GMM-based Classiﬁers for the Automatic Detection

of Obstructive Sleep Apnea

J.-A. G

omez- Garc

ıa

, J.-L. Blanco-Murillo

, J.-I. Godino-Llorente

, L. A. Hern

andez G

omez

and G. Castellanos-Dom

ınguez

Bioingenier

ıa y Optoelectr

onica group (BYO), Universidad Polit

ecnica de Madrid, 28030, Madrid, Spain

Signal Processing Applications Group (GAPS), Universidad Polit

ecnica de Madrid, 28040, Madrid, Spain

Procesamiento y Reconocimiento de Se

nal group (PRS), Manizales, Colombia

Keywords:

GMM, Supervector, GMM-SVM, Obstructive Sleep Apnea, OSA.

Abstract:

The aim of automatic pathological voice detection systems is to support a more objective, less invasive diagno-

sis of diseases. Those detection systems mostly employ an optimized representation of the spectral envelope;

whereas for classiﬁcation, Gaussian Mixture Models are typically used. However, the study of Gaussian Mix-

ture Models-based classiﬁers as well as Nuisance mitigation techniques, such as those employed in speaker

recognition, has not been widely considered in pathology detection tasks. The present work aims at consid-

ering whether such tools might improve system performance in detection of pathologies, particularly for the

Obstructive Sleep Apnea. Having this in mind, the present paper employs Linear Prediction Coding Coefﬁ-

cients, in conjunction with Gaussian Mixture Model-based classiﬁers for the detection of Obstructive Sleep

Apnea, in a database containing the sustained phonation of vowel /a/. The obtained results demonstrate subtle

improvements compared to using baseline automatic detection system.

1 INTRODUCTION

Obstructive Sleep Apnea (OSA) is a highly prevalent

disease affecting an estimated 2-4% of male popula-

tion between the ages of 40-60 (Puertas et al., 2005),

characterized by recurrent episodes of sleep-related

collapses of the Upper Airway at the level of the phar-

ynx. OSA is usually associated to loud snoring, in-

creased daytime sleepiness, poor quality of life and

impaired work performance (Puertas et al., 2005).

OSA is usually detected on the basis of the analy-

sis of the patients history and physical examinations.

Nevertheless, a full overnight sleep study involving

the recording of physiological variables, as well as

complex post-processing of collected data, is required

to conﬁrm diagnosis. This procedures is expensive

and time-consuming, and patients usually have to be

in waiting lists for years. Those issues have mo-

tivated the research of early diagnosis tools which

aim for more advantageous diagnosis of the pathol-

ogy (Alc

azar et al., 2009). For instance in (Fox et al.,

1989), acoustic cues to the automatic detection of

OSA were found. Particularly, several articulatory,

phonation and resonance characteristic were identi-

ﬁed when comparing voices from OSA patients with

those from healthy ones. With that in mind, it might

be reasonable to consider the automatic detection of

OSA by means of recorded voice signals.

The automatic detection of pathologies using

voice recordings relies on the estimation of param-

eters such as jitter and shimmer, noise measures,

among others spectral parameters such as Mel Fre-

quency Cepstral Coefﬁcients (MFCC) or Linear Pre-

diction Coding (LPC). Above features have been em-

ployed for different pathologies, obtaining different

results depending on the nature of the problem. In

particular, for OSA detection, the representation of

the spectral envelope (either from Fourier analysis

or linear prediction) has proved to be discriminative

(Fern

andez-Pozo et al., 2009). On the other hand, for

classiﬁcation purposes, the Gaussian mixture model

(GMM) has become the standard method in speech

applications, and most notably in speaker recognition

systems, due to, among others, its probabilistic frame-

work, and high-accuracy (Campbell et al., 2006).

Several variations, within the ﬁeld of speaker

recognition, have been proposed for improving the

performance of the GMM classiﬁers. Some of them,

which are to be introduced in the next section, are

364

Gómez-García J., Blanco-Murillo J., Godino-Llorente J., A. Hernández Gómez L. and Castellanos-Domínguez G..

GMM-based Classiﬁers for the Automatic Detection of Obstructive Sleep Apnea.

