AN IMPROVED APPROACH FOR REAL-TIME DETECTION

OF SLEEP APNEA

Baile Xie, Wenxun Qiu, Hlaing Minn, Lakshman Tamil and Mehrdad Nourani

Department of Electrical Engineering EC33, University of Texas at Dallas

800 W. Campbell Road, TX 75080, Richardson, U.S.A.

Keywords:

Sleep anpea, SpO

, Real-time detection, Feature selection, Cost-sensitive.

Abstract:

The traditional diagnosis of sleep apnea and hypopnea syndrome (SAHS) requires an expensive and complex

overnight procedure called polysomnography (PSG). Recently, ﬁnding valid alternatives for SAHS diagno-

sis has attracted much research attention. This paper focuses on the real-time monitoring and detection of

SAHS based on the arterial oxygen saturation signal measured by pulse oximetry (SpO

). We develop a more

comprehensive feature set and a more appropriate annotation criterion, if compared to the existing approaches

in the literature. To enjoy competitiveness in computational complexity, we also propose a reduced feature

set which provides a higher sensitivity and better adaptivity to distinct databases. The performances of 15

commonly used classiﬁers with different cost matrixes are assessed on different databases, offering detailed

insights on the diagnostic abilities of these methods.

1 INTRODUCTION

Sleep apnea and hypopnea syndrome (SAHS) is a

common sleep disorder which is estimated to affect

2% of middle-aged women and 4% of middle-aged

men (Young et al., 1993). The impacts of SAHS

include daytime sleepiness, fatigue, trafﬁc accidents

and depression. It is also blamed for linkage to is-

chemic heart disease, cardiovascular disfunction and

stroke. The clinical deﬁnition of apnea involves a

cessation of airﬂow for at least 10 seconds while hy-

popnea is deﬁned as a minimum 10-second airﬂow

reduction with either a blood oxygen desaturation of

4% or a neurological arousal (Magalang et al., 2003).

Currently, polysomnography (PSG) is regarded as the

golden standard for SAHS diagnosis. However, PSG

requires patients to sleep overnight in a sleep labora-

tory with attended technicians. A variety of recorded

signals are then analyzed by sleep specialists for ﬁnal

diagnosis. The time- and cost-consuming natures of

PSG limit its prevalence among public, which makes

a readily available, relatively inexpensive and reliable

diagnosis alternative much desirable. Existing SAHS

detection techniques have been developed based on

questionnaires (Netzer et al., 1999), ECG (McNames

and Fraser, 2000,Shinar et al., 2000, Heneghan et al.,

2008), snoring (Ng et al., 2006) and pulse oximetry

(Magalang et al., 2003,L´evy et al., 1996,Olson et al.,

1999, Zamarr´on et al., 2003, Alvarez et al., 2006,

Oliver and Flores-Mangas, 2006, Heneghan et al.,

2008,Linet al., 2008,Burgoset al., 2010), either alone

or in combination. Due to the strong reﬂection of ar-

terial oxygen saturation on the breathing airﬂow ﬂuc-

tuation and the convenience and availability of pulse

oximetry, we focus on SpO

signal in this paper for

SAHS detection.

Previous studies have proposed many quantitative

indexes derived from SpO

signal for SAHS detec-

tion. Among the commonly used time-domain in-

dexes are the accumulative time spent below a cer-

tain saturation level (Magalang et al., 2003, Olson

et al., 1999), the oxygen desaturation index (ODI,

the number of oxyhemoglobin desaturation below a

certain threshold) (Lin et al., 2008), and the satura-

tion variability index (Delta index) (Magalang et al.,

2003, L´evy et al., 1996, Olson et al., 1999). Besides,

Zamarrn et al. (2003) exploited the periodogram of

SpO

signal and discovered that the period 30s to 70s

is the interval of interest (P

30−70

). The four indexes

are related to the periodogram as the total area under

periodogram, the area enclosed in the periodogram

within P

30−70

, the area ratio of that within P

30−70

with

respect to the total periodogram area, and the peak

amplitude of the periodogram in P

30−70

, respectively.

169

Xie B., Qiu W., Minn H., Tamil L. and Nourani M..

AN IMPROVED APPROACH FOR REAL-TIME DETECTION OF SLEEP APNEA.

DOI: 10.5220/0003137101690175

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2011), pages 169-175

ISBN: 978-989-8425-35-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Later on, several non-linear parameters such as ap-

proximate entropy (ApEn), central tendency measure

(CTM) and Lempel-Ziv complexity (LZC) are also

derived from the SpO

signal as the indexes for SAHS

detection (Alvarez et al., 2006).

However, all the methods mentioned above per-

form in the context of the overnight SpO

records,

rendering a delayed off-line analysis and diagnosis.

