Unsupervised classification of respiratory sound signal into snore/no-snore classes

In this study, an automatic and online snore detection algorithm is proposed. The respiratory sound signals were recorded simultaneously with Polysomnography (PSG) data during sleep from 20 patients (10 simple snorers and 10 OSA patients). The sound signals were recorded by two tracheal and ambient microphones. The potential snoring episodes were identified using Vertical Box (V-Box) algorithm. The normalized 500Hz sub-band energy features of each episode were calculated. Principal component analysis (PCA) was applied to a 10-dimensional feature space to reduce it to a new 2-dimensional feature space. An unsupervised K-means clustering algorithm was then deployed to label the sound episodes as either snore or no-snore class. The performance of the algorithm was evaluated using manual annotation of the sound signals. The overall accuracy of the system was found to be 98.2% for the tracheal recordings and 95.5% for the sounds recorded by the ambient microphone.