV between 0.04 and 0.4 Hz.-LF/HF: the ratio of low-to-high frequency powerVascular Tone FeaturesSeveral morphology features were extracted from each PPG pulse to characterizethe vascular tone in epochs with and without A/H events.-meanAmp and stdAmp: the amplitude of each pulse was measured as thedifference between the maximum of a pulse (peak) and the previous minimum(trough). meanAmp and stdAmp were calculated as the average and standard de-viation of the amplitude of all pulses within the epoch, respectively.-meanWidthhal f and stdWidthhal f : the widthhal f of each pulse was calculatedas the width at 50% of pulse height; later, meanWidthhal f and stdWidthhal f werecalculated as the average and standard deviation of widthhal f of all pulses withinthe epoch, respectively.-meanWidth and stdWidth: the widthpulse of each pulse was calculated asthe width at 10% of pulse height; later, meanWidth and stdWidth were calculatedas the average and standard deviation of widthpulse of all pulses within the epoch,respectively.-meanTimerising and stdTimerising: the mean and standard deviation of Timerising(the time for a pulse takes to reach its peak) of all pulses within the epoch werecomputed.-meanTime f alling and stdTime f alling: the mean and standard deviation of Time f alling(the time for a pulse takes to reach its trough) of all pulses within the epoch werecomputed.-meanSlope and stdSlope: the mean and standard deviation of rising slope ofall pulses within the epoch-pwv: For each epoch, pulse wave variability (pwv) was estimated as:78pwv =max(amp)\u00E2\u0088\u0092min(amp)(max(amp)+min(amp))/2(5.1)5.2.3 SpO2 Features ExtractionFor each 30-second epoch of the SpO2 signals, the following features were ex-tracted (Table 5.2):SQI of SpO2All SpO2 values below 50% and above 100%, and the SpO2 changes betweenconsecutive sampling intervals greater than 4%, were considered as artifacts. Thefeature artspo2 was set as 0 if less than 50% of the SpO2 epoch is contaminatedby artifacts. The artspo2 was set as 1 if more than 50% of the SpO2 epoch wascontaminated by artifacts.Time Domain FeaturesSeveral oximetry indices proposed in previous studies to assess SpO2 dynamics inthe time domain, were computed here [25]:-Tr2: the number of SpO2 desaturations greater than 2% below baseline-Tr3: the number of SpO2 desaturations greater than 3% below baseline-ind96: the cumulative time spent below an SpO2 of 96%-ind94 the cumulative time spent below an SpO2 of 94%-Delta: the Delta index quantifies SpO2 variability and was computed as theaverage of absolute differences of the mean oxygen saturation between successive12-sec intervals.-SDSpO2: the standard deviation of the SpO2 within each epoch-IQR: the interquartile range of the SpO2 within each epoch-CTM: the Central Tendency Measure is a non-linear method that providesquantitative variability information was also applied to SpO2 epochs [48]79Spectral Domain FeaturesThe SpO2 signal was characterized in the spectral domain using power spectraldensity (PSD). To provide better frequency resolution, a parametric PSD was per-formed approximating the SpO2 signal through an autoregressive model using:SpO2 =\u00E2\u0088\u0092p\u00E2\u0088\u0091k=1ak.SpO2(n\u00E2\u0088\u0092 k)+ e(n) (5.2)where e(n) denotes zero-mean white noise with variance \u00CF\u00832e , ak the autoregres-sive coefficients and p the model order. Once the autoregressive coefficients andthe variance was estimated, the PSD of the autoregressive model was computed by:PSD( f ) =\u00CF\u00832e|1+\u00E2\u0088\u0091pk=1 ak.e\u00E2\u0088\u0092 j2pi f kT |2(5.3)with 1/T as the sampling frequency.The sleep apnea events happen in a pseudo periodic pattern, which modulatesthe SpO2 signal and provokes a modulation frequency peak at very low frequencyband. A significant power increase in a frequency band ranging from 0.014 to0.033 Hz was previously documented in subjects suffering from sleep apnea, dueto the modulation provoked by continuous oxygen desaturations [50]. Therefore,the following features were extracted from the PSD:-powSpO2: total spectral power of SpO2-modPow: the total power in modulation band (0.005 Hz to 0.12 Hz)-meanPow: the mean power in modulation band (0.005 Hz to 0.12 Hz)-PRatio: the ratio between the power frequency band from 0.005 to 0.12 andtotal power-powDiscr: the power in the discriminant frequency band, defined as a fre-quency interval (0.02 Hz) centered on the modulation frequency peak detected inmodulation band-SEfreq: the Shannon entropy of the power spectrum density SpO280Table 5.1: Description of the features extracted from PPGFeature DescriptionSQIartppg artppg = 0, if all pulses of the epoch have an SQI higher than 80artppg = 1, if less than four pulses of the epoch have an SQI lower than 80artppg = 2, if more than four pulses of the epoch have an SQI lower than 80Pulse Rate VariabilitymeanPP The mean of the PPIsSDPP The standard deviation of the PPIsRMSSD The root mean square of the difference of the successive PPIspow Total spectral power of PRVVLF Power of PRV in very low frequency band (0.01-0.04 Hz)nLF Normalized power of PRV in low frequency (0.04-0.15 Hz)nHF Normalized power of PRV in high frequency (0.15-1 Hz)LF/HF The ratio of low-to-high frequency power (nLF/nHF ratio)Vascular TonemeanAmppulse The average of the amplitude of all pulses within the epochstdAmppulse The standard deviation of the amplitude of all pulses within the epochmeanWidthhal f The average of the width at 50% of height of all pulses within the epochstdWidthhal f The standard deviation of the width at 50% of height of all pulses within the epochmeanWidthpulse The average of the width at 10% of height of all pulses within the epochstdWidthpulse The