DOI: 10.5220/0004252503640367

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 364-367

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

the Universal Background Models (UBM), GMM

mixed with Support Vector Machines (GMM-SVM)

or GMM-SVM with nuisance removal. However,

aforementioned techniques are mainly employed on

speaker recognition tasks, while its use on automatic

pathology detection is still under study. Having those

precedents, the aim of this paper is to explore the use-

fulness of the above classiﬁers for the automatic de-

tection of OSA when employing LPC features. The

usage of LPC is supported by the previous studies in

the same task using continuous speech (Elisha et al.,

2011). Moreover, and unlike other works in the same

topic (Blanco-Murillo et al., 2011a; Blanco-Murillo

et al., 2011b), this paper relies merely on the dis-

crimination capability of the sustained phonation of

vowel /a/. This, compared to the usage of contin-

uous speech, restricts the phonetic information that

might be obtained, but turns out to be less complex,

while attaining certain advantages such as immunity

to speaking rate, dialect and intonation of the speakers

(Fern

andez-Pozo et al., 2009).

The paper is organized as follows: Section

2 presents the theoretical background; Section 3

presents the experimental setup; Section 4 presents

the obtained results; ﬁnally Section 5 presents the dis-

cussions as well as some conclusions of the work.

2 THEORETICAL BACKGROUND

Having a data vector

x, a Gaussian Mixture Model

(GMM), deﬁned as a ﬁnite mixture of G multivariate

Gaussian components, is of the form:

x) =

∑

i=1

N (

,Σ

) (1)

where λ

are mixture weights, and N (·) are Gaussian

density functions, having mean

and covariances Σ

By training a GMM on a large speech corpus, cov-

ering most speech characteristics, a general model or

UBM is obtained. In this form, it is possible to derive

(adapt) speciﬁc models (GMM-UBM) coming from

this rather general UBM, and which might behave

better than a GMM trained directly on the dataset.

Considering a binary classiﬁcation problem; two spe-

ciﬁc models are required for representing the healthy

(control) and pathology conditions. Those models are

adapted by means of Maximum A Posteriori (MAP)

adaptation of the UBM means (as it is classically done

for speakers’ veriﬁcation), and are as follows:

(

x) =

∑

i=1

N (

,Σ

) (2a)

(

x) =

∑

i=1

N (

,Σ

) (2b)

where

and

are the adapted means for the nor-

mal GMM-UBM model, g

(

x), and the pathological

GMM-UBM model, g

(

x), respectively.

The log-likelihood decision function chosen for

discriminating if

y belongs to the OSA class is:

Λ(

y) = log(g

(

y)) − log(g

(

y)) (3)

On the other hand, a Support Vector Machine (SVM)

is a discriminative binary classiﬁer constructed from

sums of a kernel function K (·,·) such that:

f (

x) =

∑

i=1

K (

) + d (4)

where t

are ideal outputs (-1 or 1), α

are weights

such that

∑

i=1

= 0|α

> 0; d is a learned con-

stant; and

are the L support vectors obtained from

a training set by an optimization process. In order

to exploit the discriminative power of the SVM and

simultaneously the generalization capabilities of the

GMM, the supervectors are introduced. A supervec-

tor

, is a mapping ψ(·), between an utterance and

a high-dimensional vector, which is usually formed

by stacking the mean vectors of GMM-UBM models

(Kinnunen and Li, 2009). By deﬁning

and

as the supervectors for the models of equation (2), a

linear Kernel might be considered:

K (·,·) =

∑

i=1



−1/2





−1/2



Therefore, the decision function of Eq. (4), for a

test sample

y, might be rewritten:

f (

y) =



∑

i=1

ψ(

)



ψ(

y) + d =

ψ(

y) + d

In addition, some methods have been proposed to

increase performance of GMM-SVM systems, by re-

moving the directions of undesired variability in su-

pervectors before the SVM training. One of such is

the Nuisance Attribute Projection (NAP), which for a

given supervector,

is as follows:

−U





where U

U is an eigenchannel matrix, trained using a

development dataset (Kinnunen and Li, 2009). The

resulting

forms a GMM-SVM-NAP.

3 EXPERIMENTAL SETUP

3.1 Databases

Obstructive Sleep Apnea Database: Was recorded

at Hospital Cl

ınico Universitario de M

alaga, Spain. It

contains recordings of 80 male subjects, with similar

age and Body Mass Index. Half of them suffer from

GMM-basedClassifiersfortheAutomaticDetectionofObstructiveSleepApnea

365

severe OSA, and the other half are either healthy or

suffer from mild OSA. Recordings were collected at

16kHz and 16 bits per second. The speech corpus

includes four sentences, as well as recordings of the

sustained vowel /a/ (Fern

andez-Pozo et al., 2009).

The latter set is the only of interest for this paper.

