"IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 3, MAY 2012 469 Real-Time Sleep Apnea Detection by Classi?er Combination Baile Xie, Student Member, IEEE, and Hlaing Minn, Senior Member, IEEE Abstract—To ?nd an ef?cient and valid alternative of readily available, relatively inexpensive, and reliable diagnosis polysomnography (PSG), this paper investigates real-time sleep alternative is desirable for public use. apnea and hypopnea syndrome (SAHS) detection based on electro- In the past decade, several methods have been proposed cardiograph (ECG) and saturation of peripheral oxygen (SpO ) 2 as PSG alternatives for SAHS detection. For instance, detec- signals, individually and in combination. We include ten machine- tions have been developed based on questionnaires [6], snor- learning algorithms in our classi?cation experiment. It is shown that our proposed SpO features outperform the ECG features in ing [7], electrocardiograph (ECG) [8]–[12], and pulse oxime- 2 terms of diagnostic ability. More importantly, we propose classi?er try [5], [12]–[21]. Among them, ECG and saturation of oxygen combination to further enhance the classi?cation performance by measured by pulse oximeter (SpO ) are the two most extensively 2 harnessing the complementary information provided by individual studied signals. classi?ers. With our selected SpO and ECG features, the classi?er 2 Speci?cally, based on the ECG signal alone, various fre- combination using AdaBoost with Decision Stump, Bagging with REPTree, and either kNN or Decision Table achieves sensitivity, quency and time domains as well as ECG morphology features speci?city, and accuracy all around 82% for a minute-based real- have been developed. For example, McNames and Fraser [8] time SAHS detection over 25 sleep-disordered-breathing suspects’ found that the heart rate (HR), the S-pulse amplitude, and the full overnight recordings. pulse energy of the ECG signal are informative for SAHS detec- Index Terms—Classi?er combination, electrocardiograph tion. Raymond et al. [9] analyzed the power spectral features of (ECG), feature selection, hypopnea, machine learning, saturation the ECG-derived respiration (EDR) signal via wavelet transform of peripheral oxygen (SpO ), sleep apnea. 2 in addition to other features from the RR interval tachogram. Shinar et al. [10] extracted features from the changes of ECG signal’s QRS complex and spectral abnormalities of heart rate I. INTRODUCTION variability (HRV). LEEP apnea and hypopnea syndrome (SAHS) is a com- SpO is the percentage (%) of hemoglobin in the blood that 2 mon sleep disorder which is characterized by abnormal S is saturated with oxygen recorded by a pulse oximeter. Some breath pause or reduction during sleep. It is estimated to affect commonly used SpO features are the accumulative time (TSA) 2 2% of middle-aged women and 4% of middle-aged men [1]. spent below a certain saturation level [5], [14], the oxygen Sleep apnea is treatable; however, about 90% of sufferers go desaturation index (ODI, the number of oxyhemoglobin de- unidenti?ed and hence untreated [2]. They experience daytime saturation below a certain threshold) [15], and the saturation sleepiness and fatigue which can escalate to traf?c accidents, variability index (Delta index) [5], [13], [14]. Later on, several depression, and memory loss. Moreover, untreated SAHS can nonlinear parameters such as approximate entropy (ApEnt) [16], also relate to ischemic heart disease, cardiovascular disfunc- central tendency measure (CTM), and Lempel–Ziv complexity tion, and stroke [3], [4]. A common de?nition of apnea involves (LZCom) [17] were also applied to sleep apnea detection. In the a cessation of air?ow for at least 10 s while hypopnea is de- realm of spectral-domain features, Zamarron ´ et al. [18] studied ?ned as a minimum 10-s air?ow reduction with either a blood the periodogram of the SpO signal and selected four indices 2 oxygen desaturation of 4% or a neurological arousal [5]. Cur- related to the period 30–70 s for detection purpose. However, all rently, polysomnography (PSG) is considered as the standard the aforementioned methods rely on the entire overnight SpO 2 method for SAHS diagnosis. Nevertheless, PSG requires SAHS records, resulting in a delayed of?ine analysis and diagnosis. suspects to sleep in a sleep laboratory over one or two nights, In order to obtain a real-time SAHS monitoring and di- with attended technicians. During the overnight sleep, a variety agnosis, some pioneering works have emerged. Oliver and of sensors and wires are attached to the suspect’s body. The Flores-Mangas [19] implemented a real-time detection system recorded signals are then analyzed by sleep specialists for ?nal with oximetry but unfortunately lacked a performance com- diagnosis. The discomfort, inconvenience, and expensiveness of parison with the standard