standard deviation of the width at 10% of height of all pulses within the epochmeanTimerising The mean of Timerising (the time for a pulse takes to reach its peak)stdTimerising The standard deviation of Timerising (the time for a pulse takes to reach its peak)meanTime f alling The mean of Timefalling (the time for a pulse takes to reach its troughstdTime f alling The standard deviation of Timefalling (the time for a pulse takes to reach its troughmeanSlope The mean of the rising slope of all pulses within the epochstdSlope The standard deviation of the rising slope of all pulses within the epochPWV Pulse Wave Variability81Table 5.2: Description of the features extracted from SpO2Feature DescriptionSQIartspo2 artspo2 = 0, if less than 50% of the SpO2 epoch is contaminated by artifactsartspo2 = 1, if more than 50% of the SpO2 epoch is contaminated by artifactsAll SpO2 values below 50% and above 100%, and the SpO2 changes betweenconsecutive sampling intervals greater than 4% are considered as artifactsTime domain featuresSDSpO2 The standard deviation of SpO2 within each epochIQR The interquartile range of SpO2 within each epochDelta the average of absolute differences of the mean oxygen saturationbetween successive 12-sec intervalsind96 The cumulative time spent below an SpO2 of 96%ind94 The cumulative time spent below an SpO2 of 94%Tr2 The number of SpO2 desaturations greater than 2% below baselineTr3 The number of SpO2 desaturations greater than 3% below baselineCTM Central tendency measureSpectral domain featurespowSpO2 Total spectral power of SpO2modPow The total power in modulation band (0.005 Hz to 0.12 Hz)meanPow The mean power in modulation band (0.005 Hz to 0.12 Hz)PRatio The ratio between the power frequency band from 0.005 to 0.12 and total powerpowDiscr The power in the discriminant frequency band,defined as a frequency interval (0.02 Hz)centered on the modulation frequency peak detected in modulation bandSEfreq The Shannon entropy of the power spectrum density SpO25.2.4 Data AnalysisUnivariate AnalysisAll the epochs with low-quality SpO2(artspo2 = 1) were excluded from the origi-nal database and further analysis. For the rest of the epochs, the relationship be-tween each feature and the presence of A/H event(s) was assessed by comparingthe median value of the feature in the A/H and non-A/H epochs and also throughunivariate logistic regression using the OR (95% CI).82Multivariate Model DevelopmentSubject were randomly divided into training and test sets. The epochs correspond-ing to the subjects in the training set were used to train the classifiers, and theepochs corresponding to each subject in the test set were fed to the trained modelsto validate the performance of classifiers.In the training phase, all epochs of the training set were organized into twoseparate databases: 1) database1 including all epochs with high-quality PPG (artppg= 0 or artppg = 1) and high-quality SpO2 (artspo2 = 0) and 2) database2 includingall epochs with low-quality PPG (artppg = 2) and high-quality SpO2 (artspo2 = 0).For the epochs in database1, both sets of the PPG and SpO2 features wereextracted while for epochs in database2 only the SpO2 features were estimated.To classify each epoch into two classes of A/H and non-A/H a model with twobinary multivariate logistic regression classifiers was developed (Figure 5.1). Thefirst classifier was trained over the datbase1 and the second classifier was trainedover the database2.LASSO was employed to select the relevant features and to develop the classi-fiers (using the glmnet R package). The tuning parameter was adjusted through astratified 10-fold cross-validation. For each epoch, the final models estimated theprobability of belonging to a certain class.Model Classification PerformanceThe validation of the model was then performed for each subject within the testdataset, epoch-by-epoch. Individual classification results were represented usingthe area under the receiver operating characteristic (ROC) curve (AUC), accuracy,sensitivity and specificity classifying epochs with and without A/H event(s). Thegeneral performance of the model was then assessed using the distribution of theAUC, accuracy, sensitivity and specificity, for the subjects in the testing datasetthrough the mean and 95% confidence intervals (CI) of the quartiles (25, 50 [me-dian] and 75 percentile). These estimations were performed using the bootstrapmethod; 100 bootstrap samples were generated using the original AUC, accuracy,sensitivity and specificity data, through random sampling with replacement.83Figure 5.1: The proposed model has two binary classifiers. The epochs cor-responding to the subjects in the training set were used to train thesetwo classifiers, and the epochs corresponding to each subject in the testset were fed to the trained classifiers.845.3 ResultsFourteen children were excluded from analysis based on having a total sleep dura-tion, or signal data duration (from PSG or the smartphone-based pulse oximeter)shorter than 3 hours. The original dataset then included the total of 134389 epochs.The number of 1602 (about 1.1%) of epochs had a very low-quality SpO2 (artspo2= 1) and were excluded from the further analysis.5.3.1 Univariate AnalysisFor the number of 99,736 epochs with high-quality PPG and SpO2, all the PPG andSpO2 features were extracted. PPG and SpO2 derived features were significantlydifferent in epochs with A/H event(S) compared to those without A/H epochs (Ta-ble 5.3).For the number of 33,051 epochs with low-quality PPG and high-quality SpO2,the SpO2 features were extracted. SpO2 derived features were significantly differ-ent in epochs with A/H event(S) compared to those without A/H epochs (Table5.4).5.3.2 