Most discriminative atom selection for apnea-hypopnea events detection

Abstract

The sleep apnea-hypopnea syndrome is characterized by repetitive episodes of upper airway obstruction that occur while sleeping, usually associated with a reduction in blood oxygen saturation (SaO2). This work presents a novel most discriminative atom selection method to predict the occurrence of apnea-hypopnea (AH) events. First two types of dictionaries (one using class information and the other without it) are estimated, then a greedy pursuit algorithm is used in order to obtain the activation coefficients. The SHHS polysomnography database which includes nearly 1000 polysomnograms, is used for training and testing. A subset of the most discriminative coefficients is then selected for each dictionary, training a pattern recognition neural network to detect the AH events. Finally these events from a test set of 64 studies with different grades of illness are detected. Correlation coefficients of 0.90 and 0.74 are obtained for the dictionaries trained with and without class information, respectively.