UPM Database: Was recorded by Universidad

Polit

ecnica de Madrid. It contains 239 normal voices,

and 201 pathological voices with a wide variety of or-

ganic pathologies (nodules, polyps, etc.). It contains

the sustained phonation of the /a/ vowel. The distri-

bution by gender is: 226 females and 130 males. The

age range goes from 9 to 79 years. The recordings

were sampled at 50kHz, 16-bits of resolution, and 2s

long, however they were half-band ﬁltered and down-

sampled to 25kHz. In order to match the same condi-

tions of the OSA database, resampling to 16kHz was

performed, while considering only male adults from

the normal class (101 recordings).

3.2 Methodology

A general outline of the proposed automatic pathol-

ogy detection system, is shown in ﬁgure 1, while the

main stages are described next.

Recording

Preprocessing

Normalization

Windowing

Characterization

LPC

Classiﬁcation

Figure 1: Outline of the automatic voice pathological sys-

tem presented on this work.

In the Preprocessing stage, a minus one-one nor-

malization is utilized. Also, framing and windowing

is performed, by employing 40ms Hammming win-

dows, overlapped 50%.

Next, in the characterization stage a LPC

parametrization is considered by using 12, 16 and 18

coefﬁcients for both UPM and OSA databases.

Finally, classiﬁcation is carried out with GMM,

GMM-UBM, GMM-SVM, and GMM-SVM-NAP,

whose parameters where varied between 2 to 20 to

keep the same ranges analyzed in (Blanco-Murillo

et al., 2011b). For validation of results a 11-fold

cross-validation scheme was employed, assuring that

recordings from the same patients are not used in

both training and testing set. The calculated perfor-

mance measures were: Classiﬁcation accuracy (ACC)

within some conﬁdence intervals (IC), Sensitivity

(SE), Speciﬁcity (SP), ROC curves and Areas Under

ROC Curves (AUC). Assuming 95% conﬁdence, the

IC is estimated as IC = ±1.96

ACC(1 − ACC)/N,

where N is the total number of classiﬁed patterns.

4 RESULTS

The number of LPC parameters and the number of

Gaussians for which the best classiﬁcation rate was

obtained, are presented next. Also, Table 1 summa-

rizes ACC, SE, SP, for those parameters, while Figure

2 presents the corresponding ROC curves.

– GMM: 18 LPC, 16 Gaussians

– GMM-UBM: 16 LPC, 12 Gaussians.

– GMM-SVM: 16 LPCs, 16 Gaussians

– GMM-SVM-NAP: 18 LPC, 16 Gaussians

Table 1: Classiﬁcation Accuracy, Sensitivity and Speciﬁcity

for the OSA database, by using the GMM, GMM-UBM,

GMM-SVM, and GMM-SVM-NAP methodologies.

ACC ± IC SE SP AUC

GMM 54 ± 10 0,60 0,50 0,55

GMM-UBM 53 ± 10 0,52 0,53 0,57

GMM-SVM 65 ± 10 0,57 0,75 0,77

GMM-SVM-NAP 62 ± 10 0,57 0,68 0,63

0 0.2 0.4 0.6 0.8 1

0.2

0.4

0.6

0.8

ROC c urve for the tested classiﬁers

False positive rate

True positive rate

GMM. AUC=0.55

GMM−UBM. AUC=0.57

GMM−SVM. AUC=0.77 .

GMM−SVM−NAP. AUC=0.63 .

Figure 2: ROC Curve for the combination of LPC and Gaus-

sian parameters using the OSA database.

5 DISCUSSIONS AND

CONCLUSIONS

This paper has investigated the usage of GMM-based

classiﬁers, typically employed in speaker recognition,

to the issue of automatic OSA detection. LPC-based

coefﬁcients were chosen for speech parametrization

as they provide uniform resolution across the fre-

quency axis and focus on spectral resonances, which

might be suitable to characterize articulation and res-

onance abnormalities identiﬁed for OSA speakers

(Fox et al., 1989) on sustained speech records, in

a similar way as it was previously shown on vowel

segments (Elisha et al., 2011). Obtained results,

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

366

for the GMM and GMM-UBM classiﬁers (Table 1)

are comparable to previous ﬁndings using MFCC

parametrization (Blanco-Murillo et al., 2011b), re-

inforcing our understanding on the role of speech

spectral envelope for automatic detection of OSA. In

Addition, since the GMM-UBM approach on top of

a MFCC parametrization apparently outperforms the

same scheme when using LPC, but not when the al-

ternative GMM-SVM scheme is considered, the role

of the symbiosis between the features set and the clas-

siﬁcation scheme is highlighted.