Multivariate Model ValidationThe LASSO method has a tuning parameter (lambda) controlling the degree ofoverfitting. This parameter was determined by minimizing the 10-fold cross-validatedprediction error of the model, created using only the training data. The significantfeatures were then selected based on the chosen lambda. The final logistic regres-sion model selected 12 PPG and SpO2 features for the first A/H classifier (Table5.5). All the selected features had p-values < 0.0001. This model presented anAUC of 0.85 (95% CI: 0.82 - 0.87) (Figure 5.2 a). For the second A/H classifier,5 SpO2 features were chosen (Table 5.6). All the selected features had p-values <0.0001 except for Tr2 that had a p-value of 0.05. This model presented an AUC of0.77 (95% CI: 0.75 - 0.79) (Figure 5.2 b).By combing the results from two A/H classifiers, each epoch of individual sub-ject in the test dataset was assigned with the probability of containing the A/Hevent(s). To optimize the sensitivity detecting epochs with A/H based on the pre-dicted probabilities, we used a decision threshold of 0.035, which was similar to85Table 5.3: Distribution of features extracted from PPG and SpO2 for A/H andnon-A/H epochsFeatures A/H epochs non-A/H epochs p-value ORPulse Rate VariabilitymeanPP 0.71 0.74 < 2e-16 0.40SDPP 0.08 0.05 < 2e-16 14.83RMSSD 0.07 0.06 < 2e-16 27.12pow 1.43 0.66 < 2e-16 1.15VLF 0.10 0.02 < 2e-16 5.20nLF 0.34 0.15 < 2e-16 13.51nHF 0.65 0.85 < 2e-16 0.07LF/HF Ratio 0.53 0.17 < 2e-16 1.17Vascular TonemeanAmppulse 0.008 0.009 2e-06 99.66stdAmppulse 0.002 0.001 < 2e-16 86.8meanWidthhal f 23.83 23.50 3e-08 1.01stdWidthhal f 5.81 4.15 < 2e-16 1.1meanWidthpulse 38.6 40.49 < 2e-16 0.98stdWidthpulse 6.38 4.20 < 2e-16 1.08meanTimerising 11.54 10.67 < 2e-16 1.09stdTimerising 2.50 0.87 < 2e-16 1.06meanTime f alling 35.03 37.02 < 2e-16 0.98stdTime f alling 5.96 3.98 < 2e-16 1.08meanSlope 0 0 2e-14 3.64stdSlope 0 0 < 2e-16 4.40PWV 1.12 0.73 < 2e-16 3.62SpO2RRatio 0.54 0.44 < 2e-16 8.74modPow 0.14 0.02 < 2e-16 1.00powSpO2 0.28 0.04 < 2e-16 1.001SEfreq 7.03 7.37 < 2e-16 0.69meanPow 0.004 0 < 2e-16 1.01SDSpO2 0.64 0.25 < 2e-16 3.20IQR 0.88 0.32 < 2e-16 2.047Delta 0.26 0.09 < 2e-16 5.34ind96 2 0 < 2e-16 1.04CTM3 0.96 1 < 2e-16 0.00486Table 5.4: Distribution of features extracted from SpO2 for A/H and non-A/HepochsFeatures A/H epochs non-A/H epochs p-value ORSpO2RRatio 0.55 0.42 < 2e-16 23.43modPow 0.16 0.02 < 2e-16 1.03powSpO2 0.38 0.04 < 2e-16 1.04SEfreq 6.98 7.42 < 2e-16 0.54meanPow 0.004 0.00 < 2e-16 2.23SDSpO2 0.70 0.25 < 2e-16 2.73IQR 0.97 0.32 < 2e-16 1.77Delta 0.27 0.09 < 2e-16 4.50ind96 1 0 < 2e-16 1.05CTM3 0.92 1 < 2e-16 0.005the percentage of A/H epochs in our training data. The median AUC was 75% andusing the selected risk threshold, the accuracy, sensitivity and specificity valuesobtained for the subjects in the testing dataset were around 74%; even the subjectsat lowest quartile of the accuracy, sensitivity and specificity provided values above65% (Table 5.7).The model performed well identifying A/H epochs (Figure 5.3 a). However,some subjects showed unbalanced sensitivity-specificity values, with too low speci-ficity values (Figure 5.3 b and Figure 5.3 c), as a result of prioritizing model\u00E2\u0080\u0099s sen-sitivity. The most challenging cases corresponded to subjects containing low A/Hevents per night.5.4 Discussion and ConclusionThis study showed that combining the SpO2 pattern characterization and PPG anal-ysis performed using the Phone Oximeter\u00E2\u0080\u0099s measurements (SpO2 and PPG), im-proved the Phone Oximeter\u00E2\u0080\u0099s performance as a possible SDB screening tool. Inaddition, having another model trained based on the SpO2 features alone would in-clude the epochs whose PPG signal was contaminated with the artifact while theirSpO2 signals was still reliable (about 30% of the epochs in our database).87Table 5.5: Estimated coefficient and error for 12 features selected withLASSO as the significant features for A/H model trained over database1(including PPG and SpO2 features)Estimated EstimatedModel Feature Coefficient Error p-valueA/HModel(Intercept) -5.392e+00 2.981e-01 < 2e-16SDPP -1.187e+01 9.425e-01 <2e-16nLF 1.119e+00 1.169e-01 < 2e-16stdAmppulse -3.235e+01 5.344e+00 1.42e-09meanWidthhal f 4.098e-02 3.829e-03 < 2e-16stdWidthhal f -1.979e-02 9.165e-03 0.03080PWV 1.092e+00 7.145e-02 < 2e-16RRatio 1.120e+00 1.202e-01 < 2e-16SDSpO2 1.272e-01 4.285e-02 0.00300IQR 1.272e-01 4.285e-02 0.00300ind96 1.692e-02 2.332e-03 3.94e-13Tr2 2.311e-01 7.764e-02 0.00292CTM3 -1.186e+00 2.602e-01 5.14e-06Table 5.6: Estimated coefficient and error for 5 features selected with LASSOas the significant features for A/H model trained over database2 (includ-ing SpO2 features)Estimated EstimatedModel Feature Coefficient Error p-valueA/HModel(Intercept) -3.67 0.41 < 2e-16RRatio 2.17 0.18 < 2e-16SDSpO2 2.17 0.182343 < 2e-16ind96 0.017 0.002 3.94e-13Tr2 0.23 0.13 0.067CTM3 -1.00 0.41 0.01488Table 5.7: Classification results from test set represented by the mean and95% CI of the quartiles of Accuracy(Acc), Sensitivity (Sn), Speci-ficity(Sp)and the area of the ROC curve (AUC)Validation(Test set) Acc(%) Sn(%) Sp(%) AUC(%)25 Percentile 69 [66, 72] 65 [60, 70] 66 [63, 70] 73 [71, 76]50 Percentile 76 [73, 78] 72 [67, 78] 75 [72, 80] 77 [72, 81]75 Percentile 81 [79, 84] 84 [78, 89] 80 [76, 84] 80 [77, 83]The most discriminating features identifying epochs with A/H event(s) wereautomatically selected by LASSO. The selected features were related mainly to thespectral analysis of PRV, PPG pulse amplitude and width variability, SpO2 variabil-ity and modulation represented in the spectral domain. This reflects the significanteffect of intermittent apnea