You may notice some images loading slow across the Open Collections website. Thank you for your patience as we rebuild the cache to make images load faster.

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Automatic sleep arousal detection based on C-ELM and MRMR feature selection Liang, Yuemeng 2015

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

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

Item Metadata


24-ubc_2015_september_liang_yuemeng.pdf [ 978.38kB ]
JSON: 24-1.0166399.json
JSON-LD: 24-1.0166399-ld.json
RDF/XML (Pretty): 24-1.0166399-rdf.xml
RDF/JSON: 24-1.0166399-rdf.json
Turtle: 24-1.0166399-turtle.txt
N-Triples: 24-1.0166399-rdf-ntriples.txt
Original Record: 24-1.0166399-source.json
Full Text

Full Text

Automatic Sleep Arousal Detection based on C-ELMand MRMR feature selectionbyYuemeng LiangB.E., Zhejiang University, 2012A THESIS SUBMITTED IN PARTIAL FULFILMENT OFTHE REQUIREMENTS FOR THE DEGREE OFMASTER OF APPLIED SCIENCEinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Electrical and Computer Engineering)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)July 2015c© Yuemeng Liang, 2015iiAbstractSleep arousals are sudden awakenings from sleep which can be identified as an abrupt shiftin EEG frequency and can be manually scored from various physiological signals by sleepexperts. Frequent sleep arousals can degrade sleep quality, result in sleep fragmentationand lead to daytime sleepiness. Visual inspection of arousal events from PSG recordings iscumbersome, and manual scoring results can vary widely among different expert scorers.The main goal of this project is to design and evaluate the performance of an effectiveand efficient algorithm to automatically detect sleep arousals using a single channel EEG.In the first part of the thesis, a detection model based on a Curious Extreme LearningMachine (C-ELM) using a set of 22 features is proposed. The performance was evaluatedusing the term Area Under the Receiver Operating Characteristic (ROC) Curve (AUC)and the Accuracy (ACC). The proposed C-ELM based model achieved an average AUCand ACC of 0.85 and 0.79 respectively. In comparison, the average AUC and ACCof a Support Vector Machine (SVM) based model were 0.69 and 0.67 respectively. Thisindicates that the proposed C-ELM based model works well for the sleep arousal detectionproblem.In the second part of the thesis, an improved detection model is proposed by addingAbstract iiia Minimum Redundancy Maximum Relevance (MRMR) feature selection into the C-ELM based model proposed in the first part. The efficiency of the model is improvedby reducing dimensionality (reducing the number of features) of the dataset while theperformance is largely unaffected. The achieved average AUC and ACC were 0.85 and0.80 when a reduced set of 6 features were used, while the AUC and ACC were 0.86and 0.79 for a full set of 22 features. The result indicates MRMR feature selection stepis important for sleep arousal detection. By using the improved sleep arousal detectionmodel, the system runs faster and achieves a good performance for the dataset utilizedin our study.ivPrefaceI hereby declare that I am the author of this thesis. This thesis is an original, unpublishedwork under the supervision of Professor Cyril Leung. This work was supported in partby the Natural Sciences and Engineering Research Council (NSERC) of Canada underGrant RGPIN 1731-2013, and by the UBC Faculty of Applied Science.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Sleep Arousal Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1.1 What Is Sleep Arousal . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2 Sleep Arousal Identification . . . . . . . . . . . . . . . . . . . . . 31.1.3 The Significance of Sleep Arousal Detection . . . . . . . . . . . . 51.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6Table of Contents vi1.2.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.5 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 Automatic Sleep EEG Arousal Detection based on C-ELM . . . . . . 192.1 Sleep Arousal Detection Model . . . . . . . . . . . . . . . . . . . . . . . 202.2 Data Preprocessing and Segmentation . . . . . . . . . . . . . . . . . . . . 212.2.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.2.2 Band Pass Filter and Segmentation . . . . . . . . . . . . . . . . 222.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4 Classification based on C-ELM . . . . . . . . . . . . . . . . . . . . . . . 252.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.5.1 AUC and ACC Evaluation . . . . . . . . . . . . . . . . . . . . . 312.5.2 Speed Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383 Automatic Detection based on C-ELM and MRMR . . . . . . . . . . . 393.1 Improved Arousal Detection Model . . . . . . . . . . . . . . . . . . . . . 403.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.2.1 Filters and Wrappers . . . . . . . . . . . . . . . . . . . . . . . . . 423.2.2 MRMR Feature Selection . . . . . . . . . . . . . . . . . . . . . . 44Table of Contents vii3.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 544.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66A Sleep Stages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67B Basic Concept (sensitivity, etc.) . . . . . . . . . . . . . . . . . . . . . . . 69C EEG 10–20 International System . . . . . . . . . . . . . . . . . . . . . . . 71D ROC Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73viiiList of Tables2.1 The 22 features extracted from the preprocessed data . . . . . . . . . . . 262.2 Average AUC comparison of C-ELM based model and SVM based model 312.3 Average ACC comparison of C-ELM based model and SVM based model 322.4 Best performance of C-ELM based model among 50 datasets and the cor-responding performance of SVM based model of the same dataset . . . . 352.5 Best performance of SVM based model among 50 datasets and the corre-sponding performance of C-ELM based model of the same dataset . . . . 352.6 Training times for C-ELM based model and SVM based model for a datasetwith dimension of 22 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.1 MRMR feature ranking using MIQ scheme . . . . . . . . . . . . . . . . . 473.2 Average performance of different feature sets using C-ELM . . . . . . . . 523.3 Average performance of different feature sets using SVM . . . . . . . . . 53B.1 Computing accuracy, sensitivity and so on . . . . . . . . . . . . . . . . . 70ixList of Figures1.1 A 10 seconds’ sleep EEG data from 22:42:30 to 22:42:40, collected fromone EEG channel of a PSG recording downloaded from PhysioBank [1].Theselected zone shows an EEG arousal event lasting 5.6 seconds which wasmanually scored by a sleep expert. . . . . . . . . . . . . . . . . . . . . . 32.1 Sleep arousal detection model. . . . . . . . . . . . . . . . . . . . . . . . . 202.2 ROC curves for C-ELM based and SVM based models for the dataset whichgives the highest AUC for C-ELM. The red line is a random classification,the blue line is the curve of C-ELM and the green line is the curve of SVM 352.3 ROC curves for C-ELM based and SVM based models for the dataset whichgives the highest AUC for SVM. The red line is a random classification,the blue line is the curve of C-ELM and the green line is the curve of SVM 363.1 Improved sleep detection model based on MRMR feature selection. . . . 413.2 Average AUC plot for different feature sets. The x-axis is the feature setsize.Feature set of size 1 contains top 1 ranking feature, feature set of size2 contains 2 top ranked features and so on. . . . . . . . . . . . . . . . . . 50List of Figures x3.3 Average ACC plot for different feature sets. The x-axis is the feature setsize.Feature set of size 1 contains top 1 ranking feature, feature set of size2 contains 2 top ranked features and so on. . . . . . . . . . . . . . . . . . 51A.1 Sleep stages [2] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68B.1 Binary classification basic concept [3] . . . . . . . . . . . . . . . . . . . . 69C.1 EEG 10-20 system [4] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72xiList of SymbolsSymbol DescriptionC(xt) Conflict of an input vector xtf¯ Mean frequencyfi Center frequency of the bandI(h, i) The mutual information between fea-ture i and classes h, h = {h1, h2, ..., hK}I(i, j) The mutual information between twodifferent features i and jmaxVI Maximum relevance conditionminWI Minimum redundancy conditionN (xt) Novelty of an input vector xtpi The corresponding power of center fre-quency of the band|S| The number of features in the selectedfeature set SS(xt) Surprise of an input vector xtList of Figures xiiU(xt) Uncertainty of an input vector xtxt The tth M-dimensional input vector(or feature vector) of the trainingdata {(x1, c1) , · · · , (xt, ct) , · · · }, xt =[xt1, · · · , xtM ]T ∈ <MθC Initialized neuron deletion threshold forconflictθNadd Initialized neuron addition thresholdfor noveltyθN del Initialized neuron deletion threshold fornoveltyθS Initialized neuron addition or deletionthreshold for surpriseθU Initialized neuron addition thresholdfor uncertaintyxiiiList of AcronymsACC AccuracyANN Artificial Neural NetworkASDA American Sleep Disorders AssociationAUC Area Under the CurveC-ELM Curious Extreme Learning MachineECG ElectrocardiogramEDS Excessive Daytime SleepinessEEG ElectroencephalogramEMG ElectromyographyEOG ElectroocculographyFT Fourier TransformFNR False Negative RateList of Acronyms xivFPR False Positive RateFFT Fast Fourier TransformLIBSVM Library of Support Vector MachineMID Mutual Information DifferenceMIQ Mutual Information QuotientMLP Multilayer PerceptronMRMR Minimum Redundancy-Maximum RelevanceNREM Non Rapid Eye MovementOSA Obstructive Sleep ApneaPAT Peripheral Arterial Tone testPAT-AAI Peripheral Arterial Tone test based Autonomic Arousal In-dicesPD Parkinson’s DiseasePPV Positive Predictive ValuePSG PolysomnographyPSQI the Pittsburgh Sleep Quality IndexList of Acronyms xvRBF Radial Basis FunctionROC Receiver Operating CharacteristicSAS Sleep Apnea SyndromeSBS Sequential Backward SelectionSFS Sequential Forward SelectionSHHS Sleep Heart Health StudySLFN Single hidden Layer Feedforward neural NetworkSVM Support Vector MachinexviAcknowledgementsForemost, I would like to express my deepest gratitude to my research supervisor, Prof. Cyril Le-ung, for his encouragement, guidance and support during my graduate studies and re-search. I appreciate his patience, enthusiasm, and motivation. Without his help, thisthesis would not have been possible.I am thankful to my co-supervisors, Prof. Martin J. McKeown and Prof. Chun-yan Miao (Nanyang Technological University) and their research group members, whohave provided guidance and assistance during my research, especially in discussing theresearch idea and formulating the problem. I would also like to thank Dr. Qiong Wu(NTU) for her suggestions on my work and the source code for C-ELM.Special thanks to Liang Zou for his expertise in machine learning and his preciousassistance in my research. I also wish to thank Xinxin Zhang for his help in proofreadingmy thesis and discussions on mathematics. My sincere thanks also go to colleagues inour research group for sharing their knowledge and to all my friends who encouraged andsupported me during my graduate studies.This work was supported in part by the Natural Sciences and Engineering ResearchCouncil (NSERC) of Canada under Grant RGPIN 1731-2013, and by the UBC FacultyAcknowledgements xviiof Applied Science.Last but not least, I would like to express my gratitude to my beloved parents fortheir encouragement, support and greatest love of all. To them, I dedicate this thesis.1Chapter 1IntroductionThis chapter begins with an introduction to sleep arousals. In the first section, somebackground on sleep arousals including what the sleep arousals are, why detection of sleeparousals is important and the scoring rules and detection methods are discussed. Thesecond section is a literature survey of previous works on this problem. The motivationand contributions of this thesis are then discussed. The organization of the thesis isoutlined in the last part of the chapter.1.1 Sleep Arousal DetectionSleep problems are a frequent complaint among many people, especially the elderly, andhave a substantial impact on quality of their lives. Sleep arousal conventionally refers toa temporary intrusion