Proceedings of 2022 APSIPA Annual Summit and Conference 7-10 November 2022, Chiang Mai, Thailand

Obstructive Sleep Apnea Classification Using Snore

Sounds Based on Deep Learning
Apichada Sillaparaya†, Apichai Bhatranand†, Chudanat Sudthongkong††,
Kosin Chamnongthai†, and Yuttapong Jiraraksopakun†
*
Electronic and Telecommunication Engineering Department, Faculty of Engineering,
King Mongkut’s University of Technology Thonburi, Bangkok, Thailand
E-mail:[email protected], [email protected]
††
Medical and Science Media Department, School of Architecture and Design,
King Mongkut’s University of Technology Thonburi, Bangkok, Thailand

Abstract— Early screening for the Obstructive Sleep Apnea an expert and many sensors attached to a patient, hence costly
(OSA), especially the first grade of Apnea-Hypopnea Index and not convenient for a self-test. Screening OSA using the
(AHI), can reduce risk and improve the effectiveness of timely characteristics of snore sound is then a potential alternative
treatment. The current gold standard technique for OSA because it is convenient, simple, and useful for subjects to
diagnosis is Polysomnography (PSG), but the technique must be
performed in a specialized laboratory with an expert and
proceed by themselves. Snore sounds can be one of the
requires many sensors attached to a patient. Hence, it is costly information to diagnosis OSA because it is directly related to
and may not be convenient for a self-test by the patient. The abnormality of breathing conditions caused by obstruction of
characteristic of snore sounds has recently been used to screen the upper respiratory tract [4, 5].
the OSA and more likely to identify the abnormality of breathing Detecting sleep apnea has currently gained significant
conditions. Therefore, this study proposes a deep learning model interest. According to the literature review [7], ECG sensor-
to classify the OSA based on snore sounds. The snore sound data based signals are most commonly and widely used for sleep
of 5 OSA patients were selected from the opened-source PSG- apnea classification, but which sensors or signals are the best
Audio data by the Sleep Study Unit of the Sismanoglio-Amalia is still an active question. On the other hand, obstructive sleep
Fleming General Hospital of Athens [1]. 2,439 snoring and
breathing-related sound segments were extracted and divided
apnea is directly related to snore sounds. Therefore,
into 3 groups of 1,020 normal snore sounds, 1,185 apnea or obstructive sleep apnea classification using snore sounds is
hypopnea snore sounds, and 234 non-snore sounds. All sound also a potential method and interesting for study.
segments were separated into 60% training, 20% validation, and In snore sound classification, the sound characteristics
20% test sets, respectively. The mean of Mel-Frequency Cepstral between normal and abnormal snore sounds are different due
Coefficients (MFCC) of a sound segment were computed as the to different sources of sounds [8-10]. Therefore, the
feature inputs of the deep learning model. Three fully characteristic of snore sounds can be exploited for OSA
connected layers were used in this deep learning model to
screening application of the normal and abnormal snore
classify into three groups as (1) normal snore sounds, (2)
abnormal (apnea or hypopnea) snore sounds, and (3) non-snore sounds. Feature extraction is an important step before the
sounds. The result showed that the model was able to correctly learning process of sound classification. The feature can be
classify 85.2459%. Therefore, the model is promising to use divided into three main groups: the time domain features [11],
snore sounds for screening OSA. frequency domain features [8], and time and frequency
domain feature or wavelet transform [10]. Among many
I. INTRODUCTION feature extraction methods used in snore sound classification,
the Mel-Frequency Cepstral Coefficients (MFCC) is widely
Obstructive Sleep Apnea (OSA) is the most common sleep- chosen as it provides promising accuracy [12]. In addition, the
related breathing disorder and comprises up to one-seventh of various classification methods such as Bayes classifier [13],
the world’s adult population [2]. The consequences of OSA logistic regression [9], AdaBoost classifier [14], Support
both physically and mentally affect the health condition vectors machines (SVM) [15] can also be used along the
because insufficient sleep may cause hypersomnia, leading to sound features to classify the severity of OSA. However, no
microsleep and narcolepsy, diabetes mellitus, coronary artery studies clearly identify which feature extraction and
disease, heart attack, ischemic stroke, and depression, etc [3- classification method provide the optimum result.
5]. At present, Polysomnography (PSG) is the gold standard Deep learning-based classification is of great interest to
technique to screen sleep apnea [4-6] that identify sleep many classification research due to its promising accuracy
disorders from the physiological changes of the body signals, and ability to adapt to new data. In this study, a deep learning
i.e. electroencephalogram (EEG), electrocardiogram (ECG), technique to classify OSA using only snore sounds is
heart rate, eye movement, depth and breathing patterns, snore proposed.
sounds, blood oxygen levels, and skeletal muscle activity.
However, the technique requires a specialized laboratory with