events and respiratory arousals in the sympathetic andparasympathetic activity, and the recurrent desaturations in the SpO2 pattern vari-ability. The validation results, obtained for each subject within the testing dataset,provided a median AUC of 77% identifying epochs with sleep A/H event(s).Our results, obtained with the Phone OximeterTM , are comparable with previ-ous studies with more sophisticated approaches or devices. Heneghan et al. pro-posed a combined Holter-Oximeter as a portable home-based device to automati-cally assess OSA in adults with signs of SDB [30], [12]. Their system providedan automatic epoch-by-epoch estimate of OSA occurrence and calculated an AHIfor each subject. Overall the system correctly identified 85.3% of all 1-minuteepochs. Chung et al. reported that oxygen desaturation index (ODI), calculatedfrom nocturnal oximetry, was a good predictor of AHI in adult surgical patients[15]. An ODI provided an accuracy of 87%, sensitivity of 96.3% and specificityof 67.3% identifying adults with an AHI . In this study, we focused on identifyingA/H epochs in children, which is more challenging than in their adult counterparts.Yet, the Phone OximeterTM alone provided similar accuracies, maintaining a goodsensitivity-specificity balance.Considering the population under 14 years old (16% of 4,609,946 [68]) inBritish Columbia, in conjunction with SDB prevalence [64] of 2%, around 14,750children would suffer from SDB. In this study, 38% of children with signs of SDB89referred to BCCH for a PSG, were diagnosed with SDB upon analysis of a fullPSG. Therefore, approximately 38,815 children with signs of SDB may require aPSG at BCCH, where only 250 PSGs can be performed per year. The availabil-ity of PSG does not meet the demand requirements and results in long waitlists.The results of this study show that using the Phone OximeterTM as a screening toolprior to PSG could reduce the number of PSGs required, while effectively studyingthe same number of children which would result in increased coverage of medicalservices to children in British Columbia with signs of SDB, reducing wait timesand optimizing usage of hospital resources.90(a)(b)Figure 5.2: The area under the curve (AUC) of the receiver operating charac-teristic (ROC) curve of a) A/H classifier (PPG + SpO2 features) and b)A/H classifier (SpO2 features)91Figure 5.3: The estimated and observed A/H epochs for a subject with (a)high accuracy (79%), (b)low specificity with low number A/H events,and (c) low specificity with high number A/H events.92Chapter 6Conclusion and Future WorkWe evaluated the relative impact of SDB on sympathetic and parasympathetic ac-tivity in children through the characterization of PPG and we concluded that sym-patetic activity during sleep was higher in children with SDB sleep and also during30-second epochs when apnea/hypopnea events happen. We later characterizedthe SpO2 pattern in SDB and then combined SpO2 pattern characterization andPPG analysis to design and develop a method with two binary multivariate logis-tic models to automatically identify 30-s epochs with apnea/hypoeponea events.We extracted the cycles of non-REM and REM of the overnight sleep based onthe activity of cardiorespiratory system using the overnight PPG signals. We ex-tracted the relevant features associated with PRV, RR, vascular tone and movementfrom the PPG signal to build a multivariate model with a minimum set of featuresto identify wakefulness from REM and non-REM sleep. To develop and evalu-ate the proposed models, we recorded the SpO2 and PPG from 160 children usingthe Phone OximeterTM in the standard setting of overnight PSG in BC Children\u00E2\u0080\u0099shospital in Vancouver.The Phone OximeterTM provides the perfect platform to create an SDB screen-ing prototype, permitting overnight pulse oximetry recordings and allowing imple-mentation of the algorithm on a smartphone. In addition, it can wirelessly commu-nicate information (raw data, results etc.). More sophisticated analysis approachessuch as the correntropy spectral density [21], [23], could be applied to the SpO2for a more robust spectral analysis that includes nonlinear information. However,93simpler algorithms are preferred so that they can be easily implemented on a smart-phone with low computational load. By using the low cost version of the PhoneOximeterTM , which interfaces the sensor directly with the phone via the audio jack[62], the cost to monitor SDB with the phone will be reduced to that of the fingerprobe alone. The offline SpO2 and PPG analysis for the overnight study of eachsubject takes between 1 to 2 seconds. Real time performance is not required, sincewe aim to provide a final screening result after the overnight recording.6.1 Future WorkIn the remainder of this final chapter, we propose future work that may be per-formed to further develop an integrated solution for monitoring sleep and sleepbreathing disorders.6.1.1 Sleep SolutionThe ultimate goal of this research is to develop a stand-alone solution for moni-toring sleep and SDB at home using the Phone OximeterTM . After an overnightrecording at home, the 30-s will be classified into the wakefulness, non-REM andREM states using the sleep model described in chapter 4. Later the epochs withA/H events would be identified using the A/H model presented in chapter 5. Theresults of these two models would be integrated into one report to offer valuableinformation about the quality of sleep, variation of heart rate and oxygen saturationduring sleep (Figure 6.1).6.2 Limitation of the researchThe pediatric population of this study includes children with a higher likelihoodof SDB than the general population, having already been referred