Detection of Obstructive Sleep Apnea Using Non-Negative Matrix Factorization-Based Feature Extraction Approach in Eigen Spectrum Domain

Obstructive sleep apnea (OSA) is the most prevalent type sleeping disorder that highly affects people’s health impairment by threatening the cardiovascular system. As an alternative to standard polysomnography (PSG) analysis, the electrocardiogram (ECG) signal has gained high interests in recent studies of OSA diagnosis. However, the existing techniques are unable to provide high detection accuracy due to their inability of capturing the hidden patterns underlying in the EGC signal. In this context, the proposed OSA detection model is designed in terms of the singular spectrum analysis (SSA) of the ECG signal. This approach investigates the Eigen spectrum of ECG data obtained from SSA. In addition, a non-negative matrix factorization (NMF) scheme is applied to reduce the dimensionality of Eigen space by eliminating its redundant characteristics. Following this, a set of new feature predictors is extracted from the activation matrices. Moreover, the present work employs an ensemble learning method by fusing diverse and accurate base classifiers to realize a high-performance diagnostic system. This study has achieved overall 94.35% accuracy for minute-by-minute OSA detection using Physionet Apnea-ECG and University College Dublin (UCD) Database. Comparative study with the existing methods shows satisfactory performances of the proposed feature extraction scheme and ensemble learning approach.