of wakefulness into sleep or at least a sudden transient elevationof the vigilance level due to arousal stimuli or to spontaneous vigilance level oscillations[5, 6]. Sleep arousals can be induced by various sleep disorders. Thus, arousals are agood marker of sleep disruption representing a detrimental and harmful feature for sleep[5].Chapter 1. Introduction 21.1.1 What Is Sleep ArousalSpontaneous arousal is a physiological component of normal sleep and is defined as “anyincrease in Electromyography (EMG) or any channel which is accompanied by a changein pattern on any additional channel” in Rechtschaffen & Kales criteria in 1968 [7]. Thisconventional assessment of sleep is performed in epochs of 30 seconds [7, 8]. However,in other clinical conditions, frequent transient arousals of a few seconds duration werestudied [9, 10]. In 1992, the American Sleep Disorders Association (ASDA) developedscoring rules to determine arousals quantitively based on data collected from EEG andEMG channels of Polysomnography (PSG) [6]. This scoring rule is independent of R& K’s 30-second scoring system and standardizes the assessment of arousals. It hassince become the most widely used rule for manually scoring by sleep experts. TheASDA defined sleep arousals as “an abrupt shift in EEG frequency, which may includetheta,alpha and/or frequencies greater than 16 Hz but not spindles” [6]. A sleep spindleis a burst of oscillatory brain activity visible on an EEG that occurs during sleep stage2. It consists of 12–14 Hz waves that occur for at least 0.5 seconds. For definitions ofsleep stages, please refer to Appendix A. An EEG arousal event lasting for 0.56 secondsis shown in Fig. 1.1.Other scoring rules or arousal definitions have been proposed. Arousals are definedas a return of alpha or theta rhythm for at least 1.5 seconds associated with a transient(however brief) increase in EMG tone in [11]. In [12], movement arousal is defined as anabrupt appearance of an alpha rhythm in the EEG during a sleep epoch, accompaniedChapter 1. Introduction 3Figure 1.1: A 10 seconds’ sleep EEG data from 22:42:30 to 22:42:40, collected fromone EEG channel of a PSG recording downloaded from PhysioBank [1].Theselected zone shows an EEG arousal event lasting 5.6 seconds which wasmanually scored by a sleep an increase in EMG activity lasting at least 2 seconds. Simultaneous EEG and EMGchanges must last at least 2 seconds. A return of theta rhythm and “K-arousals” arenot counted. Some computer arousal detection methods have been developed whichallow identification of arousals that are not visually scorable [13]. These methods makeidentification on arousal events which last less than 3 seconds achievable.1.1.2 Sleep Arousal IdentificationTo identify sleep arousal events, one is traditionally asked to stay overnight in a hospitalor a sleep laboratory to do a standard PSG test. It is an obtrusive test which requiresthe patient to wear a variety of sensors to collect several physiological signals. ArousalChapter 1. Introduction 4events can be scored by sleep experts from one or several of the physiological changesrecorded by PSG instrument.Increases in heart rate, blood pressure during sleep are indicators for arousal identifi-cation [14]. It has been suggested that during room air breathing, arousals are stronglyassociated with periods of arterial oxygen desaturation [9].Thus, oxygen desaturation isanother physiological signals for arousal identification. Changes in EEG and EMG ac-tivities are two important physiological indicators for arousal scoring which are used inASDA standard arousal scoring rules [6].According to ASDA arousal scoring rules, arousal events can be scored from one cen-tral EEG channel (two central EEG channels are C4/A1 and C3/A2 placement) obtainedfrom C4/A1 or C3/A2 placement without accompanied EMG channel during non-REM(Rapid eye movement) sleep stages since arousals in NREM sleep may occur withoutconcurrent increases in submental EMG amplitude (For EEG channel placement, pleaserefer to Appendix C). However, arousal events should be scored only when accompaniedby concurrent increases in submental EMG amplitude in REM sleep stages. “The pres-ence of bursts of alpha or theata activity in REM sleep EEG are common phenomena.These events may or may not reflect physiological arousal from REM sleep”. Thus, sleeparousals must be scored from both EEG and EMG activities during REM sleep. Anarousal event is scored if the EEG frequency shift lasts for 3 seconds or greater in dura-tion if scored manually since identification of events of shorter durations are difficult toachieve [6].Chapter 1. Introduction 51.1.3 The Significance of Sleep Arousal DetectionSleep problem is a common complaint even among healthy people, especially the elderly.Poor sleep quality may be indicated by reduced sleep time, increased sleep stage changes, and increased arousal frequency [9]. These short arousals are usually ignored in sleepanalyses, but their impact is significant. Too many arousals during sleep can impair sleepcontinuity even when sleep efficiency is preserved [5]. Sleep efficiency can be indicatedby the ratio between the number of hours slept and the number of hours spent in bed.Frequent sleep arousals degrade the quality of sleep and lead to sleep fragmentation.Sleep fragmentation is reported to influence the impairment of cognitive functions and isoften associated with increased daytime sleepiness.[15]. Thus, sleep arousal frequency isa very important marker for sleep quality assessment.The Pittsburgh Sleep Quality Index (PSQI) is a self-rated questionnaire which assessessleep quality and disturbances over a 1-month time interval [16]. It is the most widelyused method for sleep quality assessment as far. However, it is a relatively subjectiveapproach. A more objective way based on detection of sleep arousals and other indicatorssuch as the length of each sleep stage, etc. is worth developing.Sleep arousal detection is also a key factor for research in sleep disorders includingsleep apnea, periodic leg movement [5], snoring etc. and sleep of Parkinson’s Disease(PD) patients [17].Chapter 1. Introduction 61.2 Related WorksIn this section, previous works on scoring or automatic detection of sleep arousals arereviewed and discussed. Data recordings, methodologies, and results of experiments arebriefly described.Currently, sleep arousal events are mostly diagnosed manually. Patients are asked totake an overnight PSG test which records several physiological signals. These recordingsare then analysed and scored according to some rules by highly skilled sleep experts withspecific domain knowledge. Various scoring rules and their reliabilities and validities havebeen developed and discussed in [13, 18, 19].However, visual scoring of sleep arousals is time-consuming and cumbersome. Severalautomatic or semi-automatic detection methods based on computer algorithms have beenproposed [15, 17, 20, 21, 22, 23, 24].Detection methods in addition to analysing PSG recordings have been studied on. Oneof the detection methods is based on heart rate variability of electrocardiogram (ECG)and two other methods used peripheral arterial information to detect sleep arousals [20,25, 26]. In [25, 26], patients are asked to take an overnight PSG test and a peripheralarterial tone (PAT) test simultaneously. The PAT signal and the pulse rate derived fromit are then used to detect arousals from sleep. The total number of arousals scored by thePAT device is divided by the number of hours of sleep and termed PAT-based autonomicarousal indices (PAT-AAI). It is reported in [25, 26] that the sensitivity and specificityare 0.80 and 0.79 (for definitions of sensitivity and specificity, please refer to AppendixChapter 1. Introduction 7B), respectively and area under ROC curve (AUC) is around 0.87. They only reportedthe results of patients with at least 20 arousals/hours [25, 26].Concerning the EEG based detection methods, most of these adopt two or four EEGchannels and one or two EMG channels [15, 17, 20, 21, 22, 23, 24]. Some of them alsoadd other information such as heartbeat rate, oxygen carried by hemoglobin in the blood(SaO2%) [24], airflow pressure and airflow temperature [20].In [24], an approach is developed based on statistical and data mining techniques. Itfirst defined a set of general rules to detect arousals (termed meta-rules extraction step)with a training set of 6 adult patients’ PSG recordings. The rules are then dynamicallyadjusted depending on the individual patient (called the actual-rules extraction step) anddetected arousals. The correlations between occurrences of arousals and 2 central channelEEG (For the EEG 10-20 international system, please refer to Appendix C and [4, 27]),1 channel chin-EMG, pulse (heart beat rate) and SaO2% signals were analysed on a testset of 20 patients’ PSG recordings. The sensitivity and positive predictive values (PPV)were found to be 75.2% and 76.5% respectively (for definition of PPV, please refer toAppendix B ). The sensitivity and PPV were found to be 49.4% and 82.5% when onlyEEG channels were used.An automatic detection method of EEG arousals is described in [22]. The authorsused two EEG channels (F4-C4 and C4-O2) and one EMG channel. In the first step of thestudy, a wavelet transform was used to process EEG signals and characterized the signal inthe time-frequency domain. A set of indices were obtained after the first step. The indicesChapter 1. Introduction 8obtained from the first step was then used to estimate a linear discriminant function. Each0.125-second epoch was evaluated with the function and arousals were marked when theylast 3 or more seconds. Each possible arousal event was given a score. In the third step,the PSG recordings were inspected by two sleep experts independently. They then jointlyexamined the events scored by themselves and those scored by the computer’s automaticalgorithm for all the recordings. A reference set of arousal events named as definitearousals and uncertain arousals were obtained according to the two scorers’ opinions.They defined a correctly detected arousal as an arousal event which overlapped with thereference set. They reported an overall sensitivity of 88.1% for the automatic method and72.4% and 78.4% for the two experts with a selectivity of 74.4% for the automatic methodand 83.0% and 82.0% for the experts when only definite arousals were considered. Thesensitivity decreased to 84.5% , 67.9%, 73.2% and selectivity increased to 88.4%, 96.1%,and 94.6% for the computer, expert 1 and expert 2 respectively when all possible arousalswere included in the reference set.In [20], a study was conducted on the detection of respiratory-related arousals. Inthis work, a method for automatic detection of EEG arousals in sleep apnea syndrome(SAS) patients was proposed. PSG recordings including four channels of EEG (C3-A2,C4-A1, O1-A2, O2-A1), two channels of EMG, electroocculography (EOG), ECG, airflowpressure and temperature etc. were used. First, data were segmented into 2.56 secondsfor EEG and EMG data. For respiratory data (airflow pressure and aiflow temperature),10.24 s was adopted as the segmentation length according to the definition of the durationChapter 1. Introduction 9of SAS [20]. Then, some fundamental parameters of amplitude, relative power and centralfrequency of EEG were calculated. Airflow pressure and temperature information wereused for detecting pathological events, such as obstructive sleep apnea (OSA), which werethen utilized for determining threshold values for EEG arousal detection. The authorsreported an overall accuracy defined as the percentage of (TP + TN)/(TP + TN + FP+ FN) (please see Appendix B) of 86%, a false negative rate (FNR) of 18% and a falsepositive rate (FPR) of 12%.Another automatic detection method based on the idea of segmentation, spectralfeature extraction, statistical methods and decisional rules is described in [21]. TwoEEG channels of 2 patients’ PSG recordings were utilized and three sleep experts wereasked to score the sleep arousal events in this study. An automatic detection is assumedto be a valid arousal event if there is any overlap with the manually marked events. Forone patient, the sensitivity and specificity were 82.2% and 72.4% when compared to scoreA. The sensitivities and specificities were found to be 66.4%, 81.8% and 