to the BC chil-dren\u00E2\u0080\u0099 hospital for a PSG. Although our target population for the SDB screeningtool is children with signs of SDB, the utility of the Phone Oximeter in a generalpopulation with a lower prevalence of SDB is presently unproven.The database was used to evaluate the proposed models and algorithms studywas collected performed in a hospital sleep laboratory at the BC children\u00E2\u0080\u0099 hospital.94Sleep Report User Information Patient ID: Gender : Age: BMI (kg/cm2): Study date: Start: End: Time (hour) Pulse Rate (BPM) Mean 77 Lowest 72 Highest 83 Sleep Summary Total Time in Bed 8 h 22 m Total Sleep Time 8 h 3 m 96% Wake 19 m 4% NREM 7 h 30 m 93% REM 32 m 7% Sleep Efficiency 96% 6 Disruptions Motion Artifact: 0.16% Error Signal: 22.43% Sleep Oximetry * SpO2 < 88 < 80 Duration - - Sleep % - - SpO2 Baseline (%) ** Highest 97.1 Lowest 94.6 Drift 2.5 Awake SpO2 97.2 Respiratory Disturbance *** Oxygen Desaturation Events - Oxygen Desaturation Index (ODI) - Apnea/Hypopnea Epochs - Apnea/Hypopnea Epoch Index - Figure 6.1: Sleep report provides the valuable information about the qual-ity of sleep, variation of heart rate and oxygen saturation during anovernight sleep95At-home screening is our goal for the next study. During recordings performed athome, we expect artifacts caused by sensor displacement to be more severe, whichcould degrade the performance of the Phone Oximeter as an SDB screening tool.Therefore, the implementation of an accurate artifact detection technique for thePPG and SpO2 signals, directly on the phone, is one of our main future challenges.Previous studies suggest that the indication for SDB treatment, primarily ade-notonsillectomy, is an AHI (from PSG)>5, which coincides with the current prac-tice at BC Children\u00E2\u0080\u0099s hospotal. Therefore, in this research we considered childrenwith an AHI as positive for SDB. However, there is no discrete definition of OSAbased on AHI alone, but rather a continuum from normal to abnormal. We rec-ognize that some studies consider an AHI as abnormal or mild obstructive sleepapnea (OSA). For example, The Childhood Adenotonsillectomy Trial (CHAT), de-signed to evaluate the efficacy of early adenotonsillectomy versus watchful waitingwith supportive care, defined OSA as an AHI score 2. Surgical treatment did notsignificantly improve attention or executive function in these patients, but did re-duce OSA symptoms. However, the population in the CHAT study primarily hadmild cases of OSA, reflected by the AHI interquartile range (2.5 to 8.9) in the OSApositive group, which may have affected their assessment of treatment efficacy.Therefore, we will further investigate the Phone Oximeter\u00E2\u0080\u0099s performance identify-ing children with SDB based on different AHI thresholds (AHI >= 1 , AHI >= 2), using different classifiers. An AHI will result in a recommendation for at-homemonitoring, and an AHI will result in a referral to BC Children\u00E2\u0080\u0099s hospital for aPSG.96Bibliography[1] A. Z. Abid, M. A. Gdeisat, D. R. Burton, and M. J. Lalor. Ridge extractionalgorithms for one-dimensional continuous wavelet transform: acomparison. journal of physics, 76:1\u00E2\u0080\u00938, 2007. \u00E2\u0086\u0092 pages 45, 46[2] P. S. Addison and J. Watson. Secondary transform decoupling of shifted nonstationary signal modulation components: applications tophotoplethysmography. International Journal of Wavelets, Multiresolutionand Information Processing, 2(1):43\u00E2\u0080\u009357, 2004. \u00E2\u0086\u0092 pages 39, 42[3] P. S. Addison, J. N. Watson, M. L. Mestek, and R. S. Mecca. Developing analgorithm for pulse oximetry derived respiratory rate (rroxi): a healthyvolunteer study. Journal of Clinical Monitoring and Computing, 26(1):45\u00E2\u0080\u009351, 2012. \u00E2\u0086\u0092 pages 39, 54[4] American Academy of Sleep Medicine. The International Classification ofSleep Disorders. American Academy of Sleep Medicine, 2005. \u00E2\u0086\u0092 pages 3[5] American Academy of Sleep Medicine. Rules for scoring respiratory eventsin sleep: update of the 2007 aasm manual for the scoring of sleep andassociated events. J Sleep Res, 8(5):597\u00E2\u0080\u0093619, 2012. \u00E2\u0086\u0092 pages 3, 19, 76[6] A. Baharav, S. Kotagal, B. K. Rubin, J. Pratt, and S. Akselrod. Autonomiccardiovascular control in children with obstructive sleep apnea. Clinicalautonomic research : official journal of the Clinical Autonomic ResearchSociety, 9(6):345\u00E2\u0080\u009351, 1999. \u00E2\u0086\u0092 pages 29, 31[7] H. X. Barnhart, M. J. Haber, and L. I. Lin. An overview on assessingagreement with continuous measurements. Journal of biopharmaceuticalstatistics, 17(4):529\u00E2\u0080\u009369, 2007. \u00E2\u0086\u0092 pages 48[8] R. D. Berger, S. Akselrod, D. Gordon, and R. J. Cohen. An efficientalgorithm for spectral analysis of heart rate variability. IEEE Transactionson Biomedical Engineering, 33:900\u00E2\u0080\u00934, 1986. \u00E2\u0086\u0092 pages 20, 59, 7797[9] P. Boudreau, W.-H. Yeh, G. a. Dumont, and D. B. Boivin. Circadianvariation of heart rate variability across sleep stages. Sleep, 36(12):1919\u00E2\u0080\u009328,2013. \u00E2\u0086\u0092 pages 56[10] T. D. Bradley and J. S. Floras. Sleep apnea and heart failure part ii: Centralsleep apnea. Circulation, 107:1822\u00E2\u0080\u009326, 2003. \u00E2\u0086\u0092 pages 5[11] P. H. Charlton, T. Bonnici, L. Tarassenko, D. A. Clifton, R. Beale, and P. J.Watkinson. An assessment of algorithms to estimate respiratory rate fromthe electrocardiogram and photoplethysmogram. Physiologicalmeasurement, 37:610\u00E2\u0080\u009326, 2016. \u00E2\u0086\u0092 pages 47, 53, 54[12] P. D. Chazal, C. Heneghan, and W. T. McNicholas. Multimodal detection ofsleep apnoea using electrocardiogram and oximetry signals. Philos Trans AMath Phys Eng Sci, 367:369\u00E2\u0080\u009389, 2009. \u00E2\u0086\u0092 pages 75, 89[13] R. Chervin, D. Murman, B. Malow, and V. Totten. Cost-utility of threeapproaches to the diagnosis of sleep apnea: polysomnography, home testing,and empirical therapy. Annals of Internal Medicine, 130:496\u00E2\u0080\u0093505, 1999. \u00E2\u0086\u0092pages 6, 11[14] F. Chouchou, V. Pichot, J.