Elsevier

Expert Systems with Applications

Volume 123, 1 June 2019, Pages 237-245
Expert Systems with Applications

Classification of snoring sound based on a recurrent neural network

https://doi.org/10.1016/j.eswa.2019.01.020Get rights and content

Highlights

  • The length, frequency, and period of snoring episodes (SE) differ per individual.

  • We proposed a classifier for snoring based on a recurrent neural network.

  • The classifier differentiated SEs from non-SEs by learning episode features.

  • The method can be divided to segmentation, feature extraction, and classification.

  • Despite the small dataset, the proposed classifier showed extremely high accuracy.

Abstract

Snoring is a sleep disorder that may have adverse effects on an individual's health and social activities. Polysomnography is the most common way to diagnose snoring but involves considerable time and cost. Many recent studies have attempted to classify snoring and non-snoring. However, since the length, frequency, and period of snoring episodes (SE) differ according to the individual being measured, it is very difficult to develop a general reference point to classify snoring. Therefore, in order to classify different snoring patterns and noise for different individuals, a learning-based snoring classification algorithm is essential. To this end, this study proposes a classification method based on a recurrent neural network (RNN) that can classify SEs and non-snoring episodes (NSEs) by learning the features of an individual's SEs and NSEs, measured in daily life based on the subjects’ sleep recordings using smartphone. The method proposed in this study can be largely divided into segmentation, feature extraction, and classification. The performance of this study was evaluated through statistical parameters. Despite the fact that the proposed RNN-based classifiers were trained using a relative small dataset, they exhibited an extremely high accuracy of 98.9%.

Introduction

Lack of sleep induces chronic fatigue, lethargy, and daytime sleepiness and has a considerable impact on social activities. Snoring is one of the main causes of sleep disruption and it is a breathing interference that occurs as the upper airways close during sleep or when breathing becomes blocked due to restricted airways. Snoring is a sleep disorder that influences a large part of the population, and over 60% of adult men and 44% of adult women over the age of 40 are known to snore (Dalmasso and Prota, 1996, Duckitt et al., 2006, Lugaresi et al., 1980). Moreover, snoring is not only a symptom that results from sleep disruption, but it may give rise to serious diseases, such as sleep apnea or ischemic cerebral diseases (Cavusoglu et al., 2007, Dalmasso and Prota, 1996, Duckitt et al., 2006, Lugaresi et al., 1980, Norton and Dunn, 1985, Wilkin, 1985).

Polysomnography that measures various biological variables through special equipment is the most common way to evaluate the severity of a snoring disease (Duckitt et al., 2006). However, polysomnography not only requires significant time and cost, but it is difficult to measure the same signals that the subject exhibits on a daily basis since they are measured while sleeping in an unfamiliar environment. Further, it is difficult for the test administrator to manually test snoring signals that are measured over a long period of time (Osborne et al., 1999, Osman et al., 1998). Therefore, various sleep-related applications are now being developed to measure snoring at home, and there is increased interest in snoring prevention pillows and other home care products (Bhat et al., 2015, Cistulli and Grunstein, 2005). Still, measuring snoring remains a difficult task. Electrocardiography (ECG), pulse, respiration, and other such biological signals typically have periodicity and frequency aside from special cases. However, snoring is relatively irregular, and every individual snores at a different frequency. In addition, snoring has a large range from small sounds that resemble breathing to extremely loud sounds (> 95 dB) (Emoto, Abeyratne, & Kawano et al., 2018), and the length of snoring signals also differs for each person. Since everyone has different features, developing a generalized snoring detection algorithm is a considerable challenge.

Numerous studies have been conducted to classify and detect snoring. In recent studies, Hwang et al. (2015) classified snoring signals not based on audio, but based on a support vector machine using sleep data that were stored through a contactless sensor. Ng, San Koh, Puvanendran, and Abeyratne (2008) classified snoring signals through a level—correlation - dependent threshold and a translation - invariant discrete wavelet transform. Cavusoglu et al. (2007) used various feature vectors of snoring signals to classify them through a linear regression from a two-dimensional plane of the principle component algorithm. Duckitt et al. (2006) extracted voice signal features through Mel-frequency cepstral coefficients (MFCC) and used the hidden Markov mode to classify snoring signals. Snoring sounds were also classified by applying short-time energy and zero-crossing rate (ZCR) (Abeyratne et al., 2005, Cavusoglu et al., 2007, Fiz et al., 1996). Other studies have extracted even more features and snoring from respiratory sounds through an unsupervised learning algorithm (Azarbarzin & Moussavi, 2011).

Since the features of snoring signals are different for every subject, it is very difficult to detect snoring episodes using a threshold. However, since every individual also has snoring episode lengths, periods, and frequencies that are relatively consistent, if these features are learned and used in a customized algorithm, this will increase the efficiency of detecting snoring episodes.

Therefore, this study developed an automatic classifier based on deep learning that can automatically classify snoring signals and non-snoring signals from measured signals. The ZCR, Short-Time Fourier Transform (STFT), and MFCC, which are frequently used in voice signal processing, were utilized to extract features from measured signals. The extracted features were manually labeled (annotated), then used as the training set, validation set, and test set for the recurrent neural network (RNN). The performance evaluation of the proposed deep learning-based classifier was verified using statistical values (sensitivity, specificity, precision, and accuracy) and the F1-score.

