Decision Tree Based Sleep Stage Estimation From Nocturnal Audio Signals

Noncontact sleep quality monitoring and early diagnosis of sleep disorders has attracted increasing attentions recently due to its advantages over traditional Polysomnography (PSG). In this study, we proposed a novel audio-based method to estimate four sleep stages. The method consists of three steps, breathe/snore detection based on Adaptive Vertical Box (AV-Box), nonlinear feature and sleep pattern feature extraction, and sleep stage classification utilizing a decision tree classifier. A seven-hour-long nocturnal audio signals is recorded for one healthy male subject with PSG in a sleep laboratory of Nanjing University of Science and Technology. Estimation performance of six types of decision tree classifiers is evaluated using k-fold cross-validation. The average prediction accuracy is 74.3%.