Suitable Feature Selection for OSA Classification Based on Snoring Sounds

Abstract

This research investigates suitable features of snoring sound records for Obstructive Sleep Apnea (OSA) classification. The total of 49 sound records from Thai OSA subjects was acquired with overnight multichannel polysomnography assessment. Based on apnea hypopnea index, 24 subjects were identified as severe OSA, 10 as moderate OSA, 7 as mild OSA, and 8 as non-OSA. For each record, silence sounds were removed. Breathing and snoring sounds were separated by the Fuzzy C-Mean clustering of their mean absolute amplitudes. 33 common sound features were extracted from snoring sound segments of each OSA subject. The total variation norm of each feature was computed and comprised a feature set of an OSA subject. Top 10 features with the highest weight average of Pearson‘s correlation coefficients were used to select top 10 features for further investigation on their classification performance. Based on linear discriminant analysis with the leaveone-out cross validation, most of top 10 features provided high accuracy for non-OSA/OSA classification with a single feature. On the four-class OSA classification, the accuracy using a single feature significantly dropped. The combination of SSI and MC improved the classification results to 87.8% accuracy.