modeling respiratory movement signals during central and obstructive sleep apnea events using...

UNCORRECTEDPROOF

1 Modeling Respiratory Movement Signals During Central

2 and Obstructive Sleep Apnea Events Using Electrocardiogram3

4 A. H. KHANDOKER and M. PALANISWAMI

5 Department of Electrical & Electronic Engineering, The University of Melbourne, Parkville, VIC 3010, Australia

6 (Received 15 July 2010; accepted 12 October 2010)

7 Associate Editor John H. Linehan oversaw the review of this article.8

9 Abstract—Obstructive sleep apnea (OSA) causes a pause in10 airflow with continuing breathing effort. In contrast, central11 sleep apnea (CSA) event is not accompanied with breathing12 effort. CSA is recognized when respiratory effort falls below13 15% of pre-event peak-to-peak amplitude of the respiratory14 effort. The aim of this study is to investigate whether a15 combination of respiratory sinus arrhythmia (RSA), ECG--16 derived respiration (EDR) from R-wave amplitudes and17 wavelet-based features of ECG signals during OSA and CSA18 can act as surrogate of changes in thoracic movement signal19 measured by respiratory inductance plethysmography20 (RIP). Therefore, RIP and ECG signals during 250 pre-21 scored OSA and 150 pre-scored CSA events, and 10 s22 preceding the events were collected from 17 patients. RSA,23 EDR, and wavelet decomposition of ECG signals at level 924 (0.15–0.32 Hz) were used as input to the support vector25 regression (SVR) model to recognize the RIP signals and26 classify OSA from CSA. Using cross-validation test, an27 optimal SVR (radial basis function kernel; C = 28 and28 e = 222 where C is the coefficient for trade-off between29 empirical and structural risk and e is the width of30 e-insensitive region) showed that it correctly recognized31 243/250 OSA and 139/150 CSA events (95.5% detection32 accuracy). Independent test was performed on 80 OSA and33 80 CSA events from 12 patients. The independent test34 accuracies of OSA and CSA detections were found to be 92.535 and 95.0%, respectively. Results suggest superior perfor-36 mance of SVR using ECG as the surrogate in recognizing the37 reduction of respiratory movement during OSA and CSA.38 Results also indicate that ECG-based SVR model could act39 as a potential surrogate signal of respiratory movement40 during sleep-disordered breathing.

41 Keywords—Central sleep apnea, Obstructive sleep apnea,

42 Respiratory movements, Support vector regression, Electro-

43 cardiogram.44

45

46INTRODUCTION

47The respiratory pump consists of not only the

48respiratory muscles, but also the structure of rib cage

49and abdomen controlled by the brain, spinal cord, and

50the peripheral nerves. The control mechanism of the

51respiratory pump is different from that of the upper

52airway during sleep.20 Obstructive sleep apnea (OSA)

53is a temporary closure of the upper airway during sleep

54when air is prevented from entering lungs. The onset of

55each OSA is associated with inspiratory efforts against

56a closed airway, which increase parasympathetic

57activity leading to a bradycardia, but as the physio-

58logical stress builds up during the apnea, the sympa-

59thetic activity predominates. This peaks shortly after

60the moment of arousal at which time there is systemic

61vasoconstriction, hypertension, and a tachycardia.

62Termination of OSA requires arousal from a deeper-

63to-lighter stage of sleep or wakefulness.20

64An observation during central sleep apnea (CSA),

65on the other hand, reveals an absence of respiratory

66movements which differentiate these apneas from

67OSA. These observations can be confirmed by sleep

68studies in which abdominal and chest wall movement

69recordings are combined with airflow and oximetry.

70CSA is recognized when respiratory effort falls below

7115% of pre-event peak-to-peak amplitude of the

72respiratory effort. The arousals are less frequent than

73in OSA because of the absence of any increased

74inspiratory effort as an arousal stimulus. During CSA,

75the PCO2gradually rises and when it reaches the

76apnoeic threshold, a period of hyperventilation then

77begins to lower the PCO2again.20

78Approximately, 50% of heart failure patients expe-

79rience sleep-disordered breathing,3 with either central

80sleep apnea (CSA) or obstructive sleep apnea (OSA)

81usually predominating, but both are often present.

82Optimizing the treatment for congestive heart failure is

Address correspondence to A. H. Khandoker, Department of

Electrical & Electronic Engineering, The University of Melbourne,

Parkville, VIC 3010, Australia. Electronic mail: ahsank@unimelb.

edu.au

Annals of Biomedical Engineering (� 2010)

DOI: 10.1007/s10439-010-0189-x

� 2010 Biomedical Engineering Society

Journal : ABME MS Code : 10439 PIPS No. : 189 h TYPESET h DISK h LE h CP Dispatch : 19-10-2010 Pages : 114 4

Au

tho

r P

ro

of

UNCORRECTEDPROOF

83 the first step of CSA treatment. Various forms of

84 continuous positive airways pressure (CPAP) are used

85 for CSA treatment. The use of CPAP has been shown

86 to improve cardiac function and quality of life, and

87 lessen the need for transplantation. However, it is

88 important to distinguish CSA from OSA events

89 because different types of sleep apneas may require

90 different treatment approaches. Also, this distinction is

91 important to understand the pathophysiological

92 mechanisms of different types of apnea. Different

93 algorithms on cheaper ambulatory ECG monitoring

94 technology for the detection of OSA have recently been

95 reported.15 However, the reported methods assess

96 whether the ECG could detect OSA during each min-

97 ute of the recording and hence cannot detect the actual

98 events. We have recently developed a model for

99 detection of individual apnea events and types (apnea/

100 hypopnea),10 but that model could not distinguish

101 OSA from CSA events.

102 At present, the clinical technique for respiratory

103 monitoring during sleep is the use of two respiratory

104 inductance plethysmography (RIP) measurements (rib-

105 cage and abdominal), which can be used with reason-

106 able agreement with those of the standard reference

107 methods of measurement for respiratory effort-related

108 arousals (RERA), central hypopnea–apnea, and

109 Cheyne–Stokes respiration (CSR) associated with

110 central sleep apnea CSR-CSA.1 For OSA and CSA

111 screening using RIP signals, patients are referred to a

112 sleep clinic for an overnight polysomnograpic (PSG)

113 study. This process is expensive, needing dedicated

114 systems and attendant personnel. It is not suitable for

115 mass screening; waiting periods are long, and rural

116 areas lack sleep clinics; therefore, vast majority

117 remains undiagnosed for long periods. Other disad-

118 vantage of laboratory PSG is the artificial sleep envi-

119 ronment (which may affect how the patient sleeps).

120 Fast, clinical or domiciliary, diagnostics would be

121 hugely beneficial in terms of early detection and cost.

122 This has recently been acknowledged by the US Center

123 of Medicare Services who now allow for home diag-

124 nosis reimbursement. There is evidence that for people

125 with a high probability for sleep apnea, use of facility-

126 based PSG does not result in better outcomes over an

127 ambulatory approach in terms of diagnosis and CPAP

128 titration.22 The American Sleep Disorders Association

129 classified the different monitors that have been used in

130 sleep studies into four categories, depending on which

131 channels they record and evaluate.7 Among them, type

132 IV may include monitors that record one or two bio-

133 parameters. For example, ECG-based apnea event

134 detection algorithm as proposed by our previous

135 studies10,11 fit into type IV monitor. In order to make

136 the best use of the ECG-based apnea classification

137 device, modeling surrogate respiratory signals using

138ECG signals could be useful for apnea types classifi-

139cation in the situation of unattended home setting.

140The phenomenon of respiratory modulation of the

141chest ECG has already been widely noted, and there

142are a number of studies in the literature reporting

143algorithms for estimation of respiratory signals from

144the ECG.12,13,16,21 ECG are reported to be modulated

145by respiration mainly in two different ways: (1) respi-

146ratory sinus arrhythmia (RSA) using modulations of

147beat-to-beat heart rate caused by respiration, or

148(2) ECG-derived respiration (EDR) using variations of

149the beat morphology, especially of the R-peak, induced

150by respiration.11–13,16,21

151In this study, we use discrete wavelet transform of

152ECG signals to extract the respiratory-related com-

153ponent and combine with RSA and EDR using sup-

154port vector regression (SVR). The primary objective is

155to determine whether surrogate respiratory signal ex-

156tracted from ECG signal surrounding OSA and CSA

157can correlate with respiratory signal measured by RIP.

158A secondary objective is to determine whether changes

159in mean peak-to-peak amplitude of the surrogate

160respiratory signal can discriminate CSA from OSA.

161METHODS

162Subjects and Sleep Studies

163In total, 17 and 12 PSG recordings were used to

164develop and validate our classification algorithms,

165respectively. PSGs were collected from the database of

166the Institute of sleep and breathing, Austin Hospital,

167Melbourne, Australia (a collaborative partner with the

168University of Melbourne). The research protocol was

169approved by Austin Ethics in Human Research Com-

170mittee (H2008/03252). Brief descriptions of the dat-

171abases are as follows. The PSG recordings of the

172training set of 17 sleep apnea patients [(mean ± SD)

173age 54 ± 5 years, body mass index (BMI) 31 ± 2

174kg/m2, 10 men and 7 women], and a test set of 12 sleep

175apnea patients [(mean ± SD) age 51 ± 2 years,

176body mass index (BMI) 30 ± 5 kg/m2, 10 men and 2

177women] were analyzed by Pro-Fusion software version

1783 (Compumedics Pty Melbourne, Australia). Each

179PSG study included electroencephalogram (channels

180C3–A2 and C4–A1), left- and right electrooculograms,

181leg movements, body positions, thoracic and abdomi-

182nal wall expansion (using by RIP), oronasal airflow (by

183nasal pressure), arterial oxygen saturation SaO2 (by

184pulse oximetry), and ECG (sampling frequency =

185250 Hz with a resolution of 16 bits/sample). All the

186subjects were free of any cardiac history. Diagnosis

187was based on clinical symptoms and PSG outcomes.

188Respiratory events were scored by an expert using

189criteria proposed by the AASM.1 Only full apnea

A. H. KHANDOKER AND M. PALANISWAMI


Au

tho

r P

ro

of

UNCORRECTEDPROOF

190 events were considered in this study excluding hypop-

191 neas. There has been no personal bias in the event

192 selection process. 250 pre-scored OSA events and 150

193 pre-scored CSA events were randomly selected from 10

194 patients and 7 patients, respectively, to develop the

195 algorithm. In order to validate the algorithm, 80 OSA

196 and 80 CSA events were taken from 12 patients.

197 Scoring Method

198 The OSA event was scored as 80–100% reduction of

199 oronasal airflow for >10 s using the criteria that a

200 reduction of 50–85% from peak-to-peak (mean posi-

201 tive to mean negative) amplitude of RIP signals

202 (sampling frequency = 32 Hz) must be in both the

203 thoracic and abdominal movement channels (which

204 were recorded on separate channels for this study).

205 Events were followed by either an oxygen desaturation

206 of ‡3% or an arousal.

207 The CSA event, on the other hand, was scored as

208 the absence of oronasal airflow for >10 s in using the

209 criteria that a reduction of >85% from peak-to-peak

210 (mean positive-to-mean negative) amplitude of RIP

211 signals must be in both the thoracic and abdominal

212 movement channels (which were recorded on separate

213 channels for this study). Events were followed by either

214 an oxygen desaturation of ‡3% or an arousal.

215 The range of apnea–hypopnea index (AHI) of

216 patients was 12.5–85.45. Figure 1 shows an example of

217 a PSG recording (5 min) in a patient with OSA and

218 CSA events (AHI = 48.5 events/h).

219ECG Clips

220A total of 250 (training set) and 80 (test set)

221simultaneous ECG and RIP (thoracic movement) sig-

222nal clips of 20-s duration (10-s preceding and 10-s

223following the onset of OSA events) collected from 330

224pre-scored OSA events were selected. Similarly, 150

225(training set) and 80 (test set) simultaneous ECG and

226RIP signal clips (10-s preceding and 10-s following the

227onset of CSA events) collected from 230 pre-scored

228CSA events with/without arousals were extracted for

229analysis in this study.

230Derivation of EDR Signals Through RSA, EDR, and

231Wavelet-EDR from Wavelet-Decomposed ECG Signals

232RSA

233The sampling frequency of the ECG signal was

234250 Hz. In order to calculate R–R interval with an

235error in accuracy of less than one beat per min, ECG

236signal was resampled using cubic spline interpolation

237at a sampling frequency of 4000 Hz.9

238EDR

239The method for calculation of an EDR is

240through interpolation of the amplitude values of the

241R-peaks of the ECG. A highpass (0.05 Hz cutoff

242frequency) FIR filter was used to remove the base-

243line wandering from that caused in R-peak ampli-

244tudes by respiration.

FIGURE 1. Example of digital recordings (5 min) in a patient with obstructive and central sleep apnea (AHI 5 48.5). The upper-most trace is SpO2 (%, pulse oximeter); second, nasal pressure; third, thoracic movement signals; Fourth, abdominal movementsignals, and the lowermost is that of ECG signals (Lead I); CnA 5 Central apnea; ObA 5 Obstructive apnea.

Modeling Respiratory Movement Signals


Au

tho

r P

ro

of

UNCORRECTEDPROOF

245 Wavelet-EDR

246 A discrete wavelet transform of 20-s ECG clip was

247 used to decompose the signal into a set of approximate

248 and detailed coefficients of level up to 10. Recon-

249 structed detailed coefficients up to level 10 were com-

250 puted. A symlet wavelet with order 8 was chosen as the

251 mother wavelet for decomposition. Wavelet decom-

252 position of ECG signals (during central apnea

253 breathing episodes), RSA, and EDR of a sleep apnea

254 patient (AHI = 48.5) are illustrated in Fig. 2. Recon-

255 structed decomposition level 9 (i.e., 0.15–0.32 Hz) was

256 chosen as the wavelet-EDR feature because the range

257 of breathing frequency overlaps with that range.

258 In this study, SVR model was considered to distin-

259 guish CSA from OSA events from a combination

260 (RSA, EDR, and wavelet-EDR) of ECG features. In

261 order to match the number of samples in RIP as the

262 target signal (total 640 samples in 20 s with a sampling

263 frequency, 32 Hz), RSA, EDR and wavelet-EDR

264signals were resampled using cubic spline interpolation

265(MATLAB) to produce 640 samples.

266SVR

267Support Vector Machines (SVMs) are a relatively

268new class of learning machines that have evolved from

269the concepts of structural risk minimization (SRM) (in

270the pattern recognition case) and regularization theory

271(in the regression case).23 The major difference between

272SVMs and many other Neural Network (NN)

273approaches is that instead of tackling problems using

274the traditional empirical risk minimization (ERM)

275method, SVMs use the concept of regularized ERM.

276This has enabled researchers to use SVMs with

277potentially high capabilities on smaller datasets with-

278out running into the usual difficulties of overfitting and

279poor generalization performance. Geometrically, the

280basis of SVM theory is to nonlinearly map the input

0 60 120-202

x 10-3

sample number

Thora

cic

0 20 40 60 80 100 120-2024

x 10-3

Time (seconds)

EC

G

0 60 120

-101

x 10-4

Wv1

0 60 120-505

x 10-4

Wv2

0 60 120-101

x 10-3

Wv3

0 60 120-101

x 10-3

Wv4

0 60 120-101

x 10-3

Time (seconds)

Wv5

0 60 120-505

x 10-4

Wv6

0 60 120-202

x 10-4

Wv7

0 60 120-101

x 10-4

Wv8

0 60 120-505

x 10-5

Wv9

0 60 120-505

x 10-5

Wv10

0 60 1200.8

11.2

RS

A

0 60 1201.5

22.5

x 10-3

Time (seconds)

R-w

ave

FIGURE 2. Two-minute recording of thoracic movement signal (V), ECG signals and its reconstructed wavelet-detailed decom-position up to level 10, RSA, and EDR of a sleep apnea patient (AHI 5 48.5) during four central apnea breathing episodes. RSA,EDR, and wavelet-decomposed ECG signals were resampled using cubic spline interpolation to make 3840 samples.



Au

tho

r P

ro

of

UNCORRECTEDPROOF

281 data into some (possibly infinite dimensional) feature

282 space where the problem may be treated as a linear

283 one. In particular, when tackling regression problems

284 using SVMs, the output is a linear function of position

285 in feature space.

286 e-SV Regression

287 The standard SV regression problem (e-SV Regres-

288 sion) is formulated as follows.

289 Suppose we are given a training set (xi = input;

290 zi = target):

H ¼ ðx1; z1Þ; ðx2; z2Þ; . . . ; ðxN; zNÞf g

xi 2<dL

zi 2<

292292 where dL is the dimension of xi, which is assumed

293 to have been generated based on some unknown but

294 well-defined map:

g^

: <dL ! <

296296zi ¼ g

^ðxÞ þ noise

so that

xi ¼ xi^þnoise

298298 Let us define a set of functions (implicitly, as will be

299 seen later) /j : <dL ! <; 1 £ j £ dH, which collectively

300 makes up a map from input space to feature space,

301 namely, /j : <dL ! <dH; where dH is the dimension of

302 xi in the feature space / and

/ðxÞ ¼

/1ðxÞ

/2ðxÞ

:

:

:

/dHðxÞ

2

6

6

6

6

6

6

6

6

6

4

3

7

7

7

7

7

7

7

7

7

5

304304 Using the function, /j : <dL ! <; 1 £ j £ dH, we aim

305 to find a nonlinear approximation to g^

: gðxÞ ¼306 wT/ðxÞ þ b which is a linear function of position in

307 feature space. The usual e-SVR method of selecting

308 w and b is to minimize the regularized risk functional:

minw;b;n;n�

Rðw; b; n; n�Þ ¼1

2wTwþ

C

N1Tnþ

C

N1Tn�

310310ðwT/ðxiÞ þ bÞ � zi � e� ni

Such that :

�ðwT/ðxiÞ þ bÞ � �zi � e� n�in; n� � 0

ð1Þ

312312where 12wTw characterizes the complexity of the model

313(the larger the 12wTw; the larger the gradient of g(x) in

314feature space, and hence the more the g(x) may vary

315for a given variation in input, x), and 1N1Tnþ 1

N1Tn�

316characterizes the empirical risk associated with it. A

317schematic representation of the SVR using e-insensitive

318loss function is illustrated in Fig. 3.

319The constant C> 0 controls the trade-off between

320empirical risk minimization (and potential over-fitting)

321if C is large and complexity minimization (and

322potential under-fitting) if C is small.

323The constant e> 0 term in (1) is included to give the

324model a degree of noise insensitivity. Essentially, as

325long as zi 2 e £ g(xi) £ zi + e, ni*= ni = 0, and so

326there will be no empirical risk associated with small

327perturbations (specifically, those perturbations which

328leave the point lying inside the e tube), giving the risk

329functional a degree of noise immunity (assuming that e

330is well matched to the noise present in the training

331data).

332The function K(xi,xj) = /(xi)T/(xj) is called the

333kernel function. It is not difficult to show that our

334approximation function g(x) may be written in terms

335of the kernel function:

gðyÞ ¼X

i

ðai � a�i ÞKðxi; yÞ þ b ð2Þ

337337Note that the feature map functions /j : <dL ! < are

338hidden by the kernel function. It is well known that for

339any symmetric function K : <dL � <dL ! < satisfying

340Mercer’s condition,5 there exists an associated set of

341feature maps /j : <dL ! < (although calculating what

342these maps are may not be a trivial exercise). Indeed,

343we may start with such a kernel function and, with no

344knowledge of / at all (except that it exists), optimize

345and use a SV regressor. Feature space remains entirely

346hidden throughout the process. Linear kernel defined

347as K(xi, xj) = xi.xj and Radial basis function (RBF)

FIGURE 3. Support vector regression (SVR) and its param-eters (e, n, n*). e is a precision parameter representing theradius of the tube located around the regression functiong(x) (the broken lines). The region enclosed by the tube isknown as ‘‘e-insensitive zone’’ since the loss function as-sumes a zero value in this region and does not penalize theprediction errors with magnitudes smaller than e. Two positiveslack variables, n and n*, can be used to measure the deviationfrom the boundaries of the e-insensitive zone. They representthe distance from actual values to the corresponding bound-ary values of e-insensitive zone.



Au

tho

r P

ro

of

UNCORRECTEDPROOF

348defined as Kðxi; xjÞ ¼ exp

� xi�xjk k2

2r2

� �

were used in this

349 study as the kernel function where r denotes the width

350 of the RBF. Cherkassky and Ma4 pointed out that for

351 multivariate d-dimensional problems, the RBF width

352 parameter r is set as rd = 0.1–0.5, where d is number

353 of input variables. In this study, r = 0.7 is used for all

354 the experiments.

355 Training and Testing the Linear SVR

356 A tenfold cross-validation scheme was adopted to

357 evaluate the generalization ability of the SVR as a

358 classifier for OSA and CSA. In this scheme, a data set

359 of 400 events (250 OSA and 150 CSA) set was ran-

360 domly divided into 40 subsets, of which 10 events were

361 used for testing and the remaining 390 events were

362 used to train SVR parameters. There was no fixed

363 proportion of OSA/CSA maintained during cross

364 validation. This was repeated for other subsets so that

365 all subsets were used as the cross-validation test sam-

366 ple. Only linear and RBF kernel were tested in this

367 study. All SVM architectures were trained and tested

368 on the MATLAB SVM toolbox.8

369 RESULTS

370 Maximum mean cross-validation accuracy was

371 found to be 95.5% (±3.8%) with C = 28; e = 222

372 using RBF kernel, where C is the coefficient for trade-

373 off between empirical and structural risk, and e is the

374 width of e-insensitive region. The best feature subset

375 was the combination of the wavelet decomposition

376 level (level 9; 0.15–0.32 Hz) of ECG, RSA, and EDR

377 (R-wave amplitude of ECG signals).

378 A grid search proposed by Bao and Liu2 was used in

379 this study for setting parameters of C and e. To reduce

380 the computational burden, a finer grid search on that

381 region was conducted only after identifying a better

382 region on the grid. We first used a coarse grid search

383 using linear kernel and found the best (C, e) as (26, 222)

384 with tenfold cross-validation accuracy 90.4% on 50

385 OSA and 50 CSA events. After the best (C, e) was

386 found, the model was trained with whole training set

387 (400 events) again to generate the final classifier. The

388 parameter set of C and e which generated the maximum

389 OSA/CSA classification accuracy (95.5%) was consid-

390 ered as the best parameter set (Table 1). Test Studies

391 (80 OSA and 80 CSA events) were used to provide an

392 independent test performance assessment of our model.

393 Figure 4 shows an example of RIP (thoracic

394 movement) signal and SVR-based surrogate respira-

395 tory signal of 20-s duration (10-s preceding and 10-s

396 following the onset of OSA events) from a pre-scored

397OSA event. Similarly, Fig. 5 shows RIP (thoracic

398movement) signal and SVR-based surrogate respira-

399tory signal of (10-s preceding and 10-s following the

400onset of CSA events) from a pre-scored CSA event. A

401schematic flow representation of the proposed recog-

402nition system for OSA and CSA has been shown in

403Fig. 6.

404Figure 7 shows the correlations of percentage (%)

405drop in thoracic and surrogate respiratory signals from

40610 s preceding the events for 250 OSA and 150 CSA

407events. Significant correlations (r = 0.76; p< 0.01 for

408OSA and r = 0.54; p< 0.01) were found between

409reductions in thoracic and that in surrogate respiratory

410signals. Using surrogate respiratory signals, 243 OSA

411and 139 CSA events (as cross-validation training set)

412are correctly recognized. The threshold was set at 85%

413drop from the preceding thoracic movement. An

414independent test was carried out on 80 OSA and 80

415CSA events collected from 12 patients (Table 2). Test

416results showed that 74 out of 80 OSA events and 76 out

417of 80 CSA events have correctly been identified by the

418SVR model.

419The Bland–Altman plot is the preferred method for

420assessing whether an established and a new measure-

421ment technique agree. It shows the paired difference

422between two observations on each subject against the

423mean of these two observations Figs. 8a and 8b show

424Bland–Altman-plots for OSA and CSA events,

425respectively. Percentage (%) drop in surrogate respi-

426ratory signals during OSA events were unbiased, and it

427overestimated the % drop in thoracic signals by less

428than 1% (mean bias = +0.82%; +2SD: +13.62%,

42922SD: 211.97%). Percentage (%) drop in thoracic

430signals during CSA events was overestimated by less

431than 2% mean bias = +1.23%; +2SD: + 7.68%,

43222SD: 25.23%). Visual inspection of Fig. 8b shows

433that the surrogate respiratory signal tended to estimate

434lower percentage drop in thoracic signal unbiased but

435it tended to under estimate higher percentage drop.

TABLE 1. SVR model parameter selection.

Kernel e C Accuracy (%)

Linear 221 222 79.9

221 83.2

22 84.5

222 24 88.9

26 90.4

28 88.9

224 210 87.6

212 86.9

214 83.8

RBF (r = 0.7) 222 26 91.2

28 95.5

210 94.3



Au

tho

r P

ro

of

UNCORRECTEDPROOF0 2 4 6 8 10 12 14 16 18 20

-1.5

-1

-0.5

0

0.5

1

1.5

2x 10

-3

Time (seconds)

Thora

cic

sig

nal

0 2 4 6 8 10 12 14 16 18 20-1.5

-1

-0.5

0

0.5

1

1.5

2x 10

-3

Time (seconds)

Surr

ogate

respirato

ry s

ignal

Mean positive peaks

Mean negative peaks

FIGURE 4. An example of thoracic signal and ECG-based surrogate respiratory signal 10 s preceding and 10 s after the start of anOSA event. Thick dashed line represents the start of the pre-scored OSA event.

0 2 4 6 8 10 12 14 16 18 20-6

-4

-2

0

2

x 10-3

Time (seconds)

Thora

cic

sig

nal

0 2 4 6 8 10 12 14 16 18 20-5

-4

-3

-2

-1

0

1

2

3x 10

-3

Time (seconds)

Surr

ogate

respirato

ry s

ignal

Mean negative peaks

Mean positive peaks

FIGURE 5. An example of thoracic signal and ECG-based surrogate respiratory signal 10 s preceding and 10 s after the start of aCSA event. Thick dashed line represents the start of the pre-scored CSA event.



Au

tho

r P

ro

of

UNCORRECTEDPROOF

436 In order to show the comparative performance,

437 OSA/CSA classification performances (overall accu-

438 racy, sensitivity, and specificity) on independent test set

439 (80 OSA and 80 CSA events) using (1) RSA, (2) EDR,

440 and (3) a combination of RSA, EDR, and wavelet-

441 EDR features by SVR are summarized in Table 3. The

442 results show that SVR-based OSA/CSA classifier out-

443 performs the RSA or EDR or wavelet-EDR-based

444 classification. However, wavelet-EDR-based classifi-

445 cation performance is better that RSA or EDR-based

446 classification as shown in Table 3.

447 DISCUSSION

448 SVR which is a powerful regression tool, has been

449 applied to model respiratory effort signal during sleep

450 apnea events to classify OSA and CSA events for the

451 first time in this study. Despite the rapid advances in

452 the understanding of the nature of sleep-related

453 breathing disorders over the past few years, sleep

454 apneas still often remain unrecognized and untreated.

455 The study results indicate a correlation between

456changes in amplitudes of surrogate respiratory signal

457obtained from ECG and the amplitudes of breathing

458movement during OSA and CSA. Results also showed

459that SVR with an optimal parameter set correctly

460recognized 243/250 OSA and 139/150 CSA events

461(95.5% recognition) in the cross-validation training

462set. Independent test results showed 93.75% classifi-

463cation accuracy on 160 events (80 OSA and 80 CSA).

464EDR

465The morphology and amplitude of the ECG change

466with respiration (i.e., EDR) because of the motion of

467the electrodes, the changing intra-thoracic impedance.

468However, the performance of the EDR approach is

469affected by the position of the electrical axis of the

470heart. For example, if the ECG lead and electrical axis

471of the heart are perpendicular, the project of ECG

472variations caused by respiration also decrease. Also,

473airway obstruction modulation depth on the ECG

474varies from patient to patient and throughout the night

475with changes in postural position as reported by a

476previous study.14 Moody et al.13 described a method of

477calculating EDR by measuring mean axis direction in

478two roughly orthogonal leads. No quantitative com-

479parison of their technique with reference methods was

480provided, and the effect of different body position was

481not investigated. The assessment of the respiratory

482signal by the RSA technique has been especially suit-

483able to derive the respiratory rate from ECG record-

484ings.6 However, unreliable respiratory patterns are

485observed in RSA and EDR in the presence of pre-

486mature ventricular contract (PVC) beat; therefore, the

487PVC beat needs to be excluded.

488Combining wavelet-based EDR with RSA and

489R-wave amplitude-based EDR in our study appears

490capable of providing useful discriminative information

491about respiration during OSA and CSA (Table 3).

492An important aspect of our study is that postural po-

493sition appears to have only a small effect on the

494wavelet-EDR estimates because wavelet decomposi-

495tion excludes the body position’s influence on ECG.

496Therefore, it can be inferred that single-lead ECG

497technique may be able to sufficiently identify major

498respiratory events such as OSA and CSA. If ECG

499signals get corrupted or disturbed, then SVR cannot

500derive surrogate respiratory movement signals. The

501artifacts that may corrupt ECG signals include body

502movements, improper attachment of ECG electrodes

503on chest, etc.

504Phase differences between thoracic signal and

505ECG-derived surrogate respiratory signal (as shown in

506Figs. 4, 5) indicate either paradoxical breathing

507(out-of-phase movements of the thorax and abdomi-

508nal cavities), or at least phase-shifted movements

FIGURE 6. Schematic representation of training the pro-posed diagnostic system for recognizing OSA and CSA-basedon ECG signals. Respiratory inductance plethysmogram sig-nal was used as target signal for training. Pre_sResp 5 meanpeak-to-peak amplitude of 10-s sResp preceding an OSA orCSA event. In_sResp 5 mean peak-to-peak amplitude of 10-sfollowing the onset of the OSA or CSA event.



Au

tho

r P

ro

of

UNCORRECTEDPROOF

509 (the peaks of the indicators of the chest and abdomen

510 do not line up exactly). Often, the beginning of an

511 obstructive apnea is the peak of the previous in-phase,

512 large breath. Similarly, the ending of an obstructive

513 apnea is often the beginning of the next in-phase,

514 large breath. However, an obstructive apnea does not

515 always have to have paradoxical breathing associated

516 with it. Sometimes, the movement channels indicate

517 in-phase breathing. This is because not enough nega-

518 tive airway pressure has developed to partially com-

519 press the chest cavity.1 The thoraco-abdominal

520 separation influences flow and pressure in both the

521 superior and inferior vena cava.24

522 Regression Model

523 Model parameters such as C and e are selected by

524 users based on a priori knowledge and/or user exper-

525 tise.19 Obviously, this approach is not appropriate for

526non-expert users. It is well known that SVR general-

527ization performance (estimation accuracy) depends on

528a good setting of meta-parameters parameters C, e and

529the kernel parameters. The problem of selection of the

530optimal parameter is further complicated by the fact

531that SVR model complexity (and hence its general-

532ization performance) depends on all three parameters.

533Existing software implementations of SVR usually

534treat SVR meta-parameters as user-defined inputs. In

535this article we focused on the choice of C and e, rather

536than on selecting the kernel function. OSA/CSA clas-

537sification accuracy on cross-validation data set rather

538than root mean squared error (RMSE) was considered

539the performance criterion to select the best parameter

540set. Therefore data-intensive computational burden

541was significantly reduced. Parameter C determines the

542trade-off between the model complexity (flatness) and

543the degree to which deviations larger than e are toler-

544ated in optimization formulation (Eq. 1). For example,

40 45 50 55 60 65 70 75 80 85 90 95 10050

55

60

65

70

75

80

85

90

95

100

% drop in surrogate respiratory signal

% d

rop

in

th

ora

cic

sig

na

l

OSA

CSA

FIGURE 7. Percentage (%) drops in thoracic and surrogate respiratory signals from 10 s preceding the events. A total of 250 OSAand 150 CSA events are shown. Using surrogate respiratory signals, 243 OSA and 139 CSA events are correctly recognized. 85%drop from the preceding thoracic movement was considered as the threshold.

TABLE 2. Comparisons on original OSA/CSA events and the model-based events of 17 patients in the training set and 12 patientsin the independent test set.

Number

of patients

Original ModelClassification

accuracy (%)

OSA

detection (%)

CSA

detection (%)OSA CSA OSA CSA

Training set 17 250 150 243 139 95.50 97.20 92.66

Test set 12 80 80 74 76 93.75 92.50 95.0



Au

tho

r P

ro

of

UNCORRECTEDPROOF

545 if C is too large (infinity), then the objective is to

546 minimize the empirical risk only, without regard to

547 model complexity part in the optimization formula-

548 tion. Parameter e controls the width of the e-insensitive

549 zone, used to fit the training data. The value of e can

550 affect the number of support vectors used to construct

551 the regression function. The greater the e, the fewer the

552 support vectors that are selected. On the other hand,

553 greater e-values result in more ‘‘flat’’ estimates. Hence,

554 both C and e-values affect model complexity (but in a

555 different way).

556 In this study, the RIP was used as a gold standard of

557 respiratory monitoring which measures the changes in

558 the expansion and contraction of the thoracic and

559abdominal cavities. Placing an elastic belt in series with

560a piezo-electric sensor does this. One is placed at the

561level of the upper thorax, generally right underneath

562the arms to pick up rib-cage movements, and the other

563is placed at the level of the umbilicus to pick up

564abdominal movements. This study focused on ECG

565and SVR-based surrogate respiratory signal that can

566classify OSA and CSA events. The proposed algorithm

567can be applied to recognize OSA or CSA types after

568the event is detected. For example, our earlier study10

569can be used to detect individual apnea events. Al-

570though hypopneas were not used in this study, it may,

571however, be an interesting topic to look at how SVR

572can model respiratory movements during hypopneas.

84 86 88 90 92 94 96 98-10

-5

0

5

10

Average (%dropthoracic

and %dropsurrogate

)

Diffe

rence (

%dro

p thora

cic

and %

dro

psurr

ogate

)

50 55 60 65 70 75 80 85-20

-10

0

10

20

Average (%dropthoracic

and %dropsurrogate

)

Diffe

rence (

%dro

p thora

cic

and %

dro

psurr

ogate

) (a)

(b)

mean

mean

mean+2SD

mean -2SD

mean+2SD

mean -2SD

FIGURE 8. Bland–Altman plots of the relationship of the difference between % drop in thoracic signal ands that in surrogaterespiratory signals for 250 OSA events (a) and 150 CSA events (b) vs. their average value. Mean bias (—), +2SD and 22SD lines areshown. SD 5 standard deviation.

TABLE 3. OSA/CSA classification performance (overall accuracy, sensitivity, and specificity) on independent testset (80 OSA and 80 CSA) of RSA, EDR, and a combination of RSA, EDR and wavelet-EDR features using SVR.

RSA EDR Wavelet-EDR

SVR (RSA + EDR +

wavelet-EDR)

Sens (%) Spec (%) Acc (%) Sens (%) Spec (%) Acc (%) Sens (%) Spec (%) Acc (%) Sens (%) Spec (%) Acc (%)

92.31 77.78 89.16 89.23 44.44 79.52 91.4 87.67 90.52 92.5 95 93.75

Acc, Accuracy; Sens, Sensitivity; Spec, Specificity. Sensitivity is for OSA detection accuracy and specificity is for CSA

detection accuracy.



Au

tho

r P

ro

of

UNCORRECTEDPROOF

573 In future, esophageal catheter pressure signal which

574 represents the respiratory effort (pressure) in sleep-

575 disordered breathing could be attempted to use as a

576 standard reference of respiratory signal to develop the

577 model for surrogate respiratory effort signal in sleep-

578 related breathing events. In addition to OSA/CSA

579 discrimination, other potential application of our

580 model may be to assess the severity of airway

581 obstruction in patients with obstructive airway disease

582 including asthma, bronchiolitis, and chronic obstruc-

583 tive pulmonary disease (COPD).584

585 ACKNOWLEDGMENTS

586 The authors would like to thank members of the

587 Sleep Research Team at the Institute of Breathing and

588 Sleep at Austin Hospital, Melbourne for their assis-

589 tance during sleep studies.590

591 REFERENCES

592 1American Academy of Sleep Medicine Task Force593 (AASM). Sleep-related breathing disorders in adults: rec-594 ommendations for syndrome definition and measurement595 techniques in clinical research. Sleep 22:667–689, 1999.596 2Bao, Y., and Z. Liu. A Fast Grid Search Method in597 Support Vector Regression Forecasting Time Series, Vol.598 4224. LNCS, pp. 504–511, 2006.599 3Bradley, T. D., and J. S. Floras. Sleep apnea, heart failure.600 Part I. Obstructive sleep apnea. Circulation 107:1671, 2003.601 4Cherkassky, V., and Y. Ma. Practical selection of SVM602 parameters and noise estimation for SVM regression.603 Neural Netw. 17:113–126, 2004.604 5Cortes, C., and V. Vapnik. Support vector networks.605 Mach. Learn. 20:273–297, 1995.606 6Cysarz, D., R. Zerm, H. Bettermann, M. Fruhwirth,607 M. Moser, and M. Kroz. Comparison of respiratory rates608 derived from heart rate variability, ECG amplitude, and609 nasal/oral airflow. Ann. Biomed. Eng. 36(12):2085–2094,610 2008.611 7Ferber, R., R. Millman, M. Coppola, et al. Portable612 recording in the assessment of obstructive sleep apnea.613 ASDA standards of practice. Sleep 17(4):378–392, 1994.614 8Gunn, S. Support Vector Machines for Classification and615 Regression, ISIS Tech. Rep. Southampton, UK: Univ.616 Southampton, 1998.617 9Khandoker, A. H. Regulation of Cardiac Rhythm in Avian618 Embryos and Hatchlings in Altered Environments. Ph.D.619 dissertation, Muroran Institute of Technology, Japan.

62010Khandoker, A. H., J. Gubbi, and M. Palaniswami.621Automated scoring of obstructive sleep apnea and hypo-622pnea events using short term electrocardiogram recordings.623IEEE Trans. Inf. Technol. Biomed. 13(6):1057–1067, 2009.62411Khandoker, A. H., M. Palaniswami, and C. K. Karmakar.625Support vector machines for automated recognition of626obstructive sleep apnoea syndrome from electrocardiogram627recordings. IEEE Trans. Inf. Technol. Biomed. Eng. 13(1):62837–48, 2009.62912Mazzanti, B., C. Lamberti, and J. de Bie. Validation of an630ECG-derived respiration monitoring method. Comput.631Cardiol. 30:613–616, 2003.63213Moody, G., R. Mark, A. Zoccola, and S. Mantero.633Derivation of respiratory signals from multi-lead ECGs.634Comput. Cardiol. 12:113–116, 1985.63514O’Brien, C., and C. Heneghan. A comparison of algo-636rithms for estimation of a respiratory signal from the sur-637face electrocardiogram. Comput. Biol. Med. 37:305–314,6382007.63915Penzel, T., J. McNames, P. de Chazal, B. Raymond, A.640Murray, and G. Moody. Systematic comparison of differ-641ent algorithms for apnoea detection based on electrocar-642diogram recordings. Med. Biol. Eng. Comput. 40:402–407,6432002.64416Pinciroli, F., R. Rossi, and L. Vergani. Detection of elec-645trical axis variation for the extraction of respiratory646information. Comput. Cardiol. 12:499–502, 1985.64717Roche, F., J. M. Gaspoz, I. Court-Fortune, P. Minini,648V. Pichot, D. Duverney, F. Costes, J. R. Lacour, and649J. C. Barthelemy. Screening of obstructive sleep apnea650syndrome by heart rate variability analysis. Circulation651100(13):1411–1415, 1999.65218Roche, F., V. Pichot, E. Sforza, I. Court-Fortune,653D. Duverney, F. Costes, M. Garet, and J. C. Barthelemy.654Predicting sleep apnoea syndrome from heart period: a655time-frequency wavelet analysis. Eur. Respir. J. 22(6):656937–942, 2003.65719Scholkopf, B., J. Burges, and A. Smola. Advances in658Kernel Methods: Support Vector Machine. MIT Press,6591999.66020Shineerson, J. M. Sleep Medicine: A Guide to Sleep and Its661Disorder. Blackwell Publishing, pp. 230–231, 2005.66221Travaglini, A., C. Lamberti, J. DeBie, and M. Ferri.663Respiratory signal derived from eight-lead ECG. Comput.664Cardiol. 25:65–68, 1998.66522Trikalinos, T. A., S. Ip, G. Raman, S. Cepeda, E. M. Balk,666C. D’Ambrosio, and J. Lau. Home Diagnosis of Obstruc-667tive Sleep Apnea–Hypopnea Syndrome. Technology668Assessment Program, Technical Report. Agency for669Healthcare Research and Quality, 2007.67023Vapnik, V. N. The Nature of Statistical Learning Theory.671New York: Springer, 1995.67224Wexler, L., D. H. Bergel, I. T. Gabe, G. S. Makin, and673C. J. Mills. Velocity of blood flow in normal human venae674cavae. Circ. Res. 23:349–359, 1968.

675



Au

tho

r P

ro

of

modeling respiratory movement signals during central and obstructive sleep apnea events using...

Documents