74.5%, 67.3%respectively when compared to scorer B and scorer C. For the other patient, the sensitivityand specificity were 70.1% and 71.1% when compared to score A. The sensitivities andspecificities were found to be 42.7%, 80.3% and 74.1%, 56.6% respectively when comparedto scorer B and scorer C.An approach to detect sleep EEG arousals based on signal processing and machinelearning paradigm is presented in [23]. Two channels’ EEG signals and one channel’s chin-EMG signal of each of 10 patients’ PSG sleep recordings were used. In the first phase, rawChapter 1. Introduction 10data were segmented into one-second epochs. The energies of different sleep bands weremeasured using the Fourier Transform (FT). A set of 40 features of the 3 channels’ signalswere extracted in total to train classifiers. In the second phase, several models based onthe classic Fisher’s linear Discriminant, a quadratic discriminant, several configurationsof Support Vector Machine (SVM) based on different parameters, and configurations offeed-forward Artificial Neural Networks (ANN) of different neurons in one hidden layerwere tested. The SVM and ANN models achieved better performances than the othertwo classifiers and the best overall accuracy was reported to be 0.92 which was achievedby one model of ANN.Two studies based on segmentation, feature extraction and machine learning tech-niques are reported in [15] and [17]. In [17], four channels of EEG (two central (C3-A2,C4-A1) and two occipital (O1-A2, O2-A1)) and one submental EMG channel of PSGrecordings were used. Patients recruited in this study were patients with Parkinson Dis-ease (PD). After data preprocessing, a titak if 14 features were extracted including sleepstages scored by sleep experts. Then, a two-layer feed-forward neural network with 9neurons in a hidden layer was applied to classify the arousals with features extractedpreviously. In the last step, a postprocessing step was added to combine arousals classiedin a certain proximity of each other. Arousals closer than 10 seconds from each other werecombined to one arousal event. Arousals detected but lasting less than 3 seconds were re-moved. The authors assumed correctly detected arousal events as ones which overlappedwith manually scored arousals. They reported an average sensitivity of 89.8% and PPVChapter 1. Introduction 11of 88.8%.In [15], only a single channel EEG (C3-A2) was used to automatic detect sleeparousals. Sleep data of non-REM sleep stages (wake stages and REM stages were ex-cluded) of 9 PSG recordings of patients with sleep apnea, snoring and excessive daytimesleepiness (EDS) were used. After some preprocessing of the data, time-frequency analysiswas used to extract several features. In the last step, the support vector machine (SVM)classifier was applied to features extracted based on 1-second epochs. The informationof manually scored sleep stages was also included as one of the features. The authorsreported that the proposed method achieved a sensitivity of 75.26% and specificity of93.08% compared to the sleep expert’s scores.1.2.1 DiscussionTo our knowledge, none of the works has reported a comparison between its own resultand that of other works. Maybe it is because it’s hard and unfair to do a comparison.Several possible reasons are listed as follows.• First, different dataset were used in different works. Most of the studies collaboratedwith their own hospitals to recruit patients to collect PSG data. The devices usedand patients participating in the test can vary a lot among different studies. Inaddition, various sleep experts involved in annotating the sleep arousal events indifferent studies. The results reported in [21] indicated the big difference betweendifferent scorers when doing annotation. The sensitivity can be as high as 70.1%Chapter 1. Introduction 12when compared with the annotations of scorer A and can be as low as 42.7% whencompared with scorer B for the dataset of the same patient.• Second, different physiological signals are used in different studies. For example,several studies used 2 channels of EEG signals and others used 4 channels of EEGsignals. A few works added airflow pressure or temperature information whileothers utilized heartbeat rate etc.• Last but not least, various performance evaluation methods are used in differentworks and there is no standard criteria for performance evaluation on this sleeparousal detection problem. For example, in [15], sensitivity and specificity were usedto report research results. The annotations are segmented into 1-second epochs andthe sensitivity and specificity were computed based on 1-second epochs. In anotherstudy [17], sensitivity and PPV were used and no specificity was calculated. Inaddition, they calculated the results based on arousal events lasting more than 3seconds. And they assumed correctly detected arousal events as ones which over-lapped with manually scored arousals. Based on their method, a high sensitivitycould be achieved while the duration, start and end of an arousal event may varya lot from the sleep expert’s annotations. In study [22], sensitivity and selectivitywere used as evaluation. In addition, they established a jointly reference set basedon two sleep experts’ annotations and the experiment result of the automatic de-tection algorithm and it was then used as the gold standard to compute sensitivityand selectivity instead of traditional annotation of one sleep expert. In [20], anChapter 1. Introduction 13overall accuracy, false negative rate and false positive rate were used to present theexperiment result.In addition, most of the studies utilized imbalanced dataset to evaluate performancesand they did not use imbalanced learning algorithms or make a balanced dataset. Thismay lead to a good overall accuracy while the real performance is bad.1.3 MotivationsSleep arousals are associated with various sleep disorders and can be a good indicatorfor sleep quality assessment. So far, sleep EEG arousals are mostly diagnosed by sleepexperts with specific domain knowledge and the patient is required to take an overnightsleep test in the hospital or a sleep lab. There are several disadvantages for this kind oftraditional sleep test. For example:• It is very time consuming and cumbersome for a sleep expert to manually scoresleep arousals because the expert needs to visually inspect the different channels ofa PSG recording including EEG, EMG, EOG etc.• Visual inspection is a relatively subjective way to diagnose sleep arousals. Therecan be large differences between individual sleep experts. For instance, in [22],the sensitivity of sleep expert A was 72.4% and the sensitivity of sleep expert Bwas 78.4% compared to the reference set which was jointly scored according tothe computer algorithm’s result, sleep expert A and sleep expert B’s results. AndChapter 1. Introduction 14the agreement between the two experts was only 68%. In another study[21], thesensitivities of the automatic method varies from 70.1% to 42.7% when comparedto two different scorers for the same patient. These results indicate less accuracyof manually scoring.• It has been suggested that arousals of short durations (less than 3 seconds) mayalso be significant [13]. However, identification and agreement on events of suchshort durations are difficult to achieve, if scored manually.• From a patient’s perspective, the cost for a PSG test is high (ranging from $700 to$6000). The patient need to sleep in a sleep lab or hospital for a full night with alot of electrodes attached to the patient’s body. This may disturb the sleep processof the patient which makes the test less reliable. In addition, a sleep technicianshould always be in attendance and is responsible for attaching the electrodes tothe patient and monitoring the patient during the study.Due to the above–mentioned disadvantages, research on fast, accurate computer-aidedautomatic arousal detection approaches and portable, less obtrusive detection deviceswhich allow patients to take the tests at home are of great significance. Several studieson automatic or semi-automatic sleep arousal detection are mentioned in Section 1.2.However, a number of issues still need to be solved.• Most of the previous studies utilized various physiological information collectedfrom a number of channels of PSG tests. For example, the method in [24] used 2Chapter 1. Introduction 15central channels of EEG (C4-A1, C3-A2), 1 channel of chin-EMG, pulse (heartbeatrate) and SaO2% signals. In [20], 4 channels of EEG (C3-A2,C4-A1, O1-A2, O2-A1), 2 channels of EMG, airflow pressure and airflow temperature, etc. were used.A large number of channels of information collected means more inconvenience forpatients since more electrodes need to be placed on patients. In order to manufac-ture portable, less obtrusive devices, use of fewer electrodes is desirable. However,fewer channels of information lead to lower accuracy. For instance, the sensitivitywas decreased from 75.2% to 49.4% when only 2 EEG channels were used in [24].Thus, methods which can achieve relatively high accuracy with less physiologicalinformation collected should be studied.• In sleep arousal detection, the amount of patient data is quite huge and takes longtime to be processed, even by computer algorithms. In order to analyse data moreeffectively and even make real-time display achievable, a relatively fast algorithmwith high accuracy is necessary.• Several features need to be extracted to train the classifier in machine learningbased algorithms. Features are chosen or added based on previous works or theresearcher’s own view in most of the studies. However, redundant or unimportantfeatures may be added during the feature extraction process which will lower downthe speed of the algorithm and may even decrease performance. Thus, a reliablefeature selection algorithm is crucial.Chapter 1. Introduction 161.4 ContributionsIn this thesis, an algorithm to detect non-REM sleep EEG arousals using only 1 channelEEG(C4-A1/C3-A2) is developed. The main contributions are summarized as follows:• In chapter 2, an automatic sleep arousal detection algorithm is proposed. Rawdata from [1] is used in our study. A set of 22 features different from previous workis extracted based on 1-second segmentation from the preprocessed dataset. Arecently proposed classifier named Curious Extreme Learning Machine (C-ELM),which is fast and easily implemented is adopted to do a binary classification onthe whole feature set. The widely used Support Vector Machine (SVM) classifieris also used on the feature set. The information of the accuracy and the AreaUnder the Receiver Operating Characteristic (ROC) Curve (AUC) are calculatedand compared for both our C-ELM-based and SVM-based detection algorithms.The speed of the two methods are also compared. During the process, 10-fold crossvalidation is used to avoid bias due to luckily/unluckily selected validation set, thusmaking the performance estimate less sensitive to the partitioning of the data. Theresult shows that our C-ELM based detection model has a better performance thanSVM-based model.