-C. Barthe\u00C2\u00B4le\u00C2\u00B4my, H. Bastuji, and F. Roche. Cardiacsympathetic modulation in response to apneas/hypopneas through heart ratevariability analysis. PloS One, 9(1), 2014. \u00E2\u0086\u0092 pages 31, 75[15] F. Chung, P. Liao, H. Elsaid, S. Islam, C. M. Shapiro, and Y. Sun. Oxygendesaturation index from nocturnal oximetry: A sensitive and specific tool todetect sleep-disordered breathing in surgical patients. Anesth Analg, 114:993\u00E2\u0080\u00931000, 2012. \u00E2\u0086\u0092 pages 89[16] D. Clifton, G. J. Douglas, P. S. Addison, and J. N. Watson. Measurement ofrespiratory rate from the photoplethysmogram in chest clinic patients.Journal of Clinical Monitoring and Computing, 21(1):55\u00E2\u0080\u009361, 2007. \u00E2\u0086\u0092pages 39, 54[17] I. Daubechies, J. Lu, and H. T. Wu. Synchrosqueezed wavelet transforms:An empirical mode decomposition-like tool. Applied and ComputationalHarmonic Analysis, 30(2):243\u00E2\u0080\u0093261, mar 2011. \u00E2\u0086\u0092 pages 39, 40[18] P. Dehkordi, A. Garde, W. Karlen, D. Wensley, J. M. Ansermino, and G. A.Dumont. Pulse rate variability compared with heart rate variability inchildren with and without sleep disordered breathing. In Conf Proc IEEEEng Med Biol Soc., pages 6563\u00E2\u0080\u00936, 2013. \u00E2\u0086\u0092 pages 16, 7598[19] P. Dehkordi, A. Garde, W. Karlen, D. Wensley, J. M. Ansermino, and G. A.Dumont. Pulse rate variability in children with sleep disordered breathing indifferent sleep stages. In Conf. Proc. Computing in Cardiology, pages1015\u00E2\u0080\u009318, 2013. \u00E2\u0086\u0092 pages 75[20] P. Dehkordi, A. Garde, W. Karlen, C. L. Petersen, D. Wensley, G. A.Dumont, and J. M. Ansermino. Evaluation of cardiac modulation in childrenin response to apnea/hypopnea using the Phone Oximeter . PhysiologicalMeasurement, 37(2):187\u00E2\u0080\u0093202, 2016. \u00E2\u0086\u0092 pages 38, 57, 75[21] P. Dehkordi, A. Garde, J. M. Ansermino, and G. A. Dumont.Correntropy-based pulse rate variability analysis in children with sleepdisordered breathing. Entropy, 19(6), 282:1\u00E2\u0080\u009310, 2017. \u00E2\u0086\u0092 pages 93[22] J. F. Fieselmann, M. S. Hendryx, C. M. Helms, and D. S. Wakefield.Respiratory rate predicts cardiopulmonary arrest for internal medicineinpatients. Journal of General Internal Medicine, 8(7):354\u00E2\u0080\u009360, 1993. \u00E2\u0086\u0092pages 36[23] A. Garde, L. Srnmo, R. Jan, and B. F. Giraldo. Correntropy-based spectralcharacterization of respiratory patterns in patients with chronic heart failure.IEEE Trans Biomed Eng, 57:1964\u00E2\u0080\u00931972, 2010. \u00E2\u0086\u0092 pages 93[24] A. Garde, W. Karlen, P. Dehkordi, J. Ansermino, and G. Dumont. Empiricalmode decomposition for respiratory and heart rate estimation from thephotoplethysmogram. In Computing in Cardiology, volume 40, pages799\u00E2\u0080\u0093802, 2013. \u00E2\u0086\u0092 pages 39[25] A. Garde, W. Karlen, P. Dehkordi, D. Wensley, J. M. Ansermino, and G. A.Dumont. Oxygen saturation in children with and without obstructive sleepapnea using the phone-oximeter. In 2013 35th Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society,EMBC 2013, pages 2531\u00E2\u0080\u00932534, 2013. ISBN 9781457702167. \u00E2\u0086\u0092 pages 34,75, 79[26] A. Garde, W. Karlen, J. M. Ansermino, and G. A. Dumont. EstimatingRespiratory and Heart Rates from the Correntropy Spectral Density of thePhotoplethysmogram. PloS one, 2014. \u00E2\u0086\u0092 pages 34, 39, 54[27] A. Garde, W. Karlen, P. Dehkordi, J. Ansermino, and G. Dumont. Oxygensaturation resolution influences regularity measurements. In Proceedings ofthe Annual International Conference of the IEEE Engineering in Medicineand Biology Society, 2014. \u00E2\u0086\u0092 pages 7599[28] A. Garde, P. Dekhordi, J. Ansermino, and D. GA. Identifying individualsleep apnea/hypoapnea epochs using smartphone-based pulse oximetry. InProceedings of the Annual International Conference of the IEEEEngineering in Medicine and Biology Society, 2016. \u00E2\u0086\u0092 pages 75[29] E. Gil, M. Orini, J. Bailn R snd Vergara, L. Mainardi, and P. Laguna.Photoplethysmography pulse rate variability as a surrogate measurement ofheart rate variability during non-stationary conditions. PhysiologicalMeasurement, 31(9):1271\u00E2\u0080\u009390, 2010. \u00E2\u0086\u0092 pages 16, 28, 75[30] C. Heneghan, C.-P. Chua, J. F. Garvey, P. de Chazal, R. Shouldice, P. Boyle,and W. T. McNicholas. A Portable Automated Assessment Tool for SleepApnea Using a Combined Holter-Oximeter. Sleep, 31(10):1432\u00E2\u0080\u009339, 2008.\u00E2\u0086\u0092 pages 75, 89[31] L. Horwood, R. T. Brouillette, C. D. McGregor, J. J. Manoukian, andE. Constantin. Testing for pediatric obstructive sleep apnea when health careresources are rationed. JAMA tolaryngology Head & Neck Surgery, 2014.