PSG. Heneghan et al. [12] adopted PSG set a barrier from its prevalence among public. Therefore, a both ECG and SpO signals to estimate the apnea plus hypop- 2 nea index (AHI) based on an epoch-by-epoch detection. Most recently, Burgos et al. [20] and Bsoul et al. [11] have imple- Manuscript received April 21, 2011; revised September 23, 2011 and January mented a systematic real-time SAHS detection based on SpO 2 19, 2012; accepted February 9, 2012. Date of publication February 16, 2012; alone and ECG alone, respectively. date of current version May 4, 2012. In this paper, we focus on real-time sleep apnea/hypopnea The authors are with the Department of Electrical Engineering, University of Texas at Dallas, Richardson, TX 75080 USA (e-mail: [email protected]; detection based on 1-min segments of ECG and SpO signals. 2 [email protected]). While following the comprehensive ECG features set in [11], we Digital Object Identi?er 10.1109/TITB.2012.2188299 1089-7771/$31.00 © 2012 IEEE470 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 3, MAY 2012 TABLE I PHYSIOLOGICAL PROPERTIES OF SUBJECTS IN UCD DATABASE including signal processing, feature extraction, feature selec- tion, and classi?er combination. Section III shows the results of different classi?cation strategies. In Section IV, we discuss the performances among different feature sets and classi?cation schemes, and compare our results with those of other research groups. Finally, Section V concludes this paper. II. MATERIALS AND METHODS A. Database Throughout this paper, we use the St. Vincent’s Univer- sity Hospital/University College Dublin Sleep Apnea Database (UCD database) [22] available online from PhysioNet [23] which provides a variety of physiological signals for biomedi- cal research. The availability of the UCD database offers easy validation and assessment of our approach. Twenty-?ve (21 males and 4 females) sleep-disordered-breathing suspects’ full overnight PSG recordings are recorded in the database. Each recording contains 5.9- to 7.7-h ECG and SpO signals as 2 well as an annotation ?le with detailed onset time and dura- tion of every apnea/hypopnea event. Polysomnograms were ob- tained using the Jaeger–Toennies system (Erich Jaeger GmbH, Germany). In particular, the ECG signal was recorded via a Fig. 1. Different strategies of SAHS detection. (a) SAHS detection based modi?ed lead V2 and the SpO signal was recorded using a 2 on the ECG signal only. (b) SAHS detection based on the SpO signal only. 2 ?nger pulse oximeter. General physiological properties of the (c) SAHS detection based on both ECG and SpO signals. (d) SAHS detection 2 basedonECG andSpO signals with feature selection. (e) SAHS detection 2 subjects are listed in Table I while a more detailed description based on selected ECG and SpO features with classi?er combination. 2 can be found online [22]. B. Signal Processing and Feature Extraction develop an SpO feature set of 39 features. The different strate- 2 gies of SAHS detection that we explore can be found in Fig. 1. The ECG and SpO signals in the database are originally 2 To begin with, investigations of individual diagnostic abilities of sampled by 128 and 8 Hz, respectively. In our experiment, both the ECG and SpO features are carried out and compared via an signals are segmented into 1-min episodes for signal processing 2 experiment of ten different machine-learning algorithms. Cost- and the detection/classi?cation results are returned minute by sensitive classi?cation is included to enhance the sensitivity. minute. Accordingly, the annotations need to be modi?ed to give Additionally, the classi?cation experiment extends to involving minute-based references. Given that the apnea/hypopnea hap- a full set of 150 features fused by both ECG and SpO features. pens with a minimum of 10-s air?ow change, in case the events 2 In contrast to feature selection based on individual evaluation are across two adjacent segments, we label a single minute as of each feature as used in existing works [11], [20], [21], we “Apnea” (apneic) if it contains at least 5 consecutive seconds of utilize a subset feature selection strategy that accounts not only apnea/hypopnea events; otherwise, this minute is annotated as the prediction ability of each feature, but also the redundancy “No apnea” (normal). We use MATLAB for signal processing among features. The idea of cost-sensitive weighting is also and feature extraction. incorporated in the feature selection process to favor highly Recently, Bsoul et al. [11] have proposed a quite comprehen- predictive features. Most importantly, with investigation of the sive ECG feature set which contains 111 HRV- and EDR-based performances of individual classi?ers, we introduce classi?er features in both time and spectral domains. In this paper, we combination to further enhance detection performance by