• In Chapter 3, an improved automatic sleep arousal detection model based on Min-imum Redundancy Maximum Relevance (MRMR) feature selection method andC-ELM are proposed. MRMR feature selection step is added to reduce the dimen-Chapter 1. Introduction 17sionality of the feature set and determine the subset of features which gives the bestperformance. It is shown that a subset of 17 features achieves the best performanceand a set of 6 features can still have a similar performance to the 17 feature set.A low dimension feature set can increase the speed of the sleep arousal detectionalgorithm. By the result obtained, the improved model is found to achieve a goodperformance with a reduced system complexity. This result also indicates that theMRMR feature selection step plays an important role in designing an sleep arousaldetection algorithm which is fast and accurate.1.5 Structure of the ThesisThe remainder of the thesis is organized as follows.In Chapter 2, we present our C-ELM classifier based algorithm and evaluate its clas-sification performance. The model and process are first described. Then, data prepro-cessing and segmentation are introduced. Next, feature extraction and C-ELM, SVMclassification are described . Finally, the performance of the algorithm is discussed (crossvalidation was utilized since the dataset is limit).In Chapter 3, we propose an improved algorithm with MRMR feature selection. First,a few feature selection methods including MRMR feature selection method are describedand compared. Then, binary classifications using C-ELM and SVM are performed ondifferent feature subsets according to the MRMR feature selection ranking. The per-formances are also reported in this part. Finally, a brief summary of this chapter isChapter 1. Introduction 18given.19Chapter 2Automatic Sleep EEG ArousalDetection based on C-ELMIn this Chapter, we present our C-ELM classifier based algorithm and evaluate its classi-fication performance. Support Vector Machine (SVM) has good performance on binaryclassification problems and it has been reported that it performs well when applied tosleep arousal detection problems[15, 23]. So, we also apply SVM on our feature set forcomparison with our C-ELM based model. The overall sleep arousal detection model isdescribed in Section2.1. In Section2.2, data preprocessing including the selection of rawdata and band-pass filter process and segmentation are introduced. In Section2.3, variousfeatures are described and extracted from the preprocessed data. In Section2.4, theory ofCurious Extreme Learning Machine (C-ELM) is studied. In Section2.5, the sleep arousaldetection performances using models based on C-ELM and SVM binary classificationsare compared. Cross validation is used to reduce the variance for different datasets whenwe do the performance estimate. The AUCs, ACCs and training times of C-ELM andSVM based algorithms are compared. In Section2.6, a summary is given.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 20One channel EEG signal Band pass filter Segmentation Time frequency feature extraction PSG signals Sleep stage annotation Wake Stage 1 Stage 2 Stage 3 Sleep arousal annotation C-ELM classification Non-REM 1     2  21   22        … Feature  matrix         … Figure 2.1: Sleep arousal detection model.2.1 Sleep Arousal Detection ModelOur sleep arousal detection algorithm is based on segmentation and classification. First,raw sleep dataset which contains noise is obtained from Physiobank[1]. A band pass filteris used to remove artifacts and irrelevant information. Next, the preprocessed dataset issegmented into 1-second epochs in order to do the classification. Since the input data istoo large to be processed, a feature extraction step is used to transform the raw datasetinto feature vectors which contain the relevant information. Finally, the feature vectorsof the dataset are input into the Curious Extreme Learning Machine (C-ELM) classifierand SVM classifier. The overall model is illustrated in Fig. 2.1.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 212.2 Data Preprocessing and SegmentationWe now describe the Data acquisition, data preprocessing including band pass filter andsegmentation.2.2.1 Data AcquisitionOne central EEG channel (C4-A1/C3-A2) with a sampling frequency of 250 Hz of apatient’s single overnight PSG recording is utilized in this thesis. The EEG raw data isdownloaded from Sleep Heat Health Study (SHHS) PSG DataBase of PhysioBank[1]. TheSHHS is a prospective cohort study designed to investigate the relationship between sleepdisordered breathing and cardiovascular disease. The age of the patient used in the studyis over 40, without tracheostomy, without history of treatment of sleep apnea, withoutcurrent home oxygen therapy. Other information, such as a sleep expert’s annotations ofarousal events and sleep stages, are downloaded from PhysioBank as well.According to the ASDA manually scoring rules [6], arousal events during REM sleepstages must be scored when at least one EMG channel is used since the arousal eventsduring REM sleep stages must be accompanied by an increase in submental EMG ac-cording to ASDA rules [6]. So only data of non-REM sleep stages (sleep stage 1, 2, 3)and wake stage are included in this study. Consequently, we have investigated a total of1,920,000 samples (7680 seconds) for arousal detection.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 222.2.2 Band Pass Filter and SegmentationAccording to [15, 17, 20, 21, 22, 23, 24], sleep related frequencies can be divided into 6bands: 0-0.5 Hz (gamma or slow delta), 0.5-4 Hz (delta), 4-8 Hz (theta), 8-12 Hz (alpha),12-16 Hz (sigma), 16-30 Hz (beta). Some of the above-mentioned works define the betaband as 16-64 Hz [22], 16-40 Hz [21] or >13 Hz [24]. In [6], sleep EEG arousals are relatedto the theta, alpha and beta bands. Other sleep bands are related to sleep stages or sleepspindles. In order to remove noise and frequencies non-related to sleep, we band-passfilter the raw EEG signal from 0-50 Hz.Analysis tools, such as Fast Fourier Transform (FFT), are widely used to processEEG signals. However, in this research, we want to identify sleep arousals based on1-second epochs. Thus a time-frequency representation is performed which enables us toobtain time and frequency information simultaneously. It is useful in analyzing complexphysiological signals [15]. In order to do time-frequency analysis and extract featurevectors from the signal every second, the band-pass filtered signal is segmented into 1-second epochs. Then, frequency analysis is performed for each epoch. A total of 7680epochs of sleep data are thus obtained.2.3 Feature ExtractionIn this section, features extracted from the sleep EEG data are listed and described. Inthis stage, a total of 22 features are extracted from one single channel EEG to be usedChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 23in the classification stage. Some of the 22 features may be redundant or features withless predict power; however, in this stage, this is not of immediate concern. This will bediscussed in the feature selection part in Chapter 3.All the 22 features are now described. In the feature extraction step, Fast FourierTransform (FFT) is used for the frequency and power analysis.• Power Ratio: According to the ASDA scoring rules [6], sleep EEG arousals areabrupt frequency shifts of theta, alpha and beta sleep bands. The frequency shiftcan be represented by the changes of power in time. First, for each one secondepoch, two temporal windows which contain the power information are chosen.A window of 10 seconds ending in the current epoch is used to represent priorpower information and another window starting from the current epoch is chosento provide the current or future power information. The changes of power can berepresented by the power ratio between these two windows. According to [15], wemake the “future” window 1 second in length and we also choose another “future”window of 3 seconds according to [23]. That is to say, we have 2 different powerratio frames. One is 1 second/10 second frame and the other one is 3 second/10second frame. The duration of 10 seconds as the former window length also comesfrom the ASDA scoring rules [6]. In [6], the scoring rule suggests a minimumof 10 seconds of intervening sleep is necessary to score a second arousal once aprevious arousal is detected. So we choose 10 seconds as the length of the formerwindow. Next, each of the two windows are transformed to the frequency domainChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 24using FFT, and the power of each window can be calculated. Each of the six sleepbands’ power ratios including theta ratio (4-8 Hz), alpha ratio (8-12 Hz), beta ratio(16-30 Hz), gamma ratio (0-0.5 Hz), sigma ratio (12-16 Hz) and delta ratio (0.5-4Hz). and the whole power ratio (0-50 Hz) are calculated and extracted as features.A total of 14 features (we have two frames of windows (3-second/10-second and1-second/10-second)) are extracted based on the power ratios.• Sleep Spindle: It is stated in the ASDA scoring rules [6] that arousals are abruptEEG frequency shifts which are not sleep spindles. Hence, the power ratio betweensigma and (alpha plus beta) using 3-second/10-second window frame is selected toindicate the presence of sleep spindles [15, 22].• Mean Frequency: The signal’s mean frequency of each 1-second epoch is extractedas a feature [15]. The mean frequency is computed as follows [22].f¯ =∑pi × fi∑pi, (2.1)where fi is the center frequency of the band and pi is its power.• Power and Max Power Frequency: The power of 0-50 Hz band for each 1-secondepoch is selected as a feature. Another feature is max power frequency, whichis defined as the frequency corresponding to the maximum power or maximumamplitude in the FFT amplitude spectrum.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 25• Time Domain Based Features: The mean value and standard deviation of the signalin the time domain are selected as features for each 1 second. An abrupt shift inEEG frequency may be indicated by the number of zero-crossing, so the number ofzero-crossing is another feature selected in the time domain. We choose the meanvalue of each second as ”Zero” (the baseline). A large number of zero-crossing mayindicate an abrupt shift in EEG frequency occurs, thereby an arousal may havehappened. These three features are added from our own perspective based on theASDA rules [6].• Sleep Stages: Although it is widely accepted that the scoring of sleep arousals isindependent of Rechtschaffen & Kales criteria [7], the selection of sleep stages asa feature is still necessary. First, an arousal event is easy to be incorrectly scoredduring the wake stage. Second, sleep stages are characterized by various sleep waves(theta wave, alpha wave etc.). In this thesis, the annotations of sleep stages aredownloaded from Physionet [1] which is manually scored by sleep experts.To have a brief summary of all the features extracted in this study, please see Ta-ble. Classification based on C-ELMIn this section, the Curious Extreme Learning Machine (C-ELM) algorithm is brieflydescribed. A detailed explanation can be found in [28]. Descriptions of Support VectorChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 26Table 2.1: The 22 features extracted from the preprocessed dataFeature type # of features DescriptionPower Ratio (3-sec/10-sec) 7 Power ratio between 3 secondsstarting from the current epochand prior 10 seconds of the0-50 Hz band and six individualsleep bands (alpha, beta, etc.)Power Ratio (1-sec/10-sec) 7 Power ratio between the currentepoch and prior 10 seconds ofthe 0-50 Hz band and six individualsleep bands (alpha, beta, etc.)Sleep spindle 1 Power ratio between sigma andalpha plus beta using3-second/10-second windowsMean frequency 1 The signal’s mean frequencyof each 1-second epoch Eq. (2.1)Power 1 The power (0-50 Hz) of each epochMax Power Frequency 1 The frequency correspondingto the maximum amplitudein FFT amplitude spectrumTime Domain Features 3 The zero-crossing frequency, meanvalue and corresponding standarddeviation of the each epochsleep stages 1 Annotations of sleep stagesChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 27Machine (SVM) appear in [29, 30, 31, 32].Extreme Learning Machine (ELM) is a fast, easy-to-implement machine learningalgorithm based on a single hidden layer feedforward neural network (SLFN) withoutparameter tuning. It has been reported to have good performance and generalizationability [33, 34, 35]. Details about ELM and related algorithms based on ELM can befound in [33, 36, 37, 38, 39, 40]. Curious Extreme Learning Machine (C-ELM) is a psy-chological curiosity driven algorithm based on ELM. It follows psychological theory ofcuriosity and performs curiosity appraisal towards each input data. The algorithm hasfour variables (novelty N (xt), uncertainty U(xt), conflict C(xt) and surprise S(xt)) andthree learning strategies (neuron addition, neuron deletion and parameter update). Thefour variables are computed for each input vector xt and compared with initialized thresh-olds. According to the comparison result, one corresponding learning strategy is utilizedto adjust the structure or update the parameters of the neural network automatically.The conditions for the three learning strategies [28] are briefly summarized below.