Recently, the idea of real-time SAHS monitoring and

diagnosis is proposed as a promising alternative of

PSG. The work in (Oliver and Flores-Mangas, 2006)

introduces the real-time implementation of SAHS de-

tection but lacks of a performance comparison with

the standard PSG detection. Heneghan et al. (2008)

adopt the ECG and SpO

signals jointly to estimate

the apnea plus hypopnea index (AHI) on an epoch

basis. Most recently, Burgos et al. (2010) imple-

ment a systematic real-time SAHS detection based on

the Apnea-ECG database (CinC, 2000) available on-

line from PhysioNet (Goldberger et al., 2000), attain-

ing a classiﬁcation accuracy of 93.03%, sensitivity

of 92.35% and speciﬁcity of 93.52%, upon speciﬁed

training and testing sets. Unfortunately, this database

contains only 8 recordings with SpO

signal. The lim-

ited sample number casts uncertainty on the general

applicability and robustness of this approach.

In this paper, we ﬁrst implemented the method in

(Burgos et al., 2010) (labeled as RT for short) on an-

other database St. Vincent’sUniversity Hospital / Uni-

versity College Dublin Sleep Apnea Database (UCD

Database) (UCD, 2000) which can also be found on

PhysioNet. Even with a weighted cost matrix, though

RT method gets a speciﬁcity of 96.12% and accu-

racy of 89.92%, the sensitivity drops dramatically to

33.63%, which is far from satisfactory. For the pur-

pose of SAHS detection, we would rather misclassify

a healthy person as SAHS positive, than let an SAHS

patient go unidentiﬁed. High sensitivity is preferable

over high speciﬁcity in this case. With this recogni-

tion, our paper offers contributions in the following

aspects: (1) Conversion of most of the existing in-

dexes into epoch-based (1-minute based) features. (2)

Forming a more comprehensive feature set of SpO

signal with higher sensitivity. (3) Proposal of a more

appropriate criterion of segment annotations. (4) Pro-

posal of a reduced feature set with better diagnostic

ability and computational efﬁciency. (5) Validation

of the performance of the proposed approach on two

distinct databases. (6) The performance assessment

of 15 classiﬁers with different cost-sensitivities upon

two databases.

The rest of the paper is organized as follows: In

Section 2, we introduce the two databases used and

explain the new approach in feature extraction. Sec-

tion 3 describes the experiments and discusses the re-

sults. Finally, Section 4 concludes this paper.

2 NEW INVESTIGATIONS

2.1 Database Description

PhysioNet provides a variety of physiological signals

for biomedical research. Both databases we used are

available from the web site, which offers easy valida-

tion and assessment of our approach.

• Apnea-ECG Database. This database contains

8 recordings with SpO

signals. Associated with

each signal is a reference annotation ﬁle cre-

ated by a sleep expert based on simultaneously

recorded respiration and oxygen saturation sig-

nals. The annotation is given on a 1-minute basis.

Each minute is labeled as ‘A’ when apnea was in

progress at the beginningof the associated minute,

otherwise this minute is label as ‘N’. We name this

annotation deﬁnition as AN for short. To make use

of this kind of annotation, the real-time monitor-

ing system is designed to give the detection result

minute by minute.

• UCD Database. This database comprises of 25

full overnight PSG recordings, each of which con-

tains an SpO

signal. The annotations are pre-

pared by sleep technologists who detailed the on-

set time and duration of every apnea and hypop-

nea event. In order to deﬁne the reference annota-

tion on a 1-minute basis, two labeling criteria are

used. The ﬁrst one applies the same technique in

Apnea-ECG database. Considering that the ap-

nea and hypopnea associate with a minimum of

10 second airﬂow change, in case the events are

across two adjacent segments, the second crite-

rion marks a single minute as ‘Apnea’ if it con-

tains at least 5 consecutive seconds of apnea and

hypopnea events, otherwise this minute is labeled

as ‘No apnea’. This criterion is termed as AH5C

in the following. Note that the same annotation is

also used in (Heneghan et al., 2008) except in an

overlapped epochs scenario.

2.2 Signal Processing

The SpO

signals from both databases are downsam-

pled at 1 Hz and the outliers lie in [0, 50%] are re-

moved to avoid outﬁtting. In order to inherit the mer-

its of existing metrics of the SpO

, we devise to mod-

ify the indexes and incorporate them in the real-time

detection method. To begin with, the SpO

signals

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

170

are segmented into 1-minute epochs. Then, the ex-

isting indexes are computed for each 1-minute epoch.

