Automatic detection of sleep apnea using eeg signal based on modeling of feature variation pattern in time-frequency domains

Abstract

Sleep apnea, a serious sleep disorder affecting a large population, causes disruptions in breathing during sleep. For diagnosis, sleep experts manually score the apnea events in overnight polysomnography, which is expensive, tedious, and prone to human error. To counter this problem, in this thesis, an automatic apnea detection scheme is proposed using single lead electroencephalography (EEG) signal, which can discriminate apnea patients and healthy subjects as well as the difficult task of classifying apnea and non-apnea events of an apnea patient. The main theme of the proposed method is to model the within-frame characteristic pattern of a statistical measure of EEG data and use the fitted model parameters as features in apnea detection. For this purpose, within a frame each sub-frame of EEG data is first decomposed and statistical measures, like entropy and log-variance are computed on each decomposed signal. For the purpose of decomposition, frequency domain band-pass filtering, variational mode decomposition and wavelet packet decomposition are considered because of their respective advantages. For a statistical measure, the resulting within-frame variation pattern for each decomposed signal is analyzed and we propose to utilize characteristic probability density function (PDF) to fit the pattern and use the model parameters as features in classifier. Various well-known PDFs are investigated and among them the Rician PDF offers very satisfactory feature qualities. For the purpose of classification, the K nearest neighbor classifier is adopted. Extensive experimentation is carried out considering three publicly available large EEG datasets and performance of the proposed method, in comparison to that of the existing methods, is found much superior in terms of sensitivity, specificity and accuracy.