\u00E2\u0086\u0092 pages 74[32] J. Hudson, S. M. Nguku, J. Sleiman, W. Karlen, G. A. Dumont, C. L.Petersen, C. B. Warriner, and J. M. Ansermino. Usability testing of aprototype Phone Oximeter with healthcare providers in high- andlow-medical resource environments. Anesthesia, 67:957\u00E2\u0080\u009367, 2012. \u00E2\u0086\u0092 pages10[33] P. C. Ivanov. Long-range dependence in heartbeat dynamics. Lecture Notesin Physics, 621:339\u00E2\u0080\u009368, 2003. \u00E2\u0086\u0092 pages 15, 16[34] L. J, J. J, C. X, S. W, and G. P. Comparison of respiratory-induced variationsin photoplethysmographic signals. Physiological measurement, 31(3):415\u00E2\u0080\u009325, 2010. \u00E2\u0086\u0092 pages 38, 53, 54[35] G. James, D. Witten, T. Hastie, and R. Tibshirani. An Introduction toStatistical Learning with application in R. Springer, 2013. \u00E2\u0086\u0092 pages 21, 62,64[36] K. M. C. K. and A. Blasi. Sleep-related changes in autonomic control inobstructive sleep apnea:A model-based perspective. Respiratory Physiology& Neurobiology, 188(3):267\u00E2\u0080\u009376, 2013. \u00E2\u0086\u0092 pages 17, 28, 33[37] W. Karlen and D. Floreano. Adaptive Sleep-Wake Discrimination forWearable Devices. IEEE transactions on bio-medical engineering, 58(4):920\u00E2\u0080\u0093926, 2010. \u00E2\u0086\u0092 pages 56100[38] W. Karlen, M. Turner, E. Cooke, G. A. Dumont, and J. M. Ansermino.Capnobase: Signal database and tools to collect, share and annotaterespiratory signals. In Annual Meeting of the Society for Technology inAnesthesia (STA) (2010), page 48, 2010. \u00E2\u0086\u0092 pages 41[39] W. Karlen, G. Dumont, C. Petersen, J. Gow, J. Lim, J. Sleiman, andM. Ansermino. Human-centered phone oximeter interface design for theoperating room. In Proceedings of the International Conference on HealthInformatics, pages 433\u00E2\u0080\u00938, 2011. \u00E2\u0086\u0092 pages 10[40] W. Karlen, K. Kobayashi, J. M. Ansermino, and G. A. Dumont.Photoplethysmogram signal quality estimation using repeated Gaussianfilters and cross-correlation. Physiological Measurement, 33:1617\u00E2\u0080\u009329, 2012.\u00E2\u0086\u0092 pages 19, 61[41] W. Karlen, S. Raman, J. M. Ansermino, and G. Dumont. Multi-parameterRespiratory Rate Estimation from the Photoplethysmogram. IEEETransactions on Biomedical Engineering, 60:1946\u00E2\u0080\u009353, 2013. \u00E2\u0086\u0092 pages 38,39, 53[42] C. KH, D. S, and J. K. Estimation of respiratory rate fromphotoplethysmogram data using time-frequency spectral estimation. IEEETransactions on Biomedical Engineering, 56:20542063, 2009. \u00E2\u0086\u0092 pages 39[43] A. H. Khandoker, C. K. Karmakar, and M. Palaniswam. Comparison ofpulse rate variability with heart rate variability during obstructive sleepapnea. Medical Engineering & Physics, 33(2):204\u00E2\u0080\u00939, 2011. \u00E2\u0086\u0092 pages 16, 30[44] R. E. Klabunde. Cardiovascular Physiology Concepts Second Edition.Lippincott Williams and Wilkins, 2011. \u00E2\u0086\u0092 pages 15[45] J. La\u00C2\u00B4zaro Plaza. Non-invasive techniques for respiratory informationextraction based on pulse photoplethysmogram and electrocardiogram. PhDthesis, University of Zaragoza, 2015. \u00E2\u0086\u0092 pages 46, 47[46] M. Lisenby, P. Richardson, and A. Welch. Detection of cyclic sleepphenomena using instantaneous heart rate. Electroencephalography andClinical Neurophysiology, 40(2):169\u00E2\u0080\u009377, 1976. \u00E2\u0086\u0092 pages 56[47] P. B. Lovett, J. M. Buchwald, K. Stu\u00C2\u00A8rmann, and P. Bijur. The vexatious vital:Neither clinical measurements by nurses nor an electronic monitor providesaccurate measurements of respiratory rate in triage. Annals of EmergencyMedicine, 45:68\u00E2\u0080\u009376, 2005. \u00E2\u0086\u0092 pages 36, 37101[48] D. lvarez, R. Hornero, M. Garca, F. del Campo, and C. Zamarrn. Improvingdiagnostic ability of blood oxygen saturation from overnight pulse oximetryin obstructive sleep apnea detection by means of central tendency measure.Artif Intell Med, 41:13\u00E2\u0080\u009324, 2007. \u00E2\u0086\u0092 pages 79[49] D. lvarez, R. Hornero, J. V. Marcos, and F. del Campo. Multivariate analysisof blood oxygen saturation recordings in obstructive sleep apnea diagnosis.IEEE Transactions on BioMedical Engineering, 57:2816\u00E2\u0080\u009324, 2010. \u00E2\u0086\u0092pages 74[50] D. lvarez, R. Hornero, J. V. Marcos, N. Wessel, T. Penzel, M. Glos, andF. del Campo. Assessment of feature selection and classification approachesto enhance information from overnight oximetry in the context of apneadiagnosis. International Journal of Neural Systems, 23(05), 2013. \u00E2\u0086\u0092 pages74, 80[51] A. Malhotra and D. P. White. Obstructive sleep apnoea. Lancet, 360(9328):237\u00E2\u0080\u009345, 2002. \u00E2\u0086\u0092 pages 4[52] C. L. Marcus, L. J. Brooks, S. Davidson Ward, K. A. Draper, D. Gozal, A. C.Halbower, J. Jones, C. Lehmann, M. S. Schechter, S. Sheldon, R. N.Shiffman, and K. Spruyt. Diagnosis and management of childhoodobstructive sleep apnea syndrome. Pediatrics, 130:576\u00E2\u0080\u009384, 2012. \u00E2\u0086\u0092 pages12[53] Medical Services Commission of British Columbia. Respirology. Technicalreport, 2013. \u00E2\u0086\u0092 pages 12[54] D. J. Meredith, D. Clifton, P. Charlton, J. Brooks, C. W. Pugh, andL. Tarassenko. Photoplethysmographic derivation of respiratory rate: areview of relevant physiology. Journal of medical engineering &technology, 36(1):1\u00E2\u0080\u00937, 2012. \u00E2\u0086\u0092 pages 37[55] M. Montesano, S. Miano, M. C. Paolino, A. C. Massolo, F. Ianniello,M. Forlani, and M. P. Villa. Autonomic cardiovascular tests in children withobstructive sleep apnea syndrome. Sleep, 33:1349\u00E2\u0080\u009355, 2010. \u00E2\u0086\u0092 pages 75[56] K. Nakajima, T. Tamura, and H. Miike. Monitoring of heart and respiratoryrates by photoplethysmography using a digital filtering technique. MedicalEngineering & Physics, 18:365\u00E2\u0080\u009372, 1996. \u00E2\u0086\u0092 pages 39[57] K. Narkiewicz, N. Montano, C. Cogliati, P. J. H. van de Borne, M. E. Dyken,and V. K. Somers. Altered Cardiovascular Variability in Obstructive SleepApnea. Circulation, 98(11):1071\u00E2\u0080\u009377, 