In particular, the ODI indexes, apart from the ones

in (Burgos et al., 2010), set the baseline as the mean

of the top 20% of the SpO

data within one minute,

and then sum up the number of samples which fall be-

low it. As a result, the features ODI2, ODI3, ODI4,

and ODI5 represent the ODI indexes corresponding

to 2%, 3%, 4%, and 5% below the baseline, respec-

tively. Delta index is viewed as a valid parameter for

overnight SAHS detection. To translate it into our

real-time processing, the minimal SpO

value in ev-

ery 12-second interval is picked and the Delta index

is derived as the sum of the absolute differences be-

tween two successive dips, dividing by the number of

intervals, i.e., 5 in one minute. The nonlinear meth-

ods such as ApEn, CTM, and LZC can also be easily

applied segment-wise. Speciﬁcally, we choose radii

of 0.25, 0.5, 0.75 and 1 for CTM corresponding to

CTM25, CTM50, CTM75, CTM100 features, respec-

tively.

Since the apnea/hypopnea event can last as long

as 120 seconds (Oliver and Flores-Mangas, 2006),

which exceeds the epoch length, we rule out the

frequency-domain indexes in our real-time process-

ing and focus on the ones derived directly from the

time-domain recordings.

Combined with the eight features used in (Burgos

et al., 2010), a more comprehensive feature set (la-

beled as ALL) is formed containing 19 features in all.

Classiﬁcation experiments and further feature selec-

tion are carried out based on this feature set in the

following section.

3 EXPERIMENT AND RESULT

DISCUSSION

We use WEKA (Hall et al., 2009), an open-source

machine learning software as the major tool to as-

sess the performances of 15 classic classiﬁcation al-

gorithms with their default parameter setting. Be-

sides the Bagging with ADTree suggested in (Burgos

et al., 2010), Bagging with REPTree, Support Vector

Machine (SVM), Naive Bayes, Multilayer Perceptron

(MLP), Radial Basis Function Network (RBFNet-

work), Decision Stump, J48 (C4.5) tree and so on are

all tested to ﬁnd out the most appropriate candidates

for real-time SAHS detection. All the classiﬁcation

performances, namely, the sensitivity, speciﬁcity and

accuracy are based on ten repetitions of 10-fold cross

validation for a more accurate evaluation.

Table 1: Performance of RT and ALL feature sets using

Bagging with ADTree with AN annotation.

Apnea-ECG database UCD database

RT All RT All

Sensitivity(%) 96.08 96.95 33.63 43.07

Speciﬁcity(%) 93.85 93.53 96.12 94.39

Accuracy(%) 94.88 95.11 89.92 89.30

3.1 Comparison between Two

Databases

To begin with, we take a look at the performancecom-

parison between the two feature sets, RT and ALL,

using the Bagging with ADTree algorithm recom-

mended by Burgos et al. (2010). The annotation cri-

terion of Apnea-ECG database, i.e. AN, is applied to

UCD database as well. Table 1 lists the results indi-

cating that the ALL set achieves a slightly better per-

formance than the RT set in Apnea-ECG database.

On the other hand, for the UCD database, the sen-

sitivity of the ALL set increases about 10% over that

of the RT set, but a sensitivity of 43.07% is still not

acceptable for practical detection purpose.

3.2 Comparison between Two

Annotation Criteria

The second experimentis conducted using the two an-

notation criteria: AN and AH5C on UCD database.

The classiﬁcation results of 15 classiﬁers are recorded

in Table 2 and 3, respectively. Comparing the two ta-

bles, it is observed that the AH5C gains an obvious

advantage in sensitivity over the AN for both feature

sets among all classiﬁers. In contrast to AN, the AH5C

annotation scheme is not only more physiologically

justiﬁable, but also more sensitive to those SpO

fea-

tures. Therefore, we choose this annotation criterion

for UCD database in the following.

3.3 Comparison between Two Feature

Sets ALL and RT

The results in Table 3 show that, for each classiﬁer,

using the ALL feature set returns a higher sensitivity

than the RT feature set.

To further enhance the detection sensitivity, cost

matrixes can be used to suppress the false negative er-

rors. Two cost matrixes, which penalize the false neg-

atives twice (Cost Sensitive (2)) and ﬁve times (Cost

Sensitive (5)) as the false positives respectively, are

adopted in cost-sensitive classiﬁcation experiments.

The gray area and white area of Table 4 present the

AN IMPROVED APPROACH FOR REAL-TIME DETECTION OF SLEEP APNEA

171

Table 2: Performance of RT and ALL feature sets using

different classiﬁers with AN annotation for UCD database.