• Neuron Addition Strategy: Given an input xt, the neuron addition condition is:N (xt) > θNadd AND U(xt) > θU AND S(xt) > θS , (2.2)where xt = [xt1, · · · , xtM ]T ∈ <M is the tth M-dimensional input vector (or featurevector) of the training data {(x1, c1) , · · · , (xt, ct) , · · · } (ct ∈ [1, 2, · · · , N ] is theclass label of xt, N represents the total number of distinct classes), θNadd , θU andChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 28θS are initialized neuron addition thresholds in the range of [0.1, 0.5], [0.1, 0.3],[0.2, 0.9] for novelty, uncertainty and surprise, respectively.• Neuron Deletion Strategy: Given an input xt, the neuron deletion condition is:S(xt) > θS AND C(xt) > θC AND N (xt) < θN del , (2.3)where xt is the same as in Eq. (2.2), θN del , θC and θS are initialized neuron deletionthresholds in the range of [0.1, 0.8], [0.1, 0.3], [0.2, 0.9] for novelty, conflict andsurprise, respectively.• Parameter Update Strategy: When both neuron addition and deletion conditionsnot satisfied, it indicates the new input vector is a ’familiar’ data. The number ofhidden neurons will not be changed and the output weights are updated.A pseudocode description of C-ELM is given in Algorithm1.Algorithm 1 Pseudocode for Curious Extreme Learning Machine.1: Step 1: Present an input vector (xt, ct).2: Step 2: Compute four variables (novelty N (xt), uncertainty U(xt), conflict C(xt)and surprise S(xt)) according to the input vector.3: Step 3: Select one learning strategy out of three (Neuron Addition, Neuron Deletion,Parameters Update) based on the four variables and corresponding thresholds.4: Step 4: Increment t to t+1, repeat Step 1 to Step 3.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 292.5 Performance EvaluationIn this section, we first apply C-ELM and SVM to the feature vectors of the dataset.Since our data are limited, a 10-fold cross validation is utilized to gain insight into howour model will generalize to an independent dataset (i.e., how accurately this modelwill perform in practice). Then the Area Under the Curve (AUC) and Accuracy (ACC)are computed and used as the criteria for our performance evaluation. In addition, thetraining speeds of our C-ELM based model and the SVM based model are discussed andcompared.During a patient’s overnight sleep, the number of arousal events can range from tensto hundreds. Each event can last from several seconds to more than 15 seconds (currentlyno terminal criteria is established according to ASDA scoring rules [6] ). However, thetotal duration of all arousals during one night of sleep is quite small, around 20 or 30minutes out of 8 hours. That is to say, the data can be quite imbalanced when appliedto a classifier. Thus, the accuracy could be overestimated. In this study, there are only144 epochs among the total of 7680 epochs which are labeled as positive data (arousals)by sleep experts. To solve the imbalance problem, we perform classifications using thefollowing procedure.• First, 144 negative epochs are selected randomly from a total of 7536 non-arousalepochs. The 144 positive epochs are combined with the selected negative epochsto form a balanced dataset of 288 epochs in total.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 30• Second, randomize the dataset of 288 epochs obtained in the first step and divide itinto 10 folds for cross validation. Then we apply C-ELM and SVM to the random-ized dataset. Each one of the 10 folds is used as a test set in turn, with the other9 folds used as training sets. Thus, for each test fold, decision value of each inputepoch (decision value is used to determine the predicting result, such as positive ornegative in binary classification) is obtained.• Third, a Receiver Operating Characteristic (ROC curve) is plotted according todecision values obtained from the second step. For details of ROC curve, pleaserefer to Appendix D. Finally, AUC and ACC are computed from the ROC curve.• Repeat step 1 through step 3 for 50 times. The 50 AUC and ACC results arediscussed later in this section.In this thesis, the Library of Support Vector Machine (LIBSVM)[41] is used to trainand test data in the SVM based model. In the training step, Radial Basis Function (RBF)kernel function is used for the Support Vector Machine because RBF kernel usually hasa better performance for classification problems [29]. A grid search is utilized to tuneparameters in order to optimize the performance of the SVM based model. The C-ELMbased model is trained and tested using the source code from [28]. The parameters usedare the ones that provided the best classification performance in previous experimentsaccording to [28]. The learning thresholds are set as follows.• The low threshold of novelty = 0.1;Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 31Table 2.2: Average AUC comparison of C-ELM based model and SVM based modelProperties C-ELM based model SVM-based modelAverage AUC of 50 datasets 0.85 0.69Standard deviation of 50 datasets 0.0163 0.1573• The high threshold of novelty = 0.4;• The uncertainty threshold = 0.1;• The conflict threshold = 0.3;• The surprise threshold = 0.4;2.5.1 AUC and ACC EvaluationThe average AUC and ACC results of the 50 datasets for the C-ELM based model andthe SVM-based model are listed in Table. 2.2 and Table. 2.3, respectively. The standarddeviation of the 50 AUC and ACC results are also listed in Table. 2.2 and Table. 2.3.The best C-ELM based result and its corresponding SVM result are summarized inTable. 2.4 and the ROC curves are plotted in Fig. 2.2. The best SVM based result andits corresponding C-ELM result are summarized in Table. 2.5 and the ROC curves areplotted in Fig. 2.3.According to the results shown in two tables, an average AUC of 0.8527 and ACCof 0.7903 are achieved by our C-ELM based model while an average AUC of 0.6916 andACC of 0.6719 are obtained by SVM based model. These results indicates the sleepChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 32Table 2.3: Average ACC comparison of C-ELM based model and SVM based modelProperties C-ELM based model SVM-based modelAverage ACC of 50 datasets 0.79 0.67Standard deviation of 50 datasets 0.0179 0.1218arousal detection model based on C-ELM performs very good on our datasets and theSVM based detection model is relatively poor. This comparison result is consistent withthose of other problems reported in [28]. In [28], both C-ELM and SVM are evaluatedon the benchmark problems from the UCI machine learning repository which containsthree multicategory classification problems and three binary classification problems. Itis reported that C-ELM performs better than SVM on all the six problems. The overallaccuracy of C-ELM is greater than that of SVM by 0.12 for the Vehicle problem.The standard deviation of AUC of our C-ELM model is only 0.0163 while the SVMbased model reaches 0.1573. We can see that the best AUC achieved by C-ELM basedmodel is around 0.89 from Table. 2.4 and Fig. 2.2. The similarity between the bestresult and the average AUC 0.8527 and a relative small standard deviation of 0.0163may indicate the input data of most datasets among the 50 datasets are randomizedwell and the C-ELM based model is stable on all the 50 datasets. However, the averageAUC of the SVM based model is around 0.7. It is much smaller than the best AUCwhich is around 0.89. And we can also find that the standard deviations of 50 AUCs andACCS for the SVM based model are much greater than those of C-ELM based model.Because we apply the same dataset to both C-ELM based model and SVM based modelChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 33simultaneously, and good results of C-ELM based model have suggest the datasets usedare randomized well, The above-mentioned relatively poor results of the SVM basedmodel probably indicate this model is unstable on our datasets. The possible reasons forthe relatively poor performance of SVM based model compared to the model proposedby us are discussed below.• In our study, we used RBF kernel function for SVM based model, then two pa-rameters of the model and the kernel function need to be tuned to optimize theperformance and avoid overfitting. For each one of 50 datasets, we do grid searchesto choose the best values for the two parameters. When we do cross validationfor each dataset, grid search is applied to determine values of the parameters foreach 1 of 10 folds. The complexity in tuning parameters for SVM results in bigvariance of parameter’s values which may cause the unstable performance of SVMbased model. In regarding to this problem, we might choose another parametertuning strategy. For example, we apply grid search on each 1 of the 50 datasetsand then use a major vote method to determine one optimal value for each of the 2parameters. For all the cross validation procedures we can use the fixed parametersobtained by major vote.• Choosing the kernel function is probably the most tricky part of using SVM. Thekernel function is important because it creates the kernel matrix which summarizesall the data [42]. RBF kernel function is used in SVM classifier in our study becausethis kernel function is always a good try in various problems [29, 42]. However, whatChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 34could happen is that RBF kernel is not a good choice on our data. For example,if our data is linear distributed, we used RBF kernel instead of linear kernel withpoor parameters selected, this could cause over fitting problem which leads to a lesseffective classifier. Another seperate experiment is done to observe whether SVMbased model has an over fitting problem. It is shown that the average trainingaccuracy of 50 datasets is greater than the average testing accuracy by around 0.1which indicates over fitting problem might have occur in some of the 50 datasets.• Curious Extreme Learning Machine (C-ELM) is based on Extreme Learning Ma-chine (ELM). Compared to SVM, ELM has some advantages which may lead to abetter performance in our study. The hidden node parameters can be generatedwithout the knowledge of the training data and no parameter tuning is needed forELM [38]. The constraint of the choose of kernel is much smaller on ELM thanSVM. That is to say, ELM may generalize better than SVM regardless of kernelchoosing and the distribution of the data.• Curious Extreme Learning Machine (C-ELM) is an enhanced ELM. It is reported tohave a better performance than ELM on all the 3 binary classification benchmarkproblems studied in [28]. It reduces the randomization effect of ELM mainly byproviding an optimal number of hidden neurons. The hidden neuron addition ordeletion strategy based on curiosity may helps in avoiding over fitting.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 35Table 2.4: Best performance of C-ELM based model among 50 datasets and the corre-sponding performance of SVM based model of the same datasetProperties Best C-ELM based performance SVM based performanceAUC of one dataset 0.8843 0.82710 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Positive RateTrue Positive RatesvmAUC = 0.82711   celmAUC = 0.88431Figure 2.2: ROC curves for C-ELM based and SVM based models for the dataset whichgives the highest AUC for C-ELM. The red line is a random classification,the blue line is the curve of C-ELM and the green line is the curve of SVMTable 2.5: Best performance of SVM based model among 50 datasets and the corre-sponding performance of C-ELM based model of the same datasetProperties Best SVM based performance C-ELM based performanceAUC of one dataset 0.8850 0.8629Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 360 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Positive RateTrue Positive RatesvmAUC = 0.88498   celmAUC = 0.86285Figure 2.3: ROC curves for C-ELM based and SVM based models for the dataset whichgives the highest AUC for SVM. The red line is a random classification, theblue line is the curve of C-ELM and the green line is the curve of SVMChapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 372.5.2 Speed EvaluationIn order to do a fair comparison between the training times for the C-ELM based modeland the SVM based model, we use the built-in SVM training function of MATLABR2012b to do the classification instead of the function from LIBSVM[41] since the LIB-SVM utilizes c/c++ source code (Matlab is slower than C/C++ which would make thecomparison unfair). The kernel function of the SVM classifier is RBF. A total of 288observations, each with 22 features (288*22 matrix), containing 144 positive data and144 negative data were selected randomly from the 7680 observations. The training timesfor both models are shown in Table. 