1998. \u00E2\u0086\u0092 pages 17102[58] G. M. Nixon, A. S. Kermack, G. M. Davis, J. J. Manoukian, A. Brown, R. T.Brouillette, and K. A. Brown. Planning Adenotonsillectomy in ChildrenWith Obstructive Sleep Apnea: The Role of Overnight Oximetry. Pediatrics,113(1):19\u00E2\u0080\u009325, 2014. \u00E2\u0086\u0092 pages 74[59] C. K. Peng, J. Mietus, J. M. Hausdorf, S. Havlin, H. E. Stanley, and A. L.Goldberger. Long-range anticorrelations and non-Gaussian behavior of theheartbeat. Physical Review Letters, 70:1343\u00E2\u0080\u009346, 1993. \u00E2\u0086\u0092 pages 15, 16[60] C. K. Peng, S. Havlin, J. M. Hausdorff, J. E. Mietus, H. E. Stanley, andG. AL. Fractal mechanisms and heart rate dynamics. Long-rangecorrelations and their breakdown with disease. The Journal ofElectrocardiology, 28:59\u00E2\u0080\u009365, 1995. \u00E2\u0086\u0092 pages 15, 16, 20, 21[61] T. Penzel, J. W. Kantelhardt, L. Grote, J. H. Peter, and A. Bunde.Comparison of detrended fluctuation analysis and spectral analysis for heartrate variability in sleep and sleep apnea. IEEE transactions on biomedicalengineering, 5(10):1143\u00E2\u0080\u009351, 2003. \u00E2\u0086\u0092 pages 16, 19, 21, 33, 34, 56[62] C. L. Petersen, T. P. Chen, J. M. Ansermino, and D. GA. Design andEvaluation of a Low-Cost Smartphone Pulse Oximeter. Sensors, 13(12):16882\u00E2\u0080\u009393, 2013. \u00E2\u0086\u0092 pages 10, 94[63] B. Rohit, G. J. L, P. Sairam, and Q. S. F. Comparison of nasal pressuretransducer and thermistor for detection of respiratory events duringpolysomnography in children. Sleep, 28(9):1117\u00E2\u0080\u009321, 2005. \u00E2\u0086\u0092 pages 6[64] C. L. Rosen, E. K. Larkin, H. Kirchner, J. L. Emancipator, S. F. Bivins, S. A.Surovec, R. J. Martin, and S. Redline. Prevalence and risk factors forsleep-disordered breathing in 8- to 11-year-old children: association withrace and prematurity. The Journal of Pediatrics, 142:383\u00E2\u0080\u009389, 200. \u00E2\u0086\u0092 pages4, 89[65] K. H. Shelley. Photoplethysmography: Beyond the calculation of arterialoxygen saturation and heart rate. Anesthesia and Analgesia, 105:31\u00E2\u0080\u00936, 2007.\u00E2\u0086\u0092 pages 37, 57[66] K. H. Shelley, A. A. Awad, R. G. Stout, and D. G. Silverman. The use ofjoint time frequency analysis to quantify the effect of ventilation on the pulseoximeter waveform. Journal of Clinical Monitoring and Computing, 20(2):81\u00E2\u0080\u00937, 2006. \u00E2\u0086\u0092 pages 39, 54103[67] K. Skaltsa, L. Jover, and J. L. Carrasco. Estimation of the diagnosticthreshold accounting for decision costs and sampling uncertainty.Biometrical Journal, 52(5):676\u00E2\u0080\u009397, 2010. \u00E2\u0086\u0092 pages 22[68] Stats B. 2013 sub-provincial population estimates, districts, bc regional.2014. \u00E2\u0086\u0092 pages 89[69] C. P. Subbe, R. G. Davies, E. Williams, P. Rutherford, and L. Gemmell.Effect of introducing the Modified Early Warning score on clinicaloutcomes, cardio-pulmonary arrests and intensive care utilisation in acutemedical admissions. Anaesthesia, 58(8):797\u00E2\u0080\u0093802, 2003. \u00E2\u0086\u0092 pages 36[70] G. Thakur, E. Brevdo, N. S. Fuckar, and H. T. Wu. The synchrosqueezingalgorithm for time-varying spectral analysis: Robustness properties and newpaleoclimate applications. Signal Processing, 93(5):1079\u00E2\u0080\u009394, 2013. \u00E2\u0086\u0092pages 40, 54[71] J. Thayer, J. Iii Sollers, E. Ruiz-Padial, and J. Vila. Estimating respiratoryfrequency from autoregressive spectral analysis of heart period. IEEEEngineering in Medicine and Biology Magazine, 21(4):41\u00E2\u0080\u00935, 2002. \u00E2\u0086\u0092 pages39[72] G. J. Tortora and S. R. Grabowski. Principle of Anatomy and Physiology.John Wiley & Sons, Inc., 2003. \u00E2\u0086\u0092 pages 57[73] M. K. Uar, M. R. Bozkurt, C. Bilgin, and K. Polat. Automatic sleep stagingin obstructive sleep apnea patients using photoplethysmography, heart ratevariability signal and machine learning techniques. Neural Computing andApplications, 29(8):1\u00E2\u0080\u009316, 2016. \u00E2\u0086\u0092 pages 72[74] W. Wang, S. Tretriluxana, S. Redline, S. Surovec, D. J. Gottlieb, andM. C. K. Khoo. Association of cardiac autonomic function measures withseverity of sleep-disordered breathing in a community-based sample. J SleepRes., 17(3):251\u00E2\u0080\u009362, 2008. \u00E2\u0086\u0092 pages 17[75] G. W. Webster, J. G. Webster, and R. E. Webster. Design of Pulse Oximeters.Taylor & Francis Group, 1997. \u00E2\u0086\u0092 pages xv, 8, 9[76] WHO. Pocket Book of Hospital Care for Children: Guidelines for theManagement of Common Childhood Illnesses. WHO, 2013. \u00E2\u0086\u0092 pages 36[77] B. Ylmaz, M. H. Asyal, A. Eren, S. Yetkin, and F. zgen. Sleep stage andobstructive apneaic epoch classification using single-lead ecg. BioMedicalEngineering OnLine, 2010. \u00E2\u0086\u0092 pages 72104[78] J. Zhao, F. Gonzalez, and D. Mu. Apnea of prematurity: from cause totreatment. Eur J Pediatr, 170:1097\u00E2\u0080\u0093105, 2011. \u00E2\u0086\u0092 pages 74[79] G. Y. Zou. Confidence interval estimation for the bland-altman limits ofagreement with multiple observations per individual. Statistical Methods inMedical Research, 22(2):630\u00E2\u0080\u009342, 2013. \u00E2\u0086\u0092 pages 47105"@en . "Thesis/Dissertation"@en . "2018-09"@en . "10.14288/1.0369224"@en . "eng"@en . "Biomedical Engineering"@en . "Vancouver : University of British Columbia Library"@en . "University of British Columbia"@en . "Attribution-NonCommercial-NoDerivatives 4.0 International"@* . "http://creativecommons.org/licenses/by-nc-nd/4.0/"@* . "Graduate"@en . "Monitoring sleep and sleep breathing disorders using pulse oximeter photoplethysmogram"@en . "Text"@en . "http://hdl.handle.net/2429/66591"@en .