Automatic Snore and Breathing Sound Classification based on the Signal Envelope

Abstract

Snore sound analysis has been recently proposed for diagnosing obstructive sleep apnea (OSA). The snore sounds were typically recorded simultaneously with fullnight polysomnography (PSG) resulting in many hours worth of data. Most of the automated snore-analysis techniques require reliable methods to segment out snore sounds from a recording. In this study we focus on this problem and propose a fully automated method to identify breathing and snore sounds. The method is based on the novel feature ‘positive/negative amplitude ratio (PNAR)’ to measure the shape of sound signal. The classification performance of the proposed method was evaluated using receiver operating characteristic (ROC) analysis and was compared with that of traditional zero crossing rate (ZCR). We show that PNAR-based method clearly outperforms ZCR-based method for snore-breathing classification. In particular, PNAR provided better performance for classifying apneic snore and breathing. The new feature PNAR has the potential to contribute towards developing snore-sound based non-contact technology to diagnosis OSA.