2.6. It can be seen that the training speeds for thetwo models are similar. This result is consistent with those reported in [28]. In [28],the training times for C-ELM and SVM are similar for all the three benchmark binaryclassification problems. For example, for Brest cancer problem, The training time ofSVM is 0.11 while that of C-ELM is 0.09. A total of 300 training data with dimensionof 9 are used in the Breast cancer problem. In this thesis, it is just a rough comparisonfor the specific dataset. The training time depends on the dataset, kernel function usedas well as the coding implementation of the algorithm and so on. In addition, we don’ttune parameters in our model while a grid search is utilized to optimize the SVM basedmodel. If considering the total executing time of the sleep arousal detection model, theSVM based model is slower than the C-ELM based model and the executing time of SVMbased model depends on the complexity of the grid search. A more thorough evaluationof the training times is required but this is not the main aim of this thesis.Chapter 2. Automatic Sleep EEG Arousal Detection based on C-ELM 38Table 2.6: Training times for C-ELM based model and SVM based model for a datasetwith dimension of 22C-ELM based model SVM based modeltraining time (seconds) 0.079 0.0802.6 SummaryIn this Chapter, a new model based on a new set of 22 features and Curious ExtremeLearning Machine (C-ELM) for sleep arousal detection has been proposed. Data acqui-sition, preprocessing and segmentation are first described followed by the feature extrac-tion procedure. Brief descriptions of C-ELM and SVM classification algorithms are alsoprovided. The performance of the new model is presented and compared to that of aSVM-based model. It is found that the proposed model of sleep arousal detection hasa good performance even though only one single EEG channel and limited data is used.The proposed model has a higher AUC and ACC which indicates a better ability tocorrectly classifies a random data as a sleep arousal or a non-arousal while its trainingspeed is similar to that for the SVM based model.39Chapter 3Automatic Detection based onC-ELM and MRMRIn this Chapter, we present an improved sleep arousal detection algorithm. In thisalgorithm, the Minimum Redundancy Maximum Relevance (MRMR) feature selectionmethod is added to the previous mentioned C-ELM classifier based algorithm. In Sec-tion.3.1, the improved model is illustrated and a brief introduction is provided. In Sec-tion.3.2, various feature selection methods are described and discussed including theMRMR approach utilized by our algorithm. In Section.3.3, the improved algorithm isapplied to the sleep dataset. The performance is evaluated including providing the aver-age AUC and ACC of 50 datasets. For each dataset, the AUCs and ACCs are computedwhen different feature subsets are used for C-ELM based algorithm. A brief summary ofthis chapter is provided in Section.3.4.Chapter 3. Automatic Detection based on C-ELM and MRMR 403.1 Improved Arousal Detection ModelThis improved model is based on the model proposed in Chapter2. Most of the methodsproposed for automatic sleep arousal detection do not have a feature selection step.However, it is reported in [43] has reported that the selection of different feature subsetshave a significant influence on the sleep arousal detection. All of the studies mentionedin Section 1.2 select features from previous works or add new features from their ownperspectives. The influence of different feature subsets on the performance of arousaldetection methods was not reported in these studies. In our model, a ranking of all the22 features is obtained using the MRMR feature selection method [44]. We applied theC-ELM classifier on different feature subsets according to the ranking. It is found thata subset of the 17 highest ranked features has the best performance and a feature setof the 6 highest ranked features achieves a similar performance with the 17 feature set.Dimensionality reduction of the input vectors can reduce the complexity and trainingtime of model while keeping a reasonable performance. The improved arousal detectionmodel is shown in Fig. Feature SelectionIn machine learning problems, the experimental performances can be negatively influ-enced by data dimensionality [43]. In some real problems, a small number of high-dimensional data may cause over fitting problem. Although the amount of data neededChapter 3. Automatic Detection based on C-ELM and MRMR 41One channel EEG signal Band pass filter Segmentation Time frequency feature extraction PSG signals Sleep stage annotation Wake Stage 1 Stage 2 Stage 3 Sleep arousal annotation C-ELM classification Non-REM 1     2  21   22        … Feature  matrix         … MRMR feature selection 1      2  j -1     j       … Selected feature matrix (1<=j<=22)  Figure 3.1: Improved sleep detection model based on MRMR feature selection.Chapter 3. Automatic Detection based on C-ELM and MRMR 42to properly train a model may not be obvious, dimensionality reduction of the input datamay be of benefit in our study. First, dimensionality reduction of the input data can re-duce the complexity of our model and the training time of C-ELM classifier. Second,although we do not have a huge amount of features, EEG signals obtained from patientsalways have big noise which results in noisy features which may mislead the classificationalgorithm thereby reduces the accuracy of sleep arousal detection. It is thus importantto add a feature selection step in our sleep arousal detection model. In this section, twomain categories of feature selection methods are introduced and compared. The MRMRfeature selection method which is adopted in our model is briefly described as well.3.2.1 Filters and WrappersFeature selection methods can be roughly grouped into two main categories: filters andwrappers [43]. Filter methods carry out the selection step based on intrinsic characteris-tics of the training data to determine their relevance or discriminant power with regardsto the target classes (the true label for each observation, named positive or negative inbinary classifications) [45]. Filter methods are totally independent of classifiers used inthe classification step. Wrappers use induction algorithms, e.t. Multi-Layer Perceptron(MLP) and Support Vector Machine (SVM) are used as induction algorithms in [43]to explore each subset of features. During the induction process, wrapper methods aredependent on classifiers.Filter methods give a ranking of input features or a subset of significant featuresChapter 3. Automatic Detection based on C-ELM and MRMR 43based on different measures of the input data and corresponding classes. Filters based oninformation gain, information entropy, mutual information and statistical tests (such ast-test, F-test) etc. have been developed [43, 45]. Different filter feature selection methodsare expected to have significantly different rankings of features since various measuresare used. An effective filter method is believed to improve the classification performancewhile reducing computing time. Another advantage for filter methods is that filters areindependent of the learning algorithms.For wrapper methods, feature selection is wrapped around a learning algorithm suchas SVM. The effectiveness of a feature is decided by the estimated accuracy of the learningalgorithm. Two well known strategies utilized in wrapper methods are sequential forwardselection (SFS) and sequential backward selection (SBS) [43]. SFS starts with an emptyfeature set and add features one by one while SBS starts with a full feature set anddelete features one by one [43]. Wrapper methods can give high accuracy if the learningalgorithm used in the classification step is the same one as used in the wrapper. However,wrappers require a long running time when the dataset is big because they need totrain an induction algorithm numerous times. Moreover, wrapper methods have lowergeneralization ability than filters as they depend on the learning algorithm.In our sleep arousal detection problem, time efficiency is quite important since thedataset obtained from patients is usually quite big. Thus, in our study, a filter fea-ture selection method named minimum redundancy-maximum relevance (MRMR) [45] isadopted instead of wrapper methods.Chapter 3. Automatic Detection based on C-ELM and MRMR 443.2.2 MRMR Feature SelectionIn common filter methods, a simple ranking of features is obtained based on some specificmeasures (information gain, mutual information etc.). Then we can select m features outof the total n features (m<=n). One deficiency of these approaches is that the selectedm features can be correlated among themselves. Thus, redundancy of features can stillexist in the feature set although this set has strong predict power in classification. Thisissue can lead to two main problems [45]. (1) If a feature set contains highly mutualcorrelated features, then the true unique features are fewer, and some of the featuresare wasted.(2) The feature set is “narrow” because it can only represent one or a fewdominant characteristics of the data which limits the generalization ability of the featureset.Minimum Redundancy-Maximum Relevance (MRMR) is a filter method which re-quires features to be maximally dissimilar to each other (minimum redundancy) whilekeeping the maximum relevance criteria used in other filter methods such as maximizingthe mutual information between the features and the target classes. With this approach,a smaller feature set with better representative and generalization properties which re-duces complexity of the model may be obtained.A brief description of MRMR feature selection for discrete variables is provided below.For more details, please refer to [44, 45]. The minimum redundancy condition isminWI , WI =1|S|2∑i,j∈SI(i, j), (3.1)Chapter 3. Automatic Detection based on C-ELM and MRMR 45where I(i, j) is the mutual information between two different features i and j, S is theselected feature set, |S| is the number of features in S.The maximum relevance condition ismaxVI , VI =1|S|∑i∈SI(h, i), (3.2)where I(h, i) is the mutual information between feature i and classes h = {h1, h2, ..., hK}.Two schemes - Mutual Information Difference (MID) (Eq. (3.3)) and Mutual In-formation Quotient (MIQ) (Eq. (3.4)) are utilized to optimize Eq. (3.1) and Eq. (3.2)simultaneously. Optimization of both conditions requires combining them into a singlecriterion function. Since the two conditions are equally important, two simplest combi-nation criteria (Eq. (3.3) and Eq. (3.4)) are considered [45].max(VI −WI), (3.3)max(VI/WI), (3.4)The feature selection step works as follows. The first feature is selected according toEq. (3.2), i.e. the feature with the highest I(h, i). Earlier features selected remain in theset, a new feature is selected and added according to one of the two criteria (Eq. (3.3)and Eq. (3.4)). That is to say, the last added feature is the one with the lowest rank.Finally, a ranking of all the features can be obtained.Chapter 3. Automatic Detection based on C-ELM and MRMR 463.3 Performance EvaluationIn this section, we first apply MRMR feature selection on the 7680 input epochs men-tioned in Chapter 2. Then a ranking of the 22 features is obtained using the featureselection process. The ranking of the data used in our study is listed in Table. 3.1. Ide-ally, it is a better way to do the feature selection step using a different dataset from thetraining and testing datasets. However, in our study, very limit data was used, thus, asub-optimal way was utilized (we use the 7680 data to do feature selection, and the 50datasets used to do performance evaluation are also from the 7680 data). For detailedexplanation of the features, please refer to Sec.2.3. In our study, both Mutual Informa-tion Quotient (MIQ) and Mutual Information Difference (MID) criteria are tried in thefeature selection step. It is found that MIQ criterion performs much more effective andthus MIQ criterion is utilized as the MRMR feature selection scheme in our model.According to the ranking, different feature subsets are used as input feature vectorsto train the C-ELM classifier. 