Sensitivity Speciﬁcity Accuracy

Classiﬁer RT ALL RT ALL RT ALL

SVM 0.00 0.16 1.00 0.98 0.90 0.90

RandomTree 0.21 0.31 0.96 0.92 0.89 0.86

J48 trees 0.17 0.22 0.98 0.97 0.90 0.89

NaiveBayes 0.32 0.63 0.96 0.87 0.89 0.85

Bagging.REPTree 0.18 0.22 0.98 0.98 0.90 0.90

Bagging.ADTree 0.18 0.07 0.98 0.99 0.90 0.90

MLP 0.23 0.27 0.98 0.97 0.90 0.90

FT trees 0.17 0.26 0.98 0.96 0.90 0.89

RandomForest 0.19 0.18 0.97 0.98 0.89 0.90

RBFNetwork 0.12 0.03 0.99 1.00 0.90 0.90

Decorate trees.J48 0.17 0.24 0.98 0.95 0.90 0.88

ADTree 0.24 0.10 0.97 0.99 0.90 0.90

REPTree 0.15 0.18 0.98 0.98 0.90 0.90

DecisionStump 0.00 0.00 1.00 1.00 0.90 0.90

SimpleCart 0.20 0.26 0.97 0.96 0.89 0.89

Table 3: Performance of RT and ALL feature sets using dif-

ferent classiﬁers with AH5C annotation for UCD database.

Sensitivity Speciﬁcity Accuracy

Classiﬁer RT ALL RT ALL RT ALL

SVM 0.16 0.56 0.99 0.92 0.78 0.83

RandomTree 0.43 0.58 0.94 0.85 0.81 0.78

J48 trees 0.49 0.57 0.94 0.92 0.82 0.83

NaiveBayes 0.42 0.66 0.95 0.90 0.81 0.84

Bagging.REPTree 0.50 0.62 0.94 0.92 0.83 0.84

Bagging.ADTree 0.53 0.59 0.92 0.93 0.82 0.84

MLP 0.49 0.61 0.94 0.92 0.82 0.84

FT trees 0.47 0.57 0.94 0.90 0.82 0.82

RandomForest 0.46 0.55 0.94 0.92 0.81 0.83

RBFNetwork 0.45 0.53 0.93 0.94 0.81 0.83

Decorate trees.J48 0.48 0.57 0.94 0.89 0.82 0.81

ADTree 0.52 0.57 0.93 0.93 0.82 0.84

REPTree 0.49 0.60 0.94 0.91 0.82 0.83

DecisionStump 0.58 0.81 0.87 0.79 0.80 0.80

SimpleCart 0.50 0.58 0.93 0.89 0.82 0.81

experiment results of Cost Sensitive (2) and Cost Sen-

sitive (5), respectively, while the results in Table 3

correspond to the even cost. It is veriﬁed that sen-

sitivity improves as the penalties of the false nega-

tives are added. However, the speciﬁcity is compro-

mised as sensitivity goes higher. A trade-off exists

between them. The overall accuracy also depends on

the proportion of the apnea/hypopnea minutes in one

recording. Say, if a severe SAHS patient with a great

proportion of apnea/hypopnea event undergoes in the

test, the high sensitivity schemes lead to a high ac-

curacy, and vice versa. Using the ALL feature set,

among the 15 classiﬁers, the Decision Stump and the

RBFNetwork seem to be the best candidates which

have balanced sensitivity and accuracy around 80%

under Cost Sensitive (2). In the case of Cost Sensi-

tive (5), the SVM, J48 tree, Bagging with REPTree,

Bagging with ADTree, MLP, RBFNetwork, ADTree,

Decision Stump all obtain sensitivity higher than 83%

and accuracy higher than 75%.

3.4 Feature Selection

Previous experiments demonstrate the advantages of

the ALL set over the RT set in sensitivity; never-

theless, the ALL set incorporates the features in the

RT set, potentiating a more complicated and time-

consuming classiﬁcation process, which may under-

mine the superiority of real-time monitoring. To im-

prove the computational efﬁciency, we evaluate the

Information Gain of each feature and the top three are

selected to form a 3-feature set (S3), which consists

of Delta index, ODI3, and CTM50. The performance

of S3 set will be assessed and compared below.

3.5 Comparison between the Reduced

Feature Set S3 and RT

To offer a more well-rounded assessment of the two

feature sets as well as different algorithms, the CPU

time (in seconds) spent for training and testing during

the 10-fold cross validations are also included. Note

that even with a smaller feature number, 3, the S3 set

obtains a higher sensitivity and a comparable or better

overall accuracy than the RT set of 8 features, as can

be seen in Table 5. In terms of computational com-

plexity, for most of the classiﬁers, using the S3 feature

set reduces the CPU time sometimes more than one

half of that using the RT set. However, the SVM clas-

siﬁer appears to be an exception. The reason of this

exception might be explained as below. The compu-

tational complexity of SVM depends on the number

of the support vectors (N

). For some speciﬁc algo-

rithms, such as Bunch-Kaufman training algorithm,

the complexity ranges from O(N

+ LN

+ dLN

) to

O(dL

) (Burges, 1998), where d is the number of di-

mensions, L is number of training sequences. In this

case, the S3 set may generate more support vectors

than the RT set does, resulting in the increase of the

complexity, but also provides a higher sensitivity and

better accuracy.

