An ECG-based obstructive sleep apnea detection using the combination of time-frequency decomposition techniques

Abstract:

Obstructive sleep apnea (OSA) is a common sleep-related breathing disorder characterized by repetitive events of complete or partial cessation of airflow due to physical obstruction of the upper airway. This disease can lead to disorder at night sleep, which will have harmful consequences on the quality of patients' life. Currently, polysomnography is the gold standard for diagnosing sleep apnea, which cannot provide the expectations of a fast and economical diagnosis by analyzing several signals simultaneously. In this study, with the aim of diagnosing obstructive sleep apnea events, three automatic diagnostic algorithms based on the Electrocardiogram (ECG) signal have been presented. In the first proposed algorithm, Continuous Wavelet Wransform (CWT) and Short-Time Fourier Transform Based Synchrosqueezing Transform (FSST) have been used to obtain time-frequency representations of ECG signal and then texture features have been extracted from these time-frequency representations using the gray level co-occurrence matrix.  In the second proposed algorithm, a combined decomposition method based on Empirical Mode Decomposition and Wavelet Packet Transform (WPT) have been used and then statistical and nonlinear features have been extracted from the final sub-bands. In the third proposed algorithm, Empirical Fourier Decomposition method has been used to decompose the signal, and then a combination of nonlinear and statistical features have been extracted from the fourier intrinsic band functions. Also, several classifiers including support vector machine, random forest and weighted K-nearest neighbor are used in the proposed methods.

     In order to evaluate the proposed algorithms, the Apnea-ECG database, which contains 70 recordings of single-channel ECG signals, has been used. The obtained results showed that among the proposed algorithms, the first proposed algorithm with 90.67% accuracy in minute-by-minute classification and the second proposed algorithm with 95.71% accuracy in subject-to-subject classification provided the best performance.