A Complexity Reduction Approach for Screening of Obstructive Sleep Apnea from Single Lead ECG

摘要

Sleep breathing disorders like sleep apnea are characterized by multiple episodes of breath cessation during sleep. Obstructive Sleep Apnea (OSA) is one kind of sleep apnea that occurs commonly but remains as a challenge for simpler methods of screening. Timely diagnosis and early treatment are necessary for curing this disorder which largely remains unnoticed. Sleep studies are generally conducted through Polysomnography test, which is a labor intensive and costly procedure. In recent times, extracting robust parameters from Heart Rate Variability (HRV) to detect Obstructive Sleep Apnea are emerging as an alternative to traditional sleep test. Researchers have proposed several linear and non-linear features of HRV which include time based, frequency based and non-linear features to detect the occurrence of OSA. For developing a portable, at-home screening system, a need arises to reduce HRV features for reduction of computational overheads. This paper proposes a machine learning method to screen OSA through single lead ECG recording with emphasis on reducing the dimensionality of features. Correlation based feature selection (CFS) method is used for feature dimensionality reduction and classification is attempted using three supervised classifiers such as Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and J48 decision tree. KNN performed well for the reduced feature set with an average sensitivity of 95.34%, average specificity of 94.81% and average classification accuracy of 97.3%. This paper presents the comparative study of the three classifiers combined with correlation based feature reduction done on Physionet sleepapnea database with different sample and feature sizes. The strength of the proposed method is evaluated with performance measures such as 10-fold Cross validation and conventional testing. Experiments are conducted on a maximum of 15454 one minute samples of apneic and non-apneic (normal) ECG segments. The results show that the combination of feature reduction and classification supports in the development of simpler, real time screening methods for OSA.