har- employ this feature set for the ECG signal and mainly focus on nessing the potentially complementary information provided by feature designs of the SpO signal due to its strong re?ection of 2 individual classi?ers. arterial oxygen saturation on the air?ow ?uctuation. The rest of this paper is organized as follows. In Section II, we To begin with, the SpO signal is downsampled at 1 Hz. Any 2 describe the database and the methods used in our experiment, SpO values less than 50 are considered as artifacts and hence 2XIE AND MINN: REAL-TIME SLEEP APNEA DETECTION BY CLASSIFIER COMBINATION 471 TABLE II removed from analysis (totally about 137 min; 1.3% of data are SPO FEATURES AND THEIR DESCRIPTIONS 2 removed). Basic statistics such as the minimum, mean, variance, and correlation coef?cient of SpO samples within each segment are 2 ?rst calculated and denoted as Smini, Smean, Svari, and CorreC, respectively. Then using Smean as a baseline, the number of zero crossing points of each segment is counted as feature NumZC. Within each segment, via linear regression, a regression line is ?tted. The Slope, its absolute value (AbSlope), and Bias of the regression line are measured as three additional features. Delta index is viewed as a valid parameter for overnight SAHS detection [5], [13], [14]. In our real-time processing, the mean value of the SpO signal over every 12-s interval is ?rst com- 2 puted and the Delta index (Dmean) is derived as the one-min average of the absolute differences between two successive mean values. The nonlinear methods such as ApEnt, CTM, and LZCom, which measure the regularity, variability, and complexity of a time series, have been applied to SAHS detection based on overnight SpO signals [16], [17]. These three features can also 2 be easily calculated segmentwise. In particular, for ApEnt,we choose the optimal run length of 1 and tolerance window of 0.25 times the standard deviation of each epoch data, as suggested in [16]. The CTM is calculated by selecting a radius with respect to (w.r.t.) the origin of a second-order difference plot and counting the number of points which fall within the radius [17]. We choose radii of 0.25, 0.5, 0.75, and 1, corresponding to features CTM25, CTM50, CTM75, and CTM100, respectively. Besides three ODI indices (odi2, odi3, odi4) in [20], we adopt a more general de?nition of ODI indices as in [15]. We set the baseline as the mean of the top 20% of the SpO data within 2 neighbor (kNN) [26] assigns the data the most common class 1 min. The ODI index ODIxy counts the occurrences that SpO 2 among their k closest neighbors. Decision Table classi?er [27] samples drop at least x below the baseline and last at least builds a simple hypothesis space represented by a decision table y s. In our experiment, x?{2,3,4,5} and y?{1,3,5}.For and uses it for classi?cation. Multilayer perceptron (MLP) [28] example, ODI21 presents the number of times when the SpO 2 is an arti?cial neural network consisting of multiple layers of level declines at least 2 below the baseline and lasts at least nodes; it models data pattern during training and follows the 1 s. Additionally, we count the total number of SpO samples 2 pattern to classify the testing data. Decision tree partitions which fall at least 2, 3, 4, and 5 below the baseline, contributing data into different groups recursively. Several popular deci- another four features: ODIS2, ODIS3, ODIS4, and ODIS5. sion trees and their evolved versions are included, such as C4.5 Finally, we consider ?ve TSA indices (tsa95, tsa90, tsa85, tree [29], reduced-error pruning tree (REPTree), and functional tsa80, tsa70) indicating the accumulative time that the SpO 2 trees (FT trees) [30]. We also take into consideration several level stays below 95, 90, 85, 80, and 70, respectively. 1 meta-algorithms which are used in conjunction with other (sim- As a result, totally, a set of 39 SpO features is built. The 2 ple) classi?ers to reduce prediction errors. In particular, Adap- aforementioned SpO features and their descriptions are sum- 2 tive Boosting (AdaBoost) with Decision Stump [31], Bagging marized in Table II for ease of reference. with REPTree [32], and Bagging with Alternating Decision Tree (ADTree) [33] are included in our experiment. C. Classi?cation Throughout the experiment, the default parameter settings In the classi?cation phase, we employ an open-source of each classi?er are kept, except that for SVM; we normalize machine-learning software, WEKA [24], as the major tool to all the features into the [0,1] region since SVM is sensitive to assess the performances of the aforesaid feature sets. Ten clas- the dynamic ranges of the features. And for kNN classi?er, we si?ers are included in our experiments. Speci?cally, support choose k =5 and apply an inverse distance weight function. vector machine (SVM) [25] maps data into a high-dimensional To assess the classi?cation performance, we use sensitivity, space and constructs a hyperplane to separate them. k-nearest speci?city, and accuracy as evaluation metrics. Their de?nitions are as follows: 1 We mainly focus on the time-domain features in our real-time processing TP since the apnea/hypopnea event can last as long as 120 s [19], which exceeds sensitivity = (1) the epoch length. Part of the SpO features is used in our conference paper [21]. 