978-616-590-477-3 ©2022 APSIPA APSIPA ASC 2022

1152
Authorized licensed use limited to: Universitas Indonesia. Downloaded on November 07,2024 at 02:27:01 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2022 APSIPA Annual Summit and Conference 7-10 November 2022, Chiang Mai Thailand

II. METHOD

A. Snore Dataset
The snore sound data of 5 adult OSA patients (2 female/ 3
male) were selected from the opened-source PSG-Audio data
for full-night PSG study by the Sleep Study Unit of the
Sismanoglio-Amalia Fleming General Hospital of Athens [1].
The snore sounds were recorded together with the full-night
PSG at the sleep disorders laboratory using a microphone
placed over the tracheal of the patients with the sampling rate
48 kHz. 2,439 snoring and breathing-related sound segments
with an average time duration of 18 seconds were extracted
according to the PSG annotations of nasal and respiratory
events by the medical team. All segments were divided into 3 Fig. 1. The Signal Example of A Snore Sound Segment.
groups of 1,020 normal snore sounds, 1,185 apnea or
hypopnea snore sounds, and 234 non-snore sounds as shown
in Table I. The sound segment data was then separated into 3
groups of 60% training, 20% validation, and 20% test sets for
the classification model.

TABLE I
THE NUMBER OF SNORE SOUNDS AND
BREATHING-RELATED SOUND SEGMENTS

Group/ Type of Fig. 2. The Dimension Reduction on MFCCs to Mean MFCCs for
Number of Segments
Abnormalities Model Inputs.
Hypopnea or Apnea
1,185
Snore
Normal Snore 1,020 C. Deep Learning Classification Model
A deep learning model for the snore sound classification
Non-snore 234
was constructed by three fully connected layers as shown in
Total 2,439 Fig. 3. The number of nodes in hidden layers are 100, 200,
and 100 nodes, respectively. All hidden layers were added
with 50% dropout to prevent the model overfitting. The
B. Data Preprocessing rectified linear activation function (ReLu) was used as the
activation function for all hidden layers. The softmax
Feature extraction was applied to extract the characteristic
activation was used for the output layer for a multi-class
of snore sounds as data input of the deep learning model. Fig.
classification on the three groups of snore sounds as: (1)
1 shows an example of a snore sound segment from the
normal snore sounds, (2) abnormal (apnea or hypopnea) snore
dataset. The Mel-Frequency Cepstral Coefficients (MFCC), a
sounds, and (3) non-snore sounds.
representation of the short-term power spectrum of a sound
based on a linear cosine transform of a log power spectrum on
ReLu Function Softmax Function
a nonlinear Mel-scale of frequency, was computed as the
feature of the snore sounds. The number of MFCCs were X1

typically 128 for any short window of a sound segment, hence X2

comprising a 2D array of MFCCs for one sound segment. In
X3
this study, the mean of 2D MFCCs was computed to reduce P(y = Normal Snoring |x)

the dimension of the input for the classification model as X4

show in Fig. 2. X5 P(y = Apnea or Hypopnea Snoring |x)

P(y = Non Snoring|x)

100 Nodes 100 Nodes

Xn
200 Nodes

Feature Input Hidden Layer Output Layer

n=128 (3 Layers with Dropout 50%) 3 Nodes
)

Fig. 3. The Proposed Deep Learning Classification Model.