The inﬂuence of the training database on the clas-

siﬁcation rates achieved by the GMM-UBM scheme

had been addressed in (Blanco-Murillo et al., 2011b);

concluding that better classiﬁcation results are to be

expected when the characteristics of the database used

to train the UBM match those of the ﬁnal classiﬁca-

tion task. Nevertheless, by the time the experiments

in (Blanco-Murillo et al., 2011b) were developed, the

UPM database was not available and was worth veri-

fying this conclusion on a LPC-parametrization. The

results obtained have shown that GMM-based classi-

ﬁers trained on these databases outperform those for a

speciﬁc but smaller database, matching perfectly what

had been concluded. Nonetheless, the limitations im-

posed by the apnea database are hard due to the usage

of the /a/ sound, which might not be the best choice

for OSA-related phenomena.

Moreover, as shown in Table 1 the best classiﬁca-

tion results were obtained when following the GMM-

SVM approach, outperforming the GMM and GMM-

UBM schemes (up to 10% absolute improvement,

though the large conﬁdence intervals must be taken

into account). This same pattern is observed for the

AUC. These had already been described in (Wang

et al., 2011), and has been veriﬁed for the OSA de-

tection on sustained speech. On the other hand, the

scheme including NAP technique, which was intro-

duced to minimize the effects of undesired variabil-

ity observed in the GMM-SVM classiﬁer, was found

sit in between the previous. The limited performance

of the NAP method might be explained by the dif-

ﬁculty in ﬁnding the spurious sources of variability

within the supervector space, which should have con-

tributed to an improvement in classiﬁcation. Since

the methodology for a correct discrimination of OSA-

related phenomena is still an open issue, specially re-

garding the selection of the features, accuracy rates

may be enhanced in a number of alternative ways. Re-

sults in this paper suggest that improvement should be

expected on the basis of more complex classiﬁers and

by focusing on spectral resonances analysis.

ACKNOWLEDGEMENTS

This research was carried out under grants: TEC2009-

14123-C04 from the Spanish Ministry of Education;

AL11-P(I+D)-022 and Ayudas para la realizaci

on del

doctorado (RR01/2011) from Universidad Polit

ecnica

de Madrid, Spain; and partially funded by the Span-

ish Ministry of Science and Innovation as part of the

TEC2009-14719-C02-02 (PriorSpeech) project.

REFERENCES

Alc

azar, J., Fern

andez, R., Blanco, J., Hern

andez, L.,

opez, L., Linde, F., and Torre-Toledano, D. (2009).

Automatic speaker recognition techniques: A new

tool for sleep apnoea diagnosis. Am. J. Respir. Crit.

Care Med.

Blanco-Murillo, J., Hern

andez, L., Fern

andez, R., and

Ramos, D. (2011a). Introducing non-linear analy-

sis into sustained speech characterization to improve

sleep apnea detection. Advances in Nonlinear Speech

Processing, pages 215–223.

Blanco-Murillo, J. L., Fern

andez-Pozo, R., Torre-Toledano,

D., Caminero, J., and L

opez, E. (2011b). Analyz-

ing training dependencies and posterior fusion in dis-

criminant classiﬁcation of apnea patients based on

sustained and connected speech. In INTERSPEECH,

pages 3033–3036.

Campbell, W., Campbell, J., Reynolds, D. A., Singer, E.,

and Torrescarrasquillo, P. (2006). Support vector ma-

chines for speaker and language recognition. Com-

puter Speech & Language, 20(2-3):210–229.

Elisha, O., Tarasiuk, A., and Zigel, Y. (2011). Detection of

obstructive sleep apnea using speech signal analysis.

In MAVEBA.

Fern

andez-Pozo, R., Blanco-Murillo, J. L., Hern

andez-

omez, L., L

opez-Gonzalo, E., Alc

azar Ram

ırez, J.,

and Toledano, D. T. (2009). Assessment of severe

apnoea through voice analysis, automatic speech, and

speaker recognition techniques. EURASIP J. Adv. Sig-

nal Process, 2009:6:1–6:11.

Fox, A. W., Monoson, P. K., and Morgan, C. D. (1989).

Speech dysfunction of obstructive sleep apnea. a

discriminant analysis of its descriptors. Chest,

96(3):589–95.

Kinnunen, T. and Li, H. (2009). An Overview of Text-

Independent Speaker Recognition: from Features to

Supervectors. Image Processing.

Puertas, F. J., Pin, G., Mar

ıa, J. M., and Dur

an, J. (2005).

Documento de consenso nacional sobre el s

ındrome

de apneas-hipopneas del sue

no. Grupo Espa

nol De

Sue

no.

Wang, X., Zhang, J., and Yan, Y. (2011). Discrimination

between pathological and normal voices using GMM-

SVM approach. Journal of voice, 25(1):38–43.

GMM-basedClassifiersfortheAutomaticDetectionofObstructiveSleepApnea

367