2 TP + FN472 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 3, MAY 2012 TN features with higher sensitivity are preferred. In terms of the speci?city = (2) base subset evaluator, correlation-based feature subset selection TN + FP (CfsSubsetEval) [34] is employed independently of classi?ers. TP + TN accuracy = (3) CfsSubsetEval evaluates the merit of a subset of features by P+N considering the individual predictive ability of each feature as well as the redundancy among them. To be speci?c, the merit where true positive (TP) and true negative (TN) refer to the of a subset S is calculated as [34] number of correctly detected apneic epochs and normal epochs, respectively, whereas false positive (FP) and false negative (FN) kr cf stand for the number of miss-identi?ed apneic and normal M = \u0000 (4) S epochs, respectively. P/N represents the total number of epochs (k + k(k- 1)r ) ? with/without apneic events. In other words, the sensitivity re- ?ects the ability to correctly detect the apneic epochs; speci?city where k is the number of features in S; r is the average class- cf conveys the ability to distinguish the normal epochs. A tradeoff feature correlation while r is the average feature–feature in- ? between the two usually exists in practice. In terms of SAHS tercorrelation. As a result, a subset of features which are highly detection problem, we are more interested in high sensitivity al- correlated with the class (apneic or normal) while having low gorithms which reduce the risk of missing the apnea/hypopnea intercorrelation among the features is preferred. In our experi- events that do pose threats to the patients. Therefore, we also ment, feature selection is performed on the full feature set which carry out cost-sensitive classi?cations by imposing a cost matrix consists of 111 ECG and 39 SpO features. 2 on the above ten classi?ers, to penalize the FN errors more than the FP errors, and aim for a higher sensitivity. E. Classi?er Combination As can be seen in Fig. 1, we ?rst carry out experiments to assess and compare the individual diagnostic abilities of ECG Naturally, one would want satisfactory results in all sensitiv- and SpO features. The 111 ECG features and 39 SpO features 2 2 ity, speci?city, and accuracy. However, this ideal result does not are fed into the ten classi?ers, respectively [see Fig. 1(a) and always happen given a single classi?er. Some classi?ers provide (b)]. In the next experiment, we fuse the two feature sets together high sensitivity but low speci?city while others perform the op- and perform classi?cation with a full set of 150 features [see posite way. The gap between the two metrics can be large. To Fig. 1(c)]. The experiments include both even cost (where the FN strike a balance, classi?er combination is proposed as a potential and FP are evenly weighted) and cost-sensitive classi?cations. solution. Note that all the experiment results are based on tenfold cross Classi?er combination has been proved to be a powerful validation of the entire database. In particular, the entire dataset method to improve the classi?cation performance in many is evenly divided into ten folds; each time, one fold is left out ?elds [35]. Inspired by the fact that the misclassi?ed instances for testing while the other nine folds are used for training. The of individual classi?ers do not necessarily overlap, different testing results are averaged over ten folds and then returned as classi?ers may offer different perspectives in decision making. the ?nal cross-validation results. Making use of those complementary information by classi?er combination could further improve the performance. In our combination experiment, we choose three individual D. Feature Selection classi?ers to form a group of experts. Each member classi?er Feature selection is regarded as a classic method to prevent predicts the class of every epoch and records the probabilities over?tting by eliminating redundant or even detrimental fea- associated with both classes (apneic and normal). Since the tures. In a real-time detection scenario, it is also an effective prior class distribution of a database is usually unaware before- way to reduce the computational load by requiring less sig- hand, an equiprobable prior is a common assumption. Based on nal processing in feature extraction, to lower the rate of data the predicted classes and their corresponding probabilities, we transmission and energy consumption, and more importantly, to explore four classi?er combination schemes as described next. shorten the time required for model building during the training 1) Max Probability (MP): This scheme assigns the predicted process. class as the one with the maximum probability among all clas- Feature selection can be done by either investigating the value si?ers and classes: of each features, as done in [11], [20], and [21], or using a subset evaluator to assess the merit of a group of features. With C = arg max max{P (C+),P (C-)} (5) p i i i the former method, no interinformation among features can be {C +,C-} concluded, whereas the latter strategy can explore the correlation among features. Therefore, we employ the latter subset feature where C is the ?nal predicted class after combination and i is p selection to further reduce redundancy. the classi?er index. P (C+) and P (C-) are the probabilities i i In the interest of high sensitivity, we also use a cost-sensitive that classi?er i predicts the epoch to be positive (apneic) and subset evaluator for feature selection. Cost-sensitive subset eval- negative (normal), respectively. uator is a meta subset evaluator which requires a base subset 2) Average Probability (AP): In this rule, the probabilities evaluator and a cost matrix. In our case, the same cost ma- of the positive/negative class are summed over all classi?ers trix used in cost-sensitive classi?cation is applied, and thus, the (can be viewed as the arithmetic mean criterion). The class withXIE AND MINN: REAL-TIME SLEEP APNEA DETECTION BY CLASSIFIER COMBINATION 473 TABLE III CLASSIFICATION RESULTS BY USING EITHER ECG OR SPO FEATURE SET 2 TABLE IV CLASSIFICATION RESULTS BY USING ECG AND SPO FEATURE SETS TOGETHER 2 larger summed probability is returned as the ?nal prediction: 40%, 90%, and 70% with the ECG feature set, whereas they \u0000 \u0000 are 60%, 90%, and 80% with the SpO feature set, respectively. 2 \u0000 \u0000 To enhance the sensitivity, via testing different penalty weights, C = arg max P (C+), P (C-) . (6) p i i {C +,C-} we choose a cost matrix to penalize the FN three times that i i of the FP (i.e., Cost Sensitive 3), at the expense of speci?city 3) Product of Probability (PP): Similar to the aforemen- but still maintaining an acceptable accuracy (around 70–80%). tioned approach, this scheme chooses the class with the larger Under the Cost Sensitive 3 section of Table III, the sensitivity product of positive/negative probabilities (can be viewed as the of the ECG feature set now ranges from 53.91% to 72.47% and geometric mean criterion): that of the SpO feature set spans over [70.31%, 87.63%].The 2 \u0000 \u0000 \u0000 \u0000 accuracies of using these two feature sets remain around 70% C = arg max P (C+), P (C-) . (7) p i i and 80%, respectively. {C +,C-} i i 2) Using Combined ECG and SpO Feature Set: Table IV 2 provides both the even cost and cost-sensitive classi?cation re- 4) Majority Voting (MV): The last approach chooses the sults based on the combined feature set. Given the Cost Sensitive class to which the majority of the classi?ers agree. Let 3 result in Table IV, among ten classi?ers, AdaBoost with De- C (C+) = 1 if classi?er i predicts the current epoch to be i cision Stump achieves the highest sensitivity of 87.03% but the positive; otherwise, C (C+) = 0 and apply a similar rule to i lowest speci?city (74.82%) and accuracy (77.79%). Bagging C (C-). Then, the ?nal prediction is determined as i \u0000 \u0000 with REPTree enjoys both the highest speci?city (85.89%) and \u0000 \u0000 accuracy (84.40%) but a relatively low sensitivity (79.75%). C = arg max C (C+), C (C-) . (8) p i i 2 {C +,C-} In addition, we record in Table IV the CPU time spent on i i training and testing of each classi?er during the tenfold cross In other words, it compares the numbers of classi?ers which validation. As can be seen, MLP, Bagging with ADTree, and give positive/negative predictions, and picks the class with the SVM are the most computationally intensive classi?ers based larger number of votes. on the CPU time spent on training: from about 33 to 886 s, whereas the remaining classi?ers require no more than 15 s in III. RESULTS training. A. Results of Individual Classi?ers 1) Using Either ECG or SpO Feature Set: The classi?ca- 2 tion results of using either ECG or SpO feature set are ?rst 2 2 The aforementioned results are obtained from a PC of a Linux system with tabulated in Table III. For the case of Even Cost, among all Intel Core 2 Duo CPU E6850 @ 3.00 GHz, 2G RAM. All the classi?ers are classi?ers, the sensitivity, speci?city, and accuracy are around implemented in Java.474 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 3, MAY 2012 TABLE V Since AdaBoost with Decision Stump and Bagging with SELECTED FEATURES OF REDUCED FEATURE SET FS39 REPTree achieve highest sensitivity and speci?city, respectively, we choose them as two permanent member classi?ers in our combination experiment. On the other hand, to avoid too much computation, three most computationally intensive classi?ers: SVM, Bagging with ADTree, and MLP are excluded. As a result, an additional classi?er is drawn from the remaining ?ve classi- ?ers sequentially to form a group of three experts. The reduced feature set FS39 and the SpO feature set are both considered TABLE VI 2 CLASSIFICATION RESULT OF REDUCED FEATURE SET FS39 WITH COST in the classi?er combination experiment. SENSITIVE 3 Via a thorough experiment with ?ve different choices of the third member classi?er, it is found that kNN (k =5) and Deci- sion Table collaborate best with AdaBoost with Decision Stump and Bagging with REPTree. Their combination results are illus- trated in Tables VII and VIII, respectively. As we can see, the sensitivity, speci?city, and accuracy are all around 82% and 81% for the FS39 feature set and the SpO feature set, respectively, 2 for all four combination schemes. IV. DISCUSSION A. Comparison Among Different Feature Sets Recall that in Section III-A1, if we compare the individual B. Feature Selection signals, the proposed SpO feature set achieves a much bet- 2 Applying the feature selection strategy described in ter performance than the ECG feature set in both even cost and Section II-D, 39 out of 150 features are selected as a reduced cost-sensitive cases: averagely, about 25% and 12% advantage in feature set (FS39) which consists of 8 ECG features and 31 sensitivity and 6% and 8% advantage in accuracy for Even Cost SpO features. In particular, the eight ECG features [11] are the 2 and Cost Sensitive 3, respectively. Besides, in the selected fea- second-order serial correlation coef?cient (SCrC2), the third, ture set FS39, the majority existence of SpO features (31 SpO 2 2 fourth, ?fth, and sixth discrete Fourier transform (DFT) points features versus 8 ECG features) also implies the higher diagnos- of RR intervals (rrff3, rrff4, rrff5, rrff6), the spectral variances tic ability of SpO features than the ECG features. The reason 2 of second and ninth levels of decimated wavelet transform of could be explained as follows. Since the SpO signal is a direct 2 EDR series (edrw2, edrw9), and the ?fth DFT point of EDR re?ection of the amount of oxygen that one inhales, it could di- series (edrf5). All the selected 39 features are listed in Table V. rectly capture the oxygen variation due to the pause/reduction of Cost-sensitive classi?cations are carried out to assess the per- breath when sleep apnea/hypopnea happens. On the other hand, formance of the reduced feature set FS39 and the results are the ECG signal records the electrical activity of the heart where shown in Table VI. Compared with Table IV, using FS39, the not only apnea/hypopnea, but also many other metabolic pro- computational load shrinks to around only 1/5 of the one us- cesses as well as, highly possibly, other heart-related disorders ing the full combined feature set, for almost all classi?ers, with are taking place. In this sense, the ECG signal is more compli- less than 1% decrease in sensitivity, speci?city, and accuracy. cated and the most sleep-apnea/hypopnea-relevant information AdaBoost with Decision Stump still achieves the highest sensi- might be buried in other signals. Hence, to unveil and extract tivity of 86.81% while Bagging with REPTree retains the high- the most indicative features from ECG for SAHS detection is est speci?city of 84.62% and the accuracy of 83.26% among very challenging. Consequently, if only one sensor is allowed ten classi?ers. In terms of computational ef?ciency, SVM, Bag- for SAHS detection, oximeter is preferable over ECG sensor. ging with ADTree, and MLP still consume the longest CPU time Moreover, if both ECG and SpO signals are available, us- 2 (11–71 s on training) while other classi?ers only spend less than ing the combined full feature set (see Table IV) can largely 3.7 s on training in 1tenfold cross validation. improve the performance of using ECG features alone (see Table III): about 15%, 6%, and 7% increases in maximum sen- C. Classi?er Combination sitivity, speci?city, and accuracy among ten classi?ers, respec- The result in the previous section (see Table VI) also shows tively. The reduced feature set FS39, obtained by our proposed that although AdaBoost with Decision Stump achieves the high- cost-sensitive subset feature selection, achieves approximately est sensitivity, its speci?city is the lowest among ten classi?ers; the same good classi?cation result but only requires about 1/5 of Bagging with REPTree attains the highest speci?city but its the computational load of the full feature set. Comparing FS39 sensitivity is below 80%. In the interest of a well-rounded