Additionally, the performances of the RT and S3

feature set based on the Apnea-ECG database are also

investigated. As shown in Table 6, the S3 set achieves

the same, if not better classiﬁcation result than the RT

set, even if the AN annotation is used in this database.

Note that the superiority of the S3 in computational

efﬁciency is also well established here. This result

lends evidence to the applicability and high diagnostic

ability of the S3 feature set.

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

172

Table 4: Performance of RT and ALL feature sets using cost sensitive different classiﬁers with AH5C annotation for UCD

database; gray area corresponds to Cost Sensitive (2), and white area corresponds to Cost Sensitive (5).

Classiﬁer Sensitivity(RT) Sensitivity(All) Speciﬁcity(RT) Speciﬁcity(All) Accuracy(RT) Accuracy(All)

SVM 0.52 0.69 0.72 0.87 0.88 0.75 0.85 0.71 0.79 0.73 0.81 0.75

RandomTree 0.55 0.71 0.57 0.56 0.88 0.75 0.85 0.86 0.79 0.74 0.78 0.78

J48 trees 0.61 0.83 0.69 0.83 0.88 0.69 0.85 0.72 0.81 0.72 0.81 0.75

NaiveBayes 0.42 0.43 0.66 0.68 0.95 0.95 0.90 0.89 0.81 0.81 0.84 0.84

Bagging.REPTree 0.64 0.81 0.73 0.84 0.86 0.70 0.86 0.76 0.80 0.73 0.82 0.78

Bagging.ADTree 0.60 0.82 0.76 0.88 0.89 0.71 0.84 0.71 0.82 0.73 0.82 0.75

MLP 0.62 0.81 0.74 0.87 0.87 0.68 0.86 0.72 0.80 0.71 0.82 0.76

FT trees 0.63 0.82 0.68 0.80 0.87 0.70 0.84 0.74 0.81 0.73 0.80 0.75

RandomForest 0.59 0.76 0.64 0.72 0.87 0.74 0.89 0.84 0.80 0.74 0.83 0.81

RBFNetwork 0.50 0.68 0.80 0.88 0.91 0.77 0.79 0.71 0.81 0.75 0.79 0.75

Decorate trees.J48 0.61 0.82 0.65 0.69 0.88 0.69 0.85 0.81 0.81 0.72 0.79 0.78

ADTree 0.59 0.82 0.74 0.89 0.90 0.70 0.85 0.70 0.82 0.73 0.82 0.75

REPTree 0.63 0.82 0.73 0.88 0.86 0.69 0.84 0.69 0.80 0.73 0.82 0.74

DecisionStump 0.58 0.58 0.81 0.84 0.87 0.87 0.79 0.75 0.80 0.80 0.80 0.77

SimpleCart 0.63 0.81 0.68 0.79 0.86 0.71 0.82 0.73 0.80 0.73 0.79 0.75

Table 5: Performance of RT and S3 feature sets using different classiﬁers with AH5C annotation for UCD database.

Sensitivity Speciﬁcity Accuracy CPU Time Training CPU Time Testing

Classiﬁer RT S3 RT S3 RT S3 RT S3 RT S3

SVM 0.16 0.59 0.99 0.91 0.78 0.83 4.8295 7.2387 0.4396 0.4066

RandomTree 0.43 0.57 0.94 0.86 0.81 0.79 0.0577 0.0750 0.0005 0.0007

J48 0.49 0.60 0.94 0.92 0.82 0.84 0.0918 0.0489 0.0005 0.0008

NaiveBayes 0.42 0.65 0.95 0.90 0.81 0.84 0.0154 0.0077 0.0049 0.0032

Bagging.REPTree 0.50 0.60 0.94 0.91 0.83 0.84 0.4310 0.3103 0.0013 0.0022

Bagging.ADTree 0.53 0.58 0.92 0.93 0.82 0.84 4.4422 1.7258 0.0076 0.0073

MLP 0.49 0.57 0.94 0.93 0.82 0.84 11.9256 4.5718 0.0022 0.0011

FT trees 0.47 0.59 0.94 0.92 0.82 0.84 1.0506 0.8379 0.1581 0.0404

RandomForest 0.46 0.55 0.94 0.89 0.81 0.80 0.5621 0.6591 0.0033 0.0038

RBFNetwork 0.45 0.57 0.93 0.93 0.81 0.84 0.3250 0.2916 0.0056 0.0041

Decorate trees.J48 0.48 0.61 0.94 0.91 0.82 0.84 3.7413 1.9616 0.0019 0.0009

ADTree 0.52 0.58 0.93 0.93 0.82 0.84 0.4673 0.1850 0.0006 0.0011

REPTree 0.49 0.60 0.94 0.92 0.82 0.84 0.0413 0.0299 0.0002 0.0007

DecisionStump 0.58 0.81 0.87 0.79 0.80 0.80 0.0137 0.0068 0.0008 0.0005

SimpleCart 0.50 0.57 0.93 0.90 0.82 0.82 0.7726 0.9300 0.0013 0.0009

Since we are more interested in the sensitivity and

the overall accuracy, and usually the training time

plays a major role in overall classiﬁcation time con-

sumption, we omit the speciﬁcity and CPU time for

testing in the following tables of the cost-sensitive

results to save space. It is observed that apply-

ing the cost matrix improves the sensitivity without

big change in computational complexity. Evaluating

the sensitivity, accuracy and complexity all together,

within the scope of UCD database, under the Cost

Sensitive (2), the Decision Stump and RBFNetwork

with the S3 set are good options both with 81% sen-

sitivity, 80% accuracy and little CPU time for train-

ing, as can be seen in Table 7. In the Cost Sensitive

(5) case, the Decision Stump, RERTree, J48, ADTree,

RBFNetwork, and Bagging with REPTree are all nice

choices if the S3 set is adopted. According to Table

8, for Apnea-ECG database, maybe due to the size

of the records and statistical properties of the data,

all classiﬁers work generally well in terms of accu-

racy and sensitivity. We can then choose the classiﬁer

based on the UserCPU time accordingly.

4 CONCLUSIONS

This paper provides improvements to the existing

methods of real-time SpO

signal monitoring and

SAHS detection in terms of a more comprehensive

feature set and a more appropriate segment annotation

criterion with a higher classiﬁcation sensitivity. Fur-

thermore, a feature selection technique is employed

AN IMPROVED APPROACH FOR REAL-TIME DETECTION OF SLEEP APNEA

173

Table 6: Performance of RT and S3 feature sets using different classiﬁers with AN annotation for Apnea-ECG database.

Sensitivity Speciﬁcity Accuracy CPU Time Training CPU Time Testing

Classiﬁer RT S3 RT S3 RT S3 RT S3 RT S3

SVM 0.94 0.94 0.94 0.94 0.94 0.94 0.3670 0.5009 0.0365 0.0345

RandomTree 0.92 0.90 0.94 0.94 0.93 0.92 0.0171 0.0128 0.0002 0.0003

J48 trees 0.95 0.95 0.95 0.94 0.95 0.95 0.0307 0.0111 0.0003 0.0003

NaiveBayes 0.90 0.96 0.93 0.91 0.91 0.94 0.0054 0.0025 0.0015 0.0007

Bagging.REPTree 0.95 0.95 0.95 0.94 0.95 0.94 0.1204 0.0591 0.0002 0.0006

Bagging.ADTree 0.95 0.95 0.95 0.94 0.95 0.94 1.3038 0.5981 0.0035 0.0019

MLP 0.95 0.96 0.95 0.93 0.95 0.94 4.0764 1.5718 0.0005 0.0005

FT trees 0.94 0.94 0.95 0.95 0.95 0.95 0.3584 0.223 0.0346 0.013

RandomForest 0.94 0.92 0.95 0.95 0.95 0.93 0.1486 0.1181 0.0008 0.0008

RBFNetwork 0.93 0.94 0.92 0.93 0.93 0.94 0.1303 0.0981 0.0022 0.0013

Decorate trees.J48 0.92 0.94 0.95 0.94 0.94 0.94 1.2263 0.5012 0.0008 0.0004

ADTree 0.94 0.94 0.94 0.95 0.95 0.94 0.1378 0.0606 0.0002 0.0002

REPTree 0.95 0.95 0.95 0.94 0.95 0.94 0.0126 0.0061 0.0002 0.0003

DecisionStump 0.97 0.94 0.89 0.90 0.93 0.92 0.0055 0.0018 0.0002 0.0003

SimpleCart 0.94 0.94 0.95 0.94 0.95 0.94 0.1989 0.1434 0.0002 0.0002

Table 7: Performance of RT and S3 feature sets using cost sensitive different classiﬁers with AH5C annotation for UCD

database; gray area corresponds to Cost Sensitive (2), and white area corresponds to Cost Sensitive (5).

Classiﬁer Sensitivity(RT) Sensitivity(S3) Accuracy(RT) Accuracy(S3) CPUT. Training(RT) CPUT. Training(S3)

SVM 0.52 0.69 0.75 0.87 0.79 0.73 0.84 0.76 5.7517 7.0192 9.0884 9.5365

RandomTree 0.55 0.71 0.56 0.56 0.79 0.74 0.78 0.77 0.0575 0.0566 0.0755 0.0759

J48 trees 0.61 0.83 0.72 0.87 0.81 0.72 0.82 0.75 0.0917 0.0842 0.0590 0.0589

NaiveBayes 0.42 0.43 0.69 0.73 0.81 0.81 0.84 0.83 0.0159 0.0156 0.0081 0.0088

Bagging.REPTree 0.64 0.81 0.73 0.84 0.80 0.73 0.82 0.77 0.4396 0.4297 0.3321 0.3107

Bagging.ADTree 0.60 0.82 0.73 0.88 0.82 0.73 0.82 0.75 4.5216 4.4795 1.7380 1.7005

MLP 0.62 0.81 0.73 0.88 0.80 0.71 0.82 0.75 11.9846 11.9186 4.5610 4.5643

FT trees 0.63 0.82 0.72 0.88 0.81 0.73 0.83 0.75 1.0545 1.1292 0.9349 0.9794

RandomForest 0.59 0.76 0.63 0.71 0.80 0.74 0.79 0.77 0.5643 0.5560 0.6817 0.6603

RBFNetwork 0.50 0.68 0.81 0.86 0.81 0.75 0.80 0.77 0.3382 0.3401 0.3080 0.2930

Decorate trees.J48 0.61 0.82 0.72 0.84 0.81 0.72 0.82 0.76 4.3842 4.4547 1.9924 3.1778

ADTree 0.59 0.82 0.72 0.88 0.82 0.73 0.82 0.74 0.4611 0.4672 0.1817 0.1832

REPTree 0.63 0.82 0.73 0.87 0.80 0.73 0.82 0.75 0.0378 0.0370 0.0285 0.0288

DecisionStump 0.58 0.58 0.81 0.84 0.80 0.80 0.80 0.77 0.0128 0.0148 0.0065 0.0078

SimpleCart 0.63 0.81 0.68 0.80 0.80 0.73 0.79 0.74 0.7533 0.6710 1.0750 0.9569

Table 8: Performance of RT and S3 feature sets using cost sensitive different classiﬁers with AN annotation for Apnea-ECG

database; gray area corresponds to Cost Sensitive (2), and white area corresponds to Cost Sensitive (5).

Classiﬁer Sensitivity(RT) Sensitivity(S3) Accuracy(RT) Accuracy(S3) CPUT. Training(RT) CPUT. Training(S3)

SVM 0.96 0.97 0.95 0.96 0.92 0.93 0.94 92.99 0.3719 0.3266 0.5186 0.4751

RandomTree 0.92 0.92 0.92 0.91 0.93 0.92 0.92 0.92 0.0158 0.0161 0.0128 0.0128

J48 trees 0.96 0.97 0.97 0.98 0.94 0.94 0.94 0.92 0.0273 0.0267 0.0111 0.0118

NaiveBayes 0.90 0.91 0.96 0.97 0.91 0.91 0.93 0.93 0.0057 0.0057 0.0025 0.0026

Bagging.REPTree 0.96 0.98 0.97 0.98 0.95 0.94 0.94 0.93 0.1154 0.1023 0.0591 0.0533

Bagging.ADTree 0.96 0.98 0.97 0.98 0.95 0.92 0.94 0.93 1.3119 1.2822 0.5981 0.6020

MLP 0.96 0.98 0.97 0.98 0.94 0.93 0.94 0.93 4.1073 4.0772 1.5718 1.5707

FT trees 0.96 0.97 0.97 0.98 0.94 0.93 0.94 0.93 0.3500 0.3267 0.2230 0.2086

RandomForest 0.95 0.97 0.95 0.97 0.95 0.94 0.93 0.93 0.1428 0.1257 0.1159 0.1068

RBFNetwork 0.95 0.96 0.96 0.98 0.93 0.92 0.93 0.91 0.1358 0.1394 0.0994 0.1010

Decorate trees.J48 0.95 0.96 0.96 0.97 0.94 0.93 0.93 0.92 1.2735 1.3409 0.5895 0.7642

ADTree 0.96 0.98 0.96 0.98 0.95 0.92 0.94 0.92 0.1378 0.1373 0.0598 0.0603

REPTree 0.96 0.98 0.97 0.98 0.94 0.93 0.94 0.93 0.0121 0.0118 0.0058 0.0059

DecisionStump 0.97 0.98 0.96 0.96 0.93 0.93 0.92 0.91 0.0047 0.0043 0.0018 0.0019

SimpleCart 0.96 0.97 0.97 0.98 0.94 0.93 0.94 0.93 0.1955 0.1723 0.1462 0.1364

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

174

to ﬁnd out a reduced feature set which only com-

prises of 3 indexes, namely, the Delta index, ODI3

and the CTM50. The reduced feature set not only

lowers the computational complexity, but also enjoys

a better diagnostic ability than the existing feature

sets. Moreover, cost sensitive classiﬁcations are car-

ried out among 15 popular classiﬁers based on two

distinct databases, which substantiate the effective-

ness and robustness of the proposed reduced feature

set and provide guidelines of classiﬁer selections with

the associated real-time detection strategies.

