SleepTight: Identifying Sleep Arousals Using Inter and Intra-Relation of Multimodal Signals

Abstract

Sleep arousal directly affects the quality of sleep. PhysioNet Challenge 2018 aims to correctly identify designