DETECTION

OF OBSTRUCTIVE SLEEP APNEA FROM THE

FREQUENCY ANALYSIS OF HEART RATE VARIABILITY

Abraham Otero

, Xos

e Vila

, Francisco Palacios

and Francisco J. Coves

Department of Software and Knowledge Engineering, University San Pablo CEU, 28668 Madrid, Spain

Department of Computer Science, University of Vigo, Campus As Lagoas s/n, 32004 Ourense, Spain

Hospital Universitary Hospital of Elche, Partida Huertos y Molinos, s/n, Elche, 03203, Spain

Keywords:

Obstructive sleep apnea (OSA), Biosignal processing, Heart rate variability.

Abstract:

This paper presents a new algorithm for the detection of Obstructive Sleep Apnea (OSA) from a single elec-

trocardiogram lead. It is based on the alterations that OSA patients present in the LF and HF bands of the

heart rate variability power spectrum. The algorithm calculates the power of the spectrum in two bands that

roughly corresponding with the LF and HF bands. Then the ratio between the power of the low band and the

power of the high band is obtained. If this ratio is greater than a certain threshold the patient is classiﬁed as

having OSA, otherwise he/she is classiﬁed as not having OSA. Then the algorithm was validated over the test

data set of the Apnea-ECG Database, classifying correctly 29 of 30 recordings.

1 INTRODUCTION

Obstructive Sleep Apnea (OSA) is a frequent sleep-

breathing disorder characterized by the presence of

total and/or partial cessations of respiratory airﬂow

while the patient is sleeping (K

oves, 1999). These

cessations –called apneas if they are total, and hy-

popneas if they are partial– are usually caused by a

collapse in the upper airway and they produce a dis-

ruption of the patient’s sleep architecture diminishing

the refreshing effects of nocturnal rest. The overall

effect is a decrease in the patient’s quality of life, and

an increase in mortality and morbidity. OSA has a

high prevalence –it is estimated to affect 4% of male

adults and 2% of female adults – and it is recognized

as an important public health issue.

The gold standard for the diagnosis of OSA is

based on an analysis of a polysomnogram, a record-

ing of a wide range of physiological parameters ob-

tained while the patient is asleep. These sleep stud-

ies are expensive because they require the patient to

spend a night in a Hospital Sleep Unit, which must

be equipped with sophisticated and expensive elec-

tronic systems, as well as a dedicated staff. Hence

there is an interest in developing reliable diagnostic

techniques based on a smaller number of physiolog-

ical parameters that do not require dedicated Sleep

Units. The 2000 Computers in Cardiology Challenge

encouraged the development of such techniques. It

consisted of developing an algorithm capable of clas-

sifying patients as OSA or non-OSA from the modi-

ﬁed lead V2 of the ECG (Penzel et al., 2002). For this

purpose, a database of polysomnographic recordings -

the Apnea-ECG Database- was made available to the

participants of the competition (Penzel et al., 2000).

Several clinical studies have shown that patients

with moderate-to-severe sleep apnea present alter-

ations on the heart rate variability (HRV) spectrum

(Narkiewicz et al., 1998). OSA patients present in-

creased low frequency (LF) spectral power and de-

creased high frequency (HF) spectral power. There-

fore, it is no surprise that most of the 2000 Chal-

lenge participants took advantage of information aris-

ing from the spectral analysis of the HRV (Penzel

et al., 2002). One of the simplest, but effective, al-

gorithms was the one presented by Drinnan et al.

(Drinnan et al., 2000). This algorithm calculates the

power of the HRV spectrum in two different bands,

one between 0.01 and 0.05 Hz, and another between

0.005 and 0.01 Hz. The ﬁrst one corresponds ap-

proximately with the LF band, where patients suffer-

ing from OSA present increased activity. The second

band serves as a reference level: the algorithm calcu-

lates the ratio between the power of the ﬁrst band and

the power of the second one, and it checks if this value

is greater than a certain threshold. If so, the patient

359

Otero A., Vila X., Palacios F. and J. Coves F. (2010).

DETECTION OF OBSTRUCTIVE SLEEP APNEA FROM THE FREQUENCY ANALYSIS OF HEART RATE VARIABILITY.

In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 359-362

DOI: 10.5220/0002590503590362

 SciTePress

is classiﬁed as OSA. This algorithm correctly clas-

siﬁed 90% of the training and test recordings of the

Apnea-ECG Database. Despite its relative success,

Drinnan’s algorithm only takes advantage of OSA pa-

tients presenting increased activity in the LF band and

the band’s limits were chosen based on a visual in-

spection of the recordings’ spectrum. Thus, it may be

possible to improve these results by tuning the bands’

boundaries and by merging information arising from

both LF and HF bands.

This paper presents a new algorithm for classify-

ing patients as OSA or non-OSA from a single ECG

lead. This algorithm is based on the alterations of the

heart rate variability that OSA patients present. Sec-

tion 2 presents the algorithm which enables patient

classiﬁcation. The results of the algorithm validation

over the Apnea-ECG Database are presented in Sec-

tion 3. Finally, the results obtained are discussed and

a series of conclusions on the paper are given.

2 MATERIALS AND METHODS

After heartbeat detection, the RR intervals were ﬁl-

tered, in order to eliminate the effects of false pos-

itives and false negatives during heartbeat detection.

Then they were resampled at 4 Hz. Cubic splines

were used for the signal interpolation; a linear inter-

polation would have produced more distortion –low

pass ﬁlter effect– than the cubic alternative (Vila et al.,

1997). The ﬁltered and resampled RR intervals were

stored in ASCII ﬁles.

OSA patients have an increased activity in the LF

band of the HRV power spectrum, and a decreased

activity in the HF band. Thus the ratio between the

power of the LF band and the power of the HF band

power should be larger for OSA patients than for con-

trol patients. This is the idea behind our algorithm.

Therefore, we need to ﬁnd the limits of both bands,

and a threshold that allows us to distinguish between

OSA and control patients.

A Matlab script that takes as inputs the lower and

upper limits of two bands corresponding with the LF

and HF regions was created. The script loads the

RR distances from the ASCII ﬁles and it calculates

the spectrum of the complete RR intervals time series

corresponding with each of the 10 normal patients.

For each patient it calculates the ratio between the

power of the LF band and the power of the HF band.

The maximum of these values is considered as the

boundary between control and OSA patients. Thus,

it is used as the threshold for the classiﬁcation of the

20 OSA patients of the training set. If the ratio be-

tween the two spectrum bands of a class A recording

is greater than the threshold, the patient is classiﬁed

as OSA. Otherwise, he/she is normal. After its exe-

cution, the script reports the numbers of correctly and

incorrectly classiﬁed recordings.

Given certain values for the limits of the LF and

HF bands, they will be more suitable for our purpose

the more class A recordings they classify correctly. A

second Matlab script was created to provide the bands

to the ﬁrst script, and to analyze its output. This script

starts with two initial bands selected on the basis of

a visual inspection of the spectrum of the 30 train-

ing set recordings. The initial limits for the low band

were set between 0.01 and 0.05 Hz, and for the high

band between 0.18 and 0.28 Hz. This script gradually

changes the lower and upper limits of the bands by

increasing and decreasing their values.

This second script performs an exhaustive search

looking for the optimal classiﬁcation bands. For each

band limit modiﬁcation, it invokes the ﬁrst script and

stores the number of recordings correctly classiﬁed,

as well as the bands used in this classiﬁcation. Once it

has checked all the preprogrammed bandwidths, it an-

alyzes the results and presents a list of the best bands

that have been found. This script ran for two days on

a PC with an Intel Core 2 Duo processor at 2.2 GHz

and 4 GB of RAM. The results obtained suggested

that the best classiﬁcations were obtained when the

upper limit of the low band was close to or equal to

the lower limit of the high band, and they both had a

value of approximately 0.08 Hz.

The scripts were modiﬁed in order to force the up-

per limit of the low band to always be equal to the

lower limit of the high band. The initial values of

the bands were also modiﬁed to be close to the val-

ues with which the best results were obtained in the

ﬁrst run. Finally, we decreased the magnitude of the

increases and decreases that would be applied to the

lower and upper limits of each band. This time the

initial limits for the low band were set between 0.02

and 0.08 Hz, and for the high band they were set be-

tween 0.08 and 0.30 Hz, and the script applied ﬁner

modiﬁcations to the bands. After running the script

for one day, we found that by setting the lower limit

of the low bandwidth between 0.021 and 0.048 Hz;

the upper limit of the low band (and the lower limit of

the high band) between 0.056 and 0.078 Hz; and the

upper limit of the high band between 0.232 and 0.276

Hz, all the control recordings and 28 of the 30 OSA

recordings were correctly classiﬁed.

To select the ﬁnal frequencies, we looked for the

bands which maximize the separability of the class A

and the class C test set recordings. To this end, we

subtracted the threshold used in the classiﬁcation; i.e,

the maximum of the ratios obtained for the class C

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

360

recordings, from the LF/HF ratios obtained for each

of the class A recordings, and we calculated the sum

of all these values. This value can be considered as a

measure of the separability of the class of the control

patients and the class of the OSA patients: the higher

this value is, the greater the average distance from the

OSA patients to the control patient which is “closest”

to the OSA class is, according to this metric.

The ﬁnal bands were selected in such a way that

they maximize this distance. The low band was set be-

tween 0.026 and 0.06 Hz, and the high band between

0.06 and 0.25 Hz. These bands yield a threshold value

of 0.42349; i.e., the maximum of the ratios between

the power of the low band and the power of the high

band for the 10 control patients was 0.42349. This

value will be the boundary between classes A and C.

The optimal classiﬁcation bands we have found

differ signiﬁcantly from the values traditionally asso-

ciated with the LF and HF bands. We tried to classify

the recordings using bands as close as possible to the

traditional LF and HF bands (between 0.04 and 0.14

Hz, and between 0.14 and 0.4 Hz, respectively). To

this end, we launched our training script with these

initial values for the bands’ boundaries and we al-

lowed only small changes to be made. In the best

case obtained, 6 training set and 8 testing set record-

ings were misclassiﬁed. These results suggest that the

bands in which patients suffering from OSA present

HRV alterations do not exactly match the traditional

deﬁnitions of the LF and HF bands.

3 RESULTS

Using the bands selected in the training phase, we

validated our algorithm using the test set recordings

of the Apnea-ECG Database. Before applying the

algorithm, the threshold value was rounded to 0.43.

This change did not affect the classiﬁcation results

obtained for the training set recordings. All patients

in the control test set were correctly classiﬁed as non

OSA. Of these recordings, the one which provided

the bigger LF/HF ratio value, i.e, the one which was

“closest” to the OSA class, was 17c, which yielded a

value of 0.41. Among the patients who suffered OSA,

only one was classiﬁed as healthy (the 15a); all others

were correctly classiﬁed.

Tables 1 and 2 summarize the results obtained

over the training and test sets, respectively. These ta-

bles also show the classiﬁcation results for the ﬁve

Class B recordings of the training and test sets. Given

that these recordings correspond to patients who are

on the borderline between normality and suffering

from the disorder, it is not possible to use them to

judge the quality of the results of a binary classiﬁer as

the one proposed here.

Table 1: Training set recordings classiﬁcation results.

Class Classiﬁcation

OSA non-OSA

Class A 18 2

Class B 3 2

Class C 0 10

Table 2: Test set recordings classiﬁcation results.

Class Classiﬁcation

OSA non-OSA

Class A 19 1

Class B 2 3

Class C 0 10

4 DISCUSSION

Our algorithm correctly classiﬁes 95% of the 60 OSA

and control recordings of the Apnea-ECG Database,

compared to 90% of the Drinnan et al. algorithm that

inspired ours. This has been achieved by taking ad-

vantage of the altered HRV that OSA patients present

both in the LF band and in the HF band and by in-

cluding a semi–automatic training stage to ﬁne tune

the bands’ boundaries. All the classiﬁcation errors

committed by our algorithm, and by the Drinnan et al.

algorithm, correspond to OSA patients who are mis-

classiﬁed as normal. From the viewpoint of the con-

struction of a diagnostic or a screening test, the classi-

ﬁcation of a patient who suffers from OSA as healthy

is less desirable than the classiﬁcation of a healthy

patient as having OSA. Therefore, we might prefer an

algorithm that, even though it incorrectly classiﬁes a

higher number of patients, it classiﬁes less OSA pa-

tients as healthy.

To achieve this in the training set, we would have

had to set the threshold used by our algorithm to 0.32.

In the case of the test set, the threshold would have

had to have been set to 0.35014. In the ﬁrst case, this

would lead to the classiﬁcation of 6 of the 10 con-

trol patients as having OSA; in the latter 5 control pa-

tients would have been misclassiﬁed. An algorithm

that classiﬁes half or more of the healthy patients as

having OSA does not seem a good compromise.

A visual inspection of the 3 incorrectly classiﬁed

recordings shows that they do not present altered HRV

in the LF or HF bands. Visual inspection also reveals

that some recordings only present alterations in one

of the two bands. In these cases, our algorithm is still

DETECTION OF OBSTRUCTIVE SLEEP APNEA FROM THE FREQUENCY ANALYSIS OF HEART RATE

VARIABILITY

361

able to make the correct classiﬁcation thanks to the

merging of information arising from both bands. In

the case of the Drinnan et al. algorithm, the patient

must present altered activity in the only band the al-

gorithm considers –LF.

The fact that neither the Drinnan et al. algorithm

nor ours classiﬁed any of the control patients as hav-

ing OSA suggests that both algorithms have a rather

high speciﬁcity. This speciﬁcity may have its root in

the fact that healthy patients do not usually show al-

tered HRV activity in any of the two bands, although

some OSA patients do not present the typical altered

activity, despite their disorder.

The scientiﬁc literature states that patients suffer-

ing from OSA have decreased activity in the LF band

and increased activity in the HF band. However, the

best classiﬁcation bands found by our scripts for the

Apnea-ECG Database recordings differ from the tra-

ditional LF and HF bands. Our high band –between

0.06 and 0.25 Hz– incorporates a fragment of the LF

band (between 0.06 and 0.014), and our low band –

between 0.026 and 0.06 Hz– includes only the lower

part of the LF band and the VLF band (Very Low Fre-

quency, between 0.002 and 0.004 Hz).

However, the algorithm is based on patients suf-

fering from OSA having less power in the high band

and more power in the low band, because it calculates

the ratio between the low band and the high band, and

expects that OSA patients have a higher value for this

ratio than healthy patients. Furthermore, in our at-

tempt to ﬁnd bands as close as possible to the LF and

HF bands, the classiﬁcation results are signiﬁcantly

worse than those obtained with the optimal bands –

6 recordings misclassiﬁed in the training set, and 8

in the testing test. Therefore, our results suggest that

OSA patients have increased activity only in the lower

part of the LF band and in the VLF band, and have de-

creased activity not only in part of the HF band, but

also in the 0.06 and 0.014 Hz range of the LF band.

5 CONCLUSIONS

This paper presents a simple but effective algorithm

capable of classifying patients as OSA or non-OSA

on the basis of a single ECG lead. The algorithm cal-

culates the power spectrum of the time series obtained

by ﬁltering and resampling at 4 Hz the RR intervals

extracted from the ECG lead. Then the power of the

bands between 0.026 and 0.06 Hz and between 0.06

and 0.25 Hz is calculated, and then the ratio between

the power of the low band and the high band is ob-

tained. If this ratio is greater than a certain threshold

–0.43– the patient is classiﬁed as OSA. The bound-

aries of the bands and the threshold were obtained

by means of a semi-automatic training stage where

the Apnea-ECG Database training data set was used.

The algorithm was validated on the test data set of

the same database, incorrectly classifying only 1 of

30 control and OSA recordings.

Our future work aims to obtain a screening test

for OSA patients that can be performed at low cost

and in the patient’s home. This can avoid travels to

the hospital, long waiting lists, and other inconve-

niences for the patient. To this end, the test will be

based on MEDIM, a PDA platform with the capabil-

ity of recording ECGs (Presedo et al., 2009). The

algorithm we have developed has a low demand of

computational resources, which makes it suitable to

be implemented in such a device.

ACKNOWLEDGEMENTS

This work was supported by the Spanish MEC and the

European FEDER under the grant TIN2006-15460-

C04-02.

REFERENCES

Drinnan, M., Allen, J., Langley, P., and Murray, A. (2000).

Detection of sleep apnoea from frequency analysis

of heart ratevariability. In Computers in Cardiology

2000, pages 259–262.

oves, P. (1999). Obstructive Sleep Apnea Syndrome.

Springer Verlag.

Narkiewicz, K., Montano, N., Cogliati, C., van de Borne,

P., Dyken, M., and Somers, V. (1998). Altered cardio-

vascular variability in obstructive sleep apnea. Circu-

lation, 98:1071–1077.

Penzel, T., McNames, J., de Chazal, R., Raymond, B., Mur-

ray, A., and Moody, G. (2002). Systematic compari-

son of different algorithms for apnea detection based

on electrocardiogram recordings. Med. Biol. Eng.

Comput, 40:402–407.

Penzel, T., Moody, G., Mark, R., Goldberger, A., and Peter,

J. (2000). The apnea-ecg database. In Computers in

Cardiology 2000, volume 27, pages 255–258.

Presedo, J., Tarasc

o, M., Teijeiro, T., F

elix, P., Castro, D.

(2009). A PDA-based modular and multipurpose sys-

tem for intelligent ubiquitous monitoring. In WISP

2009, pages 223–228.

Vila, J., Palacios, F., Rodr

ıguez-Presedo, J., Fernandez-

Delgado, M., F

elix, P., and Barro, S. (1997). Time-

frequency analisys of herat-rate variability. IEEE Eng.

in medicine and biology magazine, 16:119125.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

362