1153
Authorized licensed use limited to: Universitas Indonesia. Downloaded on November 07,2024 at 02:27:01 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2022 APSIPA Annual Summit and Conference 7-10 November 2022, Chiang Mai Thailand

D. Model Training and Validation stable than the other groups. The peak difference is smaller in
The mean of MFCCs of each sound segment were used as the case of apnea or hypopnea snore sounds. No peak
data input for the model with number of batch size=32. The difference is clearly observed on the non-snore sounds.
data were randomly separated into 3 groups as 60% training, B. Classification Performance
20% validation, and 20% test sets, respectively. The learning
rate and the number of epochs for training and validation sets TABLE II
was set to 0.001 and 200, respectively. The ADAptive THE CONFUSION MATRIX COMPARISON
Moment estimation (Adam) optimization algorithm was used
for network weight update during the model training. The
categorical cross entropy was used as a loss function for the Predicted
multi-class classification. The accuracy on correct estimation
Hypopnea
was tracked during the training and also evaluated on the Normal
or Apnea
Non-
Total
validation set. Snore snore
Snore
E. Classification Evaluation Normal
165 27 0 192
Snore
In this study, the performance of the deep learning model
on snore sound classification for OSA screening was Hypopnea
evaluated on the test set in terms of accuracy (ACC) and Actual or Apnea 30 211 7 248
Snore
positive predictive value (PPV). The results are shown in
terms of confusion matrix comparison between actual and Non-
6 2 40 48
predicted values and shown in terms of true positive (TP), snore
false negative (FN), false positive (FP), true negative (TN),
and overall accuracy and PPV. The accuracy and PPV were Total 201 240 47 488
calculated as:

TABLE III
THE CLASSIFICATION RESULT

Group of
TP FN FP TN ACC (%) PPV (%)
Data
III. RESULTS
Normal
165 27 36 260 85.9375 82.0896
Snore
A. Feature Extraction
Hypopnea
or Apnea 211 37 29 211 85.0806 87.9167
Snore
Non-
40 8 7 433 83.3333 95.1034
snore

Table II shows the confusion matrix comparison between

actual and predicted values on the test set of 192 normal snore
sounds, 248 apnea or hypopnea snore sounds, and 48 non-
snore sounds. The overall accuracy for snore sound
classification model is 85.2459% on the test set. The
performance of the classification model of each group of
snore sounds is summarized in Table III. The classification
accuracy on normal snore sounds is 85.9375% and slightly
drops to 85.0806% on hypopnea or apnea snore sounds. The
accuracy on non- snore sounds apparently drops to 83.3333%
Fig. 4. The Mean MFCCs of Each Group of Snore Sounds. possibly due to less number of data relative to the others. On
the other hand, the PPV is highest on the non-snore sounds at
Fig.4 shows an example of mean MFCC of each snore 95.1034% and apparently reduced on the hypopnea or apnea
sound segment type from the dataset. The mean MFCC of snore sounds and normal snore sounds at 87.9167% and
normal snore sounds has evidently larger peak difference and 82.0896, respectively.

1154
Authorized licensed use limited to: Universitas Indonesia. Downloaded on November 07,2024 at 02:27:01 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2022 APSIPA Annual Summit and Conference 7-10 November 2022, Chiang Mai Thailand