clas- and the SpO feature set, from Tables VII and VIII, for every 2 si?cation result, we apply classi?er combination proposed in combination scheme, FS39 shows advantage (about 1%) over Section II-E to balance the performances in both sensitivity and the SpO feature set in sensitivity and accuracy while most of 2 speci?city, and hence the accuracy. the time in speci?city as well. In other words, the incorporationXIE AND MINN: REAL-TIME SLEEP APNEA DETECTION BY CLASSIFIER COMBINATION 475 TABLE VII PERFORMANCES OF A CLASSIFIER COMBINATION WITH FS39 AND SPO FEATURE SETS 2 TABLE VIII PERFORMANCES OF ANOTHER CLASSIFIER COMBINATION WITH FS39 AND SPO FEATURE SETS 2 of ECG features complements SpO features by providing ad- Table as the third member gives comparable classi?cation re- 2 ditional useful information for SAHS detection. Therefore, with sults as shown in Tables VII and VIII. The difference mainly lies access to both signals, such as in hospital, classi?er combination in the computational loads of training and testing. In general, with feature set FS39 serves as a good candidate for real-time the total training/testing time for the combination approach is SAHS detection. However, given a home-based setting with the sum of those for individual classi?ers. Note that kNN is a consideration of the cost of extra sensor, classi?er combination lazy classi?er that requires no time on training; thus, the to- based on the SpO signal alone also suf?ces to offer a decent tal training time for the combination approach with kNN stays 2 performance. around 2.9 s versus 5.3 s for the combination approach with Decision Table. However, since kNN of?oads the complexity to testing phase, the corresponding testing time of the former is B. Comparison Among Different Classi?cation Strategies about 2 s, while the latter requires much less, around 3 ms. As Among the ten individual classi?ers, AdaBoost with Decision a result, the former combination is more suitable for a subject- Stump achieves the highest sensitivity and Bagging with REP- dependent (SD) application where testing is conducted based on Tree attains the highest speci?city and accuracy almost for all the model trained by the same subject’s historical data. In this case, trainings need to be carried out or updated on a subject feature sets. basis. The latter combination with Decision Table ?ts better for With classi?er combination (see Tables VII and VIII), almost a subject-independent (SI) application where training is already all three metrics improve if compared with the results of indi- done based on some (large) database before distributing to users vidual classi?ers for both FS39 and SpO feature sets (Tables 2 who are only responsible for testing. VI and III): the sensitivity increases at least 2% compared with Regarding the memory requirement of our proposed classi- Bagging with REPTree while about 7–8% improvement in speci- ?er combination for real-time SAHS detection, both training ?city and 4–5% increase in accuracy compared with AdaBoost and testing programs run freely within the default maximum with Decision Stump. Conclusively, better and more balanced allocated memory (455 MB) which is less than the memory size detection results are achieved with the proposed classi?er com- of iPhone 4 (512 MB RAM). The memory usage for the testing bination. Among four different classi?er combination schemes, phase of our proposed algorithm is much less since each time Majority Voting always wins the highest sensitivity. Because only 1 min of data are processed. In addition, for the training each member classi?er is already in favor of positive class process which generally requires more memory, current tech- due to the cost-sensitive setting, the ?nal prediction of Ma- nology allows large amount of computations to be of?oaded jority Voting further emphasizes the sensitivity by its “hard de- to servers via Internet (cloud computing) [11]. As a result, the cision” nature: based on the classes instead of the associated memory issues can be effectively solved. probabilities. Finally, it is interesting to note that performances of algo- Given a real-time detection problem, the computational com- rithms such as SVM and MLP also depend on parameter set- plexity is of concern. AdaBoost with Decision Stump and Bag- tings. However, it is usually unknown beforehand which pa- ging with REPTree are computationally ef?cient compared to rameter setting is the best for a given problem. Locating the other algorithms such as SVM and MLP, in addition to their good optimal parameter setting is often done by cross validation and detection performances. With the two being permanent mem- grid search [36]. It is also highly probable that the best param- eter setting for one database fails to cater for another database. bers for classi?er combination, choosing either kNN or Decision"