Section snippets

The overall structure of the proposed method

The snoring classification system proposed in this study can be structured into feature extraction and classification, and the overall structure of this study is shown in Fig. 1. During the feature extraction stage, after signals recorded during sleep were divided into snoring episodes (SE) and non-snoring episodes (NSE), as shown in Fig. 1(a), the ZCR, STFT, and MFCC were used to extract features as shown in Fig. 1(b). In the next stage, as shown in Fig. 1(c), the features that were extracted

Dataset

Table 2 shows the configuration of the episodes used as the input vector of RNN. Eight thousand labeled episodes for the input vector of RNN consisted of 5600 SE and 2400 NSE. As described above, the NSE is a sound excluding the snoring, and it consists of voice, deep breaths, music, silence, and others (TV, footsteps, water running, etc.). In the case of voice and music, the noise was recorded by artificially executing music or by reading a passage.

Feature extraction and representation

Feature extraction for SE and NSE took place

Conclusions

In this study, features that were extracted from SE and NSE were used to develop an automatic classifier based on an RNN with outstanding classification performance. ZCR, STFT, and MFCC, which are often used to extract features from voiced signals, were used to extract features related to the frequency of recorded snoring signals and non-snoring signals (voice, movement, door closing sound, and others) and used as the RNN's input vector to improve the classifier's classification accuracy.

Since

Declarations of interest

None.

Disclosures

All authors have approved the final article.

Acknowledgment

This work was supported by the Korea Institute of Industrial Technology [EO180016]. The funding source has no role in study design; in the collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication.

References (44)

  • S. Bhat et al.

    Is there a clinical role for smartphone sleep apps? comparison of sleep cycle detection by a smartphone application to polysomnography

    Journal of clinical sleep medicine

    (2015)
  • M. Cavusoglu et al.

    An efficient method for snore/nonsnore classification of sleep sounds

    Physiological Measurement

    (2007)
  • P.A. Cistulli et al.

    Medical devices for the diagnosis and treatment of obstructive sleep apnea

    Expert Review of Medical Devices

    (2005)
  • F. Dalmasso et al.

    Snoring: Analysis, measurement, clinical implications and applications

    European Respiratory Journal

    (1996)
  • S.B. Davis et al.

    Comparison of parametric representation for monosyllabic word recognition in continuously spoken sentences

    IEEE Transactions on Acoustics, Speech, and Signal Processing

    (1990)
  • W.D. Duckitt et al.

    Automatic detection, segmentation and assessment of snoring from ambient acoustic data

    Physiological Measurement

    (2006)
  • T. Emoto et al.

    Artificial neural networks for breathing and snoring episode detection in sleep sounds

    Physiological Measurement

    (2012)
  • J.A. Fiz et al.

    Acoustic analysis of snoring sound in patients with simple snoring and obstructive sleep apnoea

    European Respiratory Journal

    (1996)
  • İ. Güler et al.

    An expert system for detection of electrocardiographic changes in patients with partial epilepsy using wavelet based neural networks

    Expert Systems

    (2005)
  • S. Hochreiter et al.

    Long short-term memory

    Neural Computation

    (1997)
  • S.H. Hwang et al.

    Polyvinylidene fluoride sensor-based method for unconstrained snoring detection

    Physiological Measurement

    (2015)
  • J.S.R. Jang

    Audio signal processing and recognition

    (2011)
  • Cited by (28)

    • Learning and ensemble based MPC with differential dynamic programming for nuclear power autonomous control

      2023, Expert Systems with Applications
      Citation Excerpt :

      Since the time-dependent states and action samples are treated as independent in DNN, it is impossible to model the changes in time series. Therefore, a recurrent neural network (RNN) is proposed as an alternative (Lim, Jang, Lim, & Ko, 2019). In RNN, the output of neurons can directly act on themselves in the next time step, so the results of this action involve all the historical information.

    • A parameter-changing zeroing neural network for solving linear equations with superior fixed-time convergence

      2022, Expert Systems with Applications
      Citation Excerpt :

      Neural networks (NNs) provide an opportunity to solve this problem with their in-depth analysis and research. The neural network method is an efficient method for resolving LEs due to its parallel distribution characteristics, simplicity of electronic implementation, and adaptive ability (Lim et al., 2019; Yan et al., 2016). For resolving LEs, a gradient-based NN (GNN) solution method was reported in Wang (1992).

    • Automatic classification of snoring sounds from excitation locations based on prototypical network

      2022, Applied Acoustics
      Citation Excerpt :

      The learning strategy of the proposed prototypical network is a simple and effective model that uses the small dataset adequately. Mel-spectrogram expresses the real logarithm of short-term energy spectrum as the Mel-frequency scale [41] that characterizes the human ear perception of frequency by using a set of nonlinear spaced triangular bandpass filters. In this work, we extracted Mel-spectrograms of snoring sounds to feed into the prototypical network.

    View all citing articles on Scopus
    View full text