IV. DISCUSSION AND CONCLUSION Sleep Apnea Using Deep Learning," Sensors, vol. 19, no.
22, p. 4934, 2019.
In this study, the deep-learning classification based on [8] B. Calabrese et al., "A System for the Analysis of Snore
snore and breathing-related sounds for screening OSA were Signals," Procedia Computer Science, vol. 4, pp. 1101-
proposed. The results showed that the snore sound 1108, 2011/01/01/ 2011, doi: 10.1016/j.procs.2011.04.117.
characteristic obtained from the mean MFCCs of three sound [9] J.-W. Kim et al., "Prediction of Obstructive Sleep Apnea
groups can be used to classify the snore sounds with Based on Respiratory Sounds Recorded Between Sleep
promising overall accuracy of 85.2459%. The classification Onset and Sleep Offset," Clin Exp Otorhinolaryngol, vol.
12, no. 1, pp. 72-78, 2019, doi: 10.21053/ceo.2018.00388.
results on each sound group are 85.9375% accuracy and
[10] K. Qian et al., "A Bag of Wavelet Features for Snore
82.0896% PPV for normal snore sounds, 85.0806% accuracy Sound Classification," Annals of Biomedical Engineering,
and 87.9167% PPV for apnea or hypopnea snore sounds, and vol. 47, no. 4, pp. 1000-1011, 2019/04/01 2019, doi:
83.333% accuracy and 95.1034% PPV for non-snore sounds. 10.1007/s10439-019-02217-0.
The accuracy of the proposed method is also in agreement to [11] S. S. Mostafa, F. Mendonça, F. Morgado-Dias, and A.
the accuracy range of 80-95% as reported in the research Ravelo-García, "SpO2 based sleep apnea detection using
community [9, 12, 16]. Yet, it is worthy to note that the deep learning," in 2017 IEEE 21st International
accuracy might not be directly compared due to different data Conference on Intelligent Engineering Systems (INES),
sources and accuracy evaluation methods. 20-23 Oct. 2017 2017, pp. 000091-000096, doi:
10.1109/INES.2017.8118534.
The prediction error on each sound group is possibly due to
[12] F. Shen, S. Cheng, Z. Li, K. Yue, W. Li, and L. Dai,
the imbalance of the number of the data in each group. The "Detection of Snore from OSAHS Patients Based on Deep
confusion of classification between the group of normal snore Learning," Journal of Healthcare Engineering, vol. 2020, p.
and apnea or hypopnea snore might be from the similarity of 8864863, 2020/12/12 2020, doi: 10.1155/2020/8864863.
the mean MFCCs of both groups. The recent update on the [13] J. Solà-Soler, J. A. Fiz, J. Morera, and R. Jané, "Bayes
event annotation of the PSG signal from the data source [1] classification of snoring subjects with and without Sleep
can also affect to the performance of the model. The Apnea Hypopnea Syndrome, using a Kernel method," in
hyperparameters such as number of learning rate, number of 2011 Annual International Conference of the IEEE
epochs, as well as the number of layers and nodes to be used Engineering in Medicine and Biology Society, 30 Aug.-3
in the model can be investigated more to improve the Sept. 2011 2011, pp. 6071-6074, doi:
10.1109/IEMBS.2011.6091500.
classification results. [14] E. Dafna, A. Tarasiuk, and Y. Zigel, "Automatic detection
On the future work, more input data of each sound group of whole night snoring events using non-contact
for the model with optimized hyperparameters of training is microphone," PLoS One, vol. 8, no. 12, p. e84139, 2013,
one approach to improve the performance of the model. The doi: 10.1371/journal.pone.0084139.
balanced number of each snore sound group and updated [15] J. Sun, X. Hu, C. Chen, S. Peng, and Y. Ma, "Amplitude
annotation of the PSG signal on the event of snoring and spectrum trend-based feature for excitation location
breathing-related sound should be also considered. classification from snore sounds," Physiol Meas, vol. 41,
no. 8, p. 085006, Sep 10 2020, doi: 10.1088/1361-
6579/abaa34.
[16] T. Kim, J. W. Kim, and K. Lee, "Detection of sleep
REFERENCES disordered breathing severity using acoustic biomarker and
[1] K. Georgia et al. "PSG-Audio", Science Data Bank, March machine learning techniques," Biomed Eng Online, vol.
14, 2022. Available: 17, no. 1, p. 16, Feb 1 2018, doi: 10.1186/s12938-018-
http://doi.org/10.11922/sciencedb.00345 0448-x.
[2] M. M. Lyons, N. Y. Bhatt, A. I. Pack, and U. J. Magalang,
"Global burden of sleep-disordered breathing and its
implications," Respirology, vol. 25, no. 7, pp. 690-702, Jul
2020, doi: 10.1111/resp.13838.
[3] Chulalongkornhospital. "Obstructive sleep apnea"
Retrieved from http://chulalongkornhospital.go.th/
sleepcenter
[4] Mayo Clinic Staff. "Obstructive sleep apnea." Retrieved
from https://www.mayoclinic.org/diseases-
conditions/obstructive-sleep-apnea/symptoms-causes/syc-
20352090
[5] N. M. Punjabi, "The epidemiology of adult obstructive
sleep apnea," Proceedings of the American Thoracic
Society, vol. 5, no. 2, pp. 136-143, 2008.
[6] Easmed. "Sleep Apnea Diagnosis." Retrieved from
https://www.easmed.com/sleep-apnea-diagnosis/
[7] S. S. Mostafa, F. Mendonça, A. G. Ravelo-García, and
F. Morgado-Dias, "A Systematic Review of Detecting

1155
Authorized licensed use limited to: Universitas Indonesia. Downloaded on November 07,2024 at 02:27:01 UTC from IEEE Xplore. Restrictions apply.