50 datasets are randomly chosen using the same steps inChapter 2 to evaluate the performance (the same logic as Chapter 2). AUC and ACCare used as the criteria for our performance evaluation. The steps for the performanceevaluation are described in Algorithm 2.For each of the 50 datasets, we have 22 AUCs and ACCs computed for 22 featuresets. The first feature set only contains the highest ranked feature; the second featureset contains the top 2 features and so on. The 22nd feature set contains a full set of 22features. The average performance of the 50 datasets for each feature set is computed.Chapter 3. Automatic Detection based on C-ELM and MRMR 47Table 3.1: MRMR feature ranking using MIQ schemeFeature rank Feature name1 Mean frequency2 3-sec delta power ratio3 1-sec alpha power ratio4 max power frequency5 1-sec power (0-50 Hz) ratio6 sleep stage annotation7 zero-crossing frequency8 3-sec theta ratio9 mean value10 1-sec power11 3-sec alpha power ratio12 standard deviation13 1-sec sigma ratio14 3-sec power (0-50 Hz) ratio15 3-sec sigma power ratio16 3-sec gamma power ratio17 sleep spindle18 1-sec beta power ratio19 3-sec beta power ratio20 1-sec delta power ratio21 1-sec theata ratio22 1-sec gamma ratioChapter 3. Automatic Detection based on C-ELM and MRMR 48Algorithm 2 The algorithm for the improved model’s performance evaluation.for iteration i := 1 to 50 doselect 144 negative epochs (each epoch is a 22-dimension vector) randomly from atotal of 7536 non-arousal epochs. 144 positive epochs are combined with the selectednegative epochs to form a balanced dataset of 288 epochs in total.for iteration j := 1 to 22 dochoose the top j features according to the MRMR ranking for each input vector.Thus, the current dataset becomes a 288*j input dataset.randomize the dataset obtained in the previous step and divide it into 10 folds forcross validation. Then C-ELM classifier is applied on the dataset. For each testfold, decision values (probabilities of positive or negative) are obtained.A ROC curve is plotted according to decision values and AUC, ACC are computedfrom the ROC curve.end forend forChapter 3. Automatic Detection based on C-ELM and MRMR 49Results are shown in Fig. 3.2 and Fig. 3.3.From Fig. 3.2, it can be seen that the best average AUC achieved is 0.86 when aset of 17 top-ranking features are selected. We also observe that a set of the 6 top-ranking features can achieve a reasonable good performance with an average AUC of0.85. Fig. 3.3 shows that the best average ACC (0.80) is obtained with the 6 top rankedfeatures. Using the full set of 22 features, the average AUC is 0.85 while the averageACC is 0.79. Thus, the MRMR feature selection successfully improves the average AUCfrom 0.85 to 0.86 while reducing the number of features from 22 to 17. Moreover, wecan reduce the number of features to 6 and thereby reduce the training time of ourmodel while maintaining a similar performance. In order to find out the training timesof dataset with different number of features. Another experiment is done on a trainingdata set of 288 observations (each observation is a m-dimensional input vector, here m isthe number of features which is between 1 and 22). It is observed that the training timeis 0.043 seconds for the dataset with 6 features while the training time achieves 0.079secons for the dataset with 22 features. However, the relationship between the trainingtime and the number of features is not a simple linear correlation because the trainingtime depends on the kernel function you used, the convergence time of the method usedto find a separating hyperplane (in SVM) and so on. This may be a good topic to workon, however, in our study, it is not the main topic to discuss. The average AUC andACC of input data with different feature set size are listed in Table. 3.2.As above-mentioned, we have big noise in EEG signal. Thus, we may have noisyChapter 3. Automatic Detection based on C-ELM and MRMR 500 5 10 15 20 250.820.8250.830.8350.840.8450.850.8550.860.865 Average AUC for different feature setsfeature set sizeAverage AUCAUC=0.85number offeatures:6AUC=0.86number offeatures:17Figure 3.2: Average AUC plot for different feature sets. The x-axis is the feature setsize.Feature set of size 1 contains top 1 ranking feature, feature set of size2 contains 2 top ranked features and so on.Chapter 3. Automatic Detection based on C-ELM and MRMR 510 5 10 15 20 250.780.7850.790.7950.80.8050.81 Average ACC for different feature setsfeature set sizeAverage ACCACC=0.80number of features : 6Figure 3.3: Average ACC plot for different feature sets. The x-axis is the feature setsize.Feature set of size 1 contains top 1 ranking feature, feature set of size2 contains 2 top ranked features and so on.Chapter 3. Automatic Detection based on C-ELM and MRMR 52Table 3.2: Average performance of different feature sets using C-ELMNumber of features Average AUC Average ACC1 0.82 0.786 0.85 0.8017 0.86 0.7922 0.85 0.79features in the feature set. In our study, we only use one channel of EEG signal, so thepredictive power of a feature is important. It is observed from Table. 3.2 that the AUCand ACC of the 6th feature set is higher than those of others which indicates that on topof the 6 top ranked features, the remaining features does not provide as much predictpower as the 6 top ranked features. The 6 top features are mean frequency, 3-sec/10-sec delta power ratio, 1-sec/10-sec alpha ratio, max power frequency, 1-sec/10-sec powerratio (0-50 Hz), and sleep stage annotations. From the 6 features we can see that powerratios have made good contributions to a effective feature set which indicates the powerchanges could represent EEG frequency shift to some extent.As a supplementary, we also applied MRMR feature selection to the SVM basedmodel. The best average AUC of 0.69 and ACC of 0.67 are achieved by the set of 6 topranked features (the feature ranking is the same as above-mentioned), while the averageAUC and ACC for a full set of features are the same as those of the set with 6 features.part of the AUCs and ACCs for different feature sets are listed in Table. 3.3. This resultindicates a good generalization of the 6 top ranked features regardless of the classifier.Chapter 3. Automatic Detection based on C-ELM and MRMR 53Table 3.3: Average performance of different feature sets using SVMNumber of features Average AUC Average ACC1 0.65 0.656 0.69 0.6722 0.69 0.673.4 SummaryIn this Chapter, an improved non-REM sleep arousal detection model with a MinimumRedundancy- Maximum Relevance (MRMR) feature selection step is presented. Severalfeature selection methods are briefly introduced, followed by an illustration of the MRMRfeature selection. The performance of this proposed model with MRMR is evaluatedusing the AUC and ACC criteria. It is found that the sleep arousal detection modelcan provide a similar performance with a reduced feature set size. It should be notedthat only one single EEG channel and limited data were used in our simulations. Betterperformance may be achieved if the model is trained on a larger dataset or more channelsof physiological signals are utilized.54Chapter 4Conclusions and Future WorkIn this chapter, we conclude this thesis by summarizing the research results and contri-butions. Future research topics are suggested as well.4.1 ConclusionsWe studied the problem of automatic detection of sleep arousals. A new detection modelwas proposed based on a set of 22 features and Curious Extreme Learning Machine (C-ELM) in Chapter 2. This model was found to provide a good performance on the datasetused in our study. A Support Vector Machine (SVM) based model was also evaluated forcomparison with our model. In Chapter 3, an improved detection model was presented,in which a Minimum Redundancy Maximum Relevance (MRMR) feature selection stepis added to the model proposed in Chapter 2. The improved model allows a reduction ofthe size of the feature set, and have a decreased training time while maintaining a similardetection performance.• In Chapter.2, we presented a new model for sleep arousal detection. In this model,data was first preprocessed and segmented. Then, a proposed set of 22 features areChapter 4. Conclusions and Future Work 55extracted, followed by the use of a Curious Extreme Learning Machine (C-ELM)classifier. An average Area Under the ROC Curve (AUC) of 0.85 (an AUC of 1corresponds to perfect classification, whereas an AUC of 0.5 corresponds to randomclassification) and an average accuracy (ACC) of 0.79 was achieved by the proposedmodel while an average AUC and ACC of 0.69 and 0.67 respectively was achievedfor the SVM based model. The results indicates that our system for sleep arousaldetection has a high performance on the dataset utilized in this study. In addition,the training speed of C-ELM is similar to that of SVM and the total executing timeof our model is less than SVM based model since a grid search was done to optimizethe SVM based model which increased the total running time. The detailed runningtime of the SVM based model varies a lot based on the complexity of grid search.• In Chapter.3, we proposed an improved model for sleep arousal detection basedon the model proposed in Chapter.2 by adding a Minimum Redundancy MaximumRelevance feature selection step to remove redundant features. Using this improvedmodel, it was found that the size of the feature set could be reduced to 6 from 22without a significant performance change while reducing the training time of theclassifier. The average AUC and ACC achieved by a 6-feature model are 0.85 and0.80 respectively while the average AUC and ACC obtained by a model with the fullset of 22 features were 0.85 and 0.79 respectively. The best average AUC achievedwas 0.86 with an average ACC of 0.79 when using a set of 17 features. The resultsof this chapter suggests adding an effective feature selection step (such as MRMRChapter 4. Conclusions and Future Work 56feature selection) in automatic sleep arousal detection system is significant.4.2 Future WorkSome possible extensions of the research work on automatic sleep arousal detection areoutlined below based on what have been observed and learnt in this project.• The data used in our study is limited. A better performance may be expected ifwe have a larger dataset to train the classifier. It would be interesting to apply theproposed model on a big dataset obtained from real patients.• During one patient’s overnight sleep, arousal events happen frequently. However,the events usually only last several seconds and the total time of all arousals duringa night is quite short, typically 20 or 30 minutes out of 8 hours. Thus, the imbalancebetween positive and negative data is a big issue no matter in training a classifieror doing a performance evaluation. In our study, we simply choose datasets withequal number of positive and negative data. Other methods can be explored tosolve the imbalance problem such as using over or under sampling strategies [46].• In our study, we applied a bandpass filter to the dataset for preprocessing in ourmodel. EEG signals can have a lot of noise caused by the movement during sleep.A considerable range of methods have been proposed to remove artifacts if multi-channel EEG recordings are used. However, few methods have been proposed toChapter 4. Conclusions and Future Work 57remove artifacts of a single channel EEG [47]. Thus, studies on artifact removal fora single channel EEG would be useful for sleep arousal detection.• In our work, we added a MRMR feature selection step to the model proposed andhave observed a good performance. This result indicates the significance of featureselection step. Thus, how to choose an effective feature selection method for sleeparousal detection would be another interesting topic.58Bibliography[1] A. L. Goldberger, L. A. N. Amaral, et al., “PhysioBank, PhysioToolkit, andPhysioNet: Components of a new research resource for complex physiologic signals,”Circulation, vol. 101, no. 23, pp. e215–e220, 2000 (June 13), circulation ElectronicPages: PMID:1085218;doi: 10.1161/01.CIR.101.23.e215. [Online]. Available:[2] Wikipedia, “Sleep — wikipedia the free encyclopedia,” 2015, online; accessed22-June-2015. [Online]. Available:[3] ——, “Receiver operating characteristic — wikipedia the free encyclopedia,” 2015,[Online; accessed 22-June-2015]. [Online]. Available: operating characteristic&oldid=667584494[4] “Report of the committee on methods of clinical examination in elec-troencephalography: 1957,” Electroencephalography and Clinical Neuro-physiology, vol. 10, no. 2, pp. 370–375, 1958. [Online]. Available: 59[5] P. Halsz, M. Terzano, L. Parrino, and R. Bdizs, “The nature of arousal in sleep,”Journal of Sleep Research, vol. 13, no. 1, pp. 1–23, 2004. [Online]. Available:[6] M. Bonnet, D. Carley, et al., “EEG arousals: Scoring rules and examples: A pre-liminary report from The Sleep Disorders Atlas Task Force Of The American SleepDisorders Association,” Sleep, vol. 15, pp. 173–184, 1992.