REFERENCES

Alvarez, D., Hornero, R., Ab´asolo, D., Campo, F., and Za-

marr´on, C. (2006). Nonlinear characteristics of blood

oxygen saturation from nocturnal oximetry for ob-

structive sleep apnoea detection. Physiological Mea-

surement, 27:399.

Burges, C. (1998). A tutorial on support vector machines

for pattern recognition. Data mining and knowledge

discovery, 2(2):121–167.

Burgos, A., Goni, A., Illarramendi, A., and Bermudez, J.

(2010). Real-time detection of apneas on a PDA. In-

formation Technology in Biomedicine, IEEE Transac-

tions on, 14(4):995–1002.

CinC (2000). CinC Challenge 2000 data

sets: Data for development and eval-

uation of ECG-based apnea detectors.

http://www.physionet.org/physiobank/database/apnea-

ecg/.

Goldberger, A. L., Amaral, L. A. N., Glass, L., Haus-

dorff, J. M., Ivanov, P. C., Mark, R. G., Mietus,

J. E., Moody, G. B., Peng, C.-K., and Stanley, H. E.

(2000). PhysioBank, PhysioToolkit, and PhysioNet:

Components of a new research resource for com-

plex physiologic signals. Circulation, 101(23):e215–

e220. Circulation Electronic Pages: http://

circ.ahajournals.org/cgi/content/full/101/23/e215.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,

P., and Witten, I. (2009). The WEKA data mining

software: An update. ACM SIGKDD Explorations

Newsletter, 11(1):10–18.

Heneghan, C., Chua, C., Garvey, J., De Chazal, P.,

Shouldice, R., Boyle, P., and McNicholas, W. (2008).

A portable automated assessment tool for sleep ap-

nea using a combined Holter-oximeter. Sleep,

31(10):1432.

L´evy, P., P´epin, J., Deschaux-Blanc, C., Paramelle, B., and

Brambilla, C. (1996). Accuracy of oximetry for de-

tection of respiratory disturbances in sleep apnea syn-

drome. Chest, 109(2):395.

Lin, C. L., Yeh, C., Yen, C. W., Hsu, W. H., and Hang, L. W.

(2008). Comparison of the indices of oxyhemoglobin

saturation by pulse oximetry in obstructive sleep ap-

nea hypopnea syndrome. Chest, 135(1):86–93.

Magalang, U., Dmochowski, J., Veeramachaneni, S.,

Draw, A., Mador, M., El-Solh, A., and Grant,

B. (2003). Prediction of the Apnea-Hypopnea

Index From Overnight Pulse Oximetry*. Chest,

124(5):1694.

McNames, J. and Fraser, A. (2000). Obstructive sleep ap-

nea classiﬁcation based on spectrogram patterns in the

electrocardiogram. Computers in Cardiology, pages

749–752.

Netzer, N., Stoohs, R., Netzer, C., Clark, K., and Strohl,

K. (1999). Using the Berlin Questionnaire to identify

patients at risk for the sleep apnea syndrome. Annals

of Internal Medicine, 131(7):485.

Ng, A., Koh, T., Baey, E., and Puvanendran, K. (2006).

Speech-like Analysis of Snore Signals for the Detec-

tion of Obstructive Sleep Apnea. In International

Conference on Biomedical and Pharmaceutical Engi-

neering, 2006. ICBPE 2006, pages 99–103. IEEE.

Oliver, N. and Flores-Mangas, F. (2006). HealthGear: a

real-time wearable system for monitoring and analyz-

ing physiological signals. Technical report, MSR-TR-

2005-182, Microsoft Corporation.

Olson, L., Ambrogetti, A., and Gyulay, S. (1999). Pre-

diction of sleep-disordered breathing by unattended

overnight oximetry. Journal of sleep research,

8(1):51–55.

Shinar, Z., Baharav, A., and Akselrod, S. (2000). Obstruc-

tive sleep apnea detection based on electrocardiogram

analysis. In Computers in Cardiology 2000, pages

757–760. IEEE.

UCD (2000). St. Vincent’s University Hospital /

University College Dublin Sleep Apnea Database.

http://www.physionet.org/pn3/ucddb/.

Young, T., Palta, M., Dempsey, J., Skatrud, J., Weber,

S., and Badr, S. (1993). The occurrence of sleep-

disordered breathing among middle-aged adults. New

England Journal of Medicine, 328(17):1230.

Zamarr´on, C., Gude, F., Barcala, J., Rodriguez, J., and

Romero, P. (2003). Utility of Oxygen Saturation and

Heart Rate Spectral Analysis Obtained From Pulse

Oximetric Recordings in the Diagnosis of Sleep Ap-

nea Syndrome*. Chest, 123(5):1567.

AN IMPROVED APPROACH FOR REAL-TIME DETECTION OF SLEEP APNEA

175