[7] A. Rechtschaffen, A. Kales, et al., A Manual of Standardized Terminology,Techniques and Scoring System for Sleep Stages of Human Subjects, ser. Publication.Brain Information Service/Brain Research Institute, University of California, 1968.[Online]. Available:[8] L. Gennaro, M. Ferrara, and M. Bertini, “EEG arousals in normal sleep: Variationsinduced by total and selective slow-wave sleep deprivation,” Sleep, vol. 24, no. 6,2001.[9] J. Fleetham, P. West, et al., “Sleep, arousals, and oxygen desaturation in chronicobstructive pulmonary disease. the effect of oxygen therapy,” The American reviewof respiratory disease, vol. 126, no. 3, p. 429433, September 1982. [Online].Available:[10] C. Guilleminault, R. Stoohs, et al., “A cause of excessive daytime sleepiness. theupper airway resistance syndrome.” Chest, vol. 104, no. 3, pp. 781–787, 1993.[Online]. Available: + 60[11] K. Cheshire, H. Engleman, et al., “Factors impairing daytime performance in pa-tients with sleep apnea/hypopnea syndrome,” Archives of Internal Medicine, vol.152, no. 3, pp. 538–541, 1992.[12] P. Collard, M. Dury, et al., “Movement arousals and sleep-related disordered breath-ing in adults.” American Journal of Respiratory and Critical Care Medicine, vol. 154,no. 2, pp. 454–459, 1996.[13] M. Bonnet, K. Doghramji, et al., “The scoring of arousal in sleep: reliability, validity,and alternatives,” J Clin Sleep Med, vol. 3, no. 2, pp. 133–145, 2007.[14] M. Drinnan, A. Murray, et al., “Automated recognition of EEG changesaccompanying arousal in respiratory sleep disorders,” Sleep, vol. 19, no. 4, p. 296303,May 1996. [Online]. Available:[15] S.-P. Cho, J. Lee, H. Park, and K. Lee, “Detection of arousals in patients withrespiratory sleep disorders using a single channel EEG,” in Engineering in Medicineand Biology Society, 2005. IEEE-EMBS 2005. 27th Annual International Conferenceof the, Jan 2005, pp. 2733–2735.[16] D. J. Buysse et al., “The Pittsburgh sleep quality index: A new instrument forpsychiatric practice and research,” Psychiatry Research, vol. 28, no. 2, pp. 193 –213, 1989. [Online]. Available: 61[17] G. Sorensen, J. Kempfner, P. Jennum, and H. Sorensen, “Detection of arousalsin parkinson’s disease patients,” in Engineering in Medicine and Biology Society,EMBC, 2011 Annual International Conference of the IEEE, Aug 2011, pp. 2764–2767.[18] R. J. Thomas, “Arousals in sleep-disordered breathing: patterns and implications,”Sleep, vol. 26, no. 8, pp. 1042–1048, 2003.[19] M. Grigg-Damberger, D. Gozal, et al., “The visual scoring of sleep and arousal ininfants and children,” J Clin Sleep Med, vol. 3, no. 2, pp. 201–240, 2007.[20] T. Sugi, F. Kawana, and M. Nakamura, “Automatic EEG arousal detection forsleep apnea syndrome,” Biomedical Signal Processing and Control, vol. 4, no. 4, pp.329 – 337, 2009, special Issue on Biomedical Systems, Signals and Control ExtendedSelected papers from the {IFAC} World Congress, Seoul, July 2008. [Online].Available:[21] R. Agarwal, “Automatic detection of micro-arousals,” IEEE Engineering inMedicine and Biology Society.Annual Conference, vol. 2, p. 1158, 2005.[22] F. De Carli, L. Nobili, P. Gelcich, and F. Ferrillo, “A method for the automaticdetection of arousals during sleep,” Sleep, vol. 22, no. 5, p. 561, 1999.[23] D. A´lvarez Este´vez and V. Moret-Bonillo, “Model comparison for the detectionof EEG arousals in sleep apnea patients,” in Proceedings of the 10th Interna-Bibliography 62tional Work-Conference on Artificial Neural Networks: Part I: Bio-Inspired Systems:Computational and Ambient Intelligence, ser. IWANN ’09, 2009, pp. 997–1004.[24] O. Shmiel, T. Shmiel, et al., “Data mining techniques for detection of sleep arousals,”Journal of Neuroscience Methods, vol. 179, no. 2, pp. 331 – 337, 2009. [Online].Available:[25] G. Pillar, A. Bar, et al., “Autonomic arousal index: an automated detection basedon peripheral arterial tonometry.” Sleep, vol. 25, no. 5, pp. 543–549, 2002.[26] ——, “An automatic ambulatory device for detection of AASM defined arousalsfrom sleep: the WP100,” Sleep Medicine, vol. 4, no. 3, pp. 207–212, 2003.[27] G. H. Klem, H. O. Lu¨ders, H. Jasper, and C. Elger, “The ten-twenty electrodesystem of the International Federation of Clinical Physiology,” ElectroencephalogrClin Neurophysiol, vol. 52, no. suppl. 52, p. 3, 1999.[28] Q. Wu and C. Miao, “C-ELM: A curious extreme learning machine for classificationproblems,” in Proceedings of ELM-2014 Volume 1, ser. Proceedings in Adaptation,Learning and Optimization. Springer International Publishing, 2015, vol. 3, pp.355–366. [Online]. Available: 30[29] C.-W. Hsu, C.-C. Chang, et al., “A practical guide to support vector classification,”2003. [Online]. Available:∼cjlin/papers.htmlBibliography 63[30] T. Joachims, Introduction to support vector machines. Cambridge University Press,2002.[31] N. Cristianini and J. Shawe-Taylor, An introduction to support vector machines andother kernel-based learning methods. Cambridge University Press, 2000.[32] D. Fradkin and I. Muchnik, “Support vector machines for classification,” Discretemethods in epidemiology, vol. 70, pp. 13–20, 2006.[33] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: theory andapplications,” Neurocomputing, vol. 70, no. 1, pp. 489–501, 2006.[34] Q. Yuan, W. Zhou, S. Li, and D. Cai, “Epileptic EEG classification based on extremelearning machine and nonlinear features,” Epilepsy research, vol. 96, no. 1, pp. 29–38,2011.[35] X.-K. Wei, Y.-H. Li, and Y. Feng, “Comparative study of extreme learning machineand support vector machine,” in Advances in Neural Networks-ISNN 2006. Springer,2006, pp. 1089–1095.[36] R. Rajesh and J. S. Prakash, “Extreme learning machines: a review and state-of-the-art,” International Journal of Wisdom Based Computing, vol. 1, no. 1, pp. 35–49,2011.Bibliography 64[37] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: a new learningscheme of feedforward neural networks,” in Neural Networks, 2004. Proceedings.2004 IEEE International Joint Conference on, vol. 2. IEEE, 2004, pp. 985–990.[38] G.-B. Huang, “Extreme learning machine: learning without iterative tuning,” 2010.[Online]. Available:[39] S. Suresh, S. Saraswathi, and N. Sundararajan, “Performance enhancement of ex-treme learning machine for multi-category sparse data classification problems,” En-gineering Applications of Artificial Intelligence, vol. 23, no. 7, pp. 1149–1157, 2010.[40] Z. Bai, G.-B. Huang, et al., “Sparse extreme learning machine for classification,”Cybernetics, IEEE Transactions on, vol. 44, no. 10, pp. 1858–1870, 2014.[41] C. C. Chang and C. J. Lin, “LIBSVM: A library for support vector machines,”ACM Transactions on Intelligent Systems and Technology, vol. 2, pp. 27:1–27:27,2011, software available at∼cjlin/libsvm.[42] M. Law, “A simple introduction to Support Vector Machines,” Dept.of Computer Science and Eng. Michigan State Univ. [Online]. Available: svm new.pdf[43] D. lvarez Estvez, N. Snchez-Maroo, et al., “Reducing dimensionality in a databaseof sleep EEG arousals,” Expert Systems with Applications, vol. 38, no. 6, pp. 7746 –7754, 2011. [Online]. Available: 65[44] H. Peng, F. Long, and C. Ding, “Feature selection based on mutual informationcriteria of max-dependency, max-relevance, and min-redundancy,” Pattern Analysisand Machine Intelligence, IEEE Transactions on, vol. 27, no. 8, pp. 1226–1238, 2005.[45] C. Ding and H. Peng, “Minimum redundancy feature selection from microarray geneexpression data,” Journal of bioinformatics and computational biology, vol. 3, no. 02,pp. 185–205, 2005.[46] N. V. Chawla, “Data mining for imbalanced datasets: An overview,” in Data miningand knowledge discovery handbook. Springer, 2005, pp. 853–867.[47] C. H. Xun Chen and H. Peng, “Removal of muscle artifacts from single-channel EEGbased on ensemble empirical mode decomposition and multiset canonical correlationanalysis,” Journal of Applied Mathematics, p. 10, 2014.[48] Wikipedia, “10-20 system (eeg) — wikipedia the free encyclopedia,” 2015, [Online;accessed 22-June-2015]. [Online]. Available: system (EEG)&oldid=66471623866Appendices67Appendix ASleep StagesUsually sleepers pass through five sleep stages: 1, 2, 3, 4 and REM (rapid eye movement)sleep. Sleep stage 3 and 4 are always combined into one sleep stage (we use sleep stage3 to represent sleep stage 3 and 4). Sleep stage 1 to 3 are called non-REM sleep stages.These stages progress cyclically from 1 through REM stage. An overnight sleep of ahealthy individual usually contains 4 to 5 sleep cycles as shown in Fig. A.1. Sleep stage1 is known as a transitional stage usually occurs between sleep and wakefulness. In thisstage, the brain produces high amplitude and low frequency theta waves. Brain wavesduring Sleep stage 2 are mainly in the theta wave range. This sleep stage is characterizedby two phenomena: sleep spindles and K-Complex. Sleep stage 3 is known as slow wavesleep or deep sleep characterized by delta wave along with sleep spindles, although muchfewer than sleep stage 2. REM sleep stage is a stage during which EOG shows a rapideye movement. Dreams often occur in this stage.Appendix A. Sleep Stages 68Figure A.1: Sleep stages [2]69Appendix BBasic Concept (sensitivity, etc.)In a two-class prediction problem (binary classification), in which the outcomes are la-beled either as positive (p) or negative (n). There are four possible outcomes from abinary classifier. If the outcome from a prediction is p and the actual value is also p,then it is called a true positive (TP); however if the actual value is n then it is saidto be a false positive (FP). Conversely, a true negative (TN) has occurred when boththe prediction outcome and the actual value are n, and false negative (FN) is when theprediction outcome is n while the actual value is p [3]. Please see Fig. B.1.Please see Table. B.1 for computing sensitivity, specificity, accuracy, selectivity, posi-tive predictive value (PPV) and so on.Figure B.1: Binary classification basic concept [3]Appendix B. Basic Concept (sensitivity, etc.) 70Table B.1: Computing accuracy, sensitivity and so onThe name of the measure Computing formulaAccuracy (ACC) (TP+TN)/Total populationTrue positive rate (TPR), Sensitivity TP/Condition positiveTrue negative rate (TNR), Specificity TN/Condition negativeFalse positive rate (FPR), Fall-out FP/Condition negativeFalse negative rate (FNR), Miss rate FN/Condition positivePositive predictive value (PPV), Precision TP/Test outcome positiveSelectivity TP/(TP+FP)71Appendix CEEG 10–20 International SystemThe 10-20 system or International 10-20 system is an internationally recognized methodto describe and apply the location of scalp electrodes in the context of an EEG test orexperiment. This method was developed to ensure standardized reproducibility so thata subject’s studies could be compared over time and subjects could be compared to eachother. This system is based on the relationship between the location of an electrode andthe underlying area of cerebral cortex. The “10” and “20” refer to the fact that the actualdistances between adjacent electrodes are either 10% or 20% of the total front-back orright-left distance of the skull [4, 27, 48]. Please see Fig. C.1.Appendix C. EEG 10–20 International System 72Figure C.1: EEG 10-20 system [4]73Appendix DROC CurveA receiver operating characteristic (ROC), or ROC curve, is a graphical plot that il-lustrates the performance of a binary classifier system as its discrimination threshold isvaried. The curve is created by plotting the True positive rate (Sensitivity) against theFalse positive rate (1-Specificity) at various threshold settings [3]. The Area Under theROC Curve indicates the ability of a classifier to discriminate a positive data from anegative data. A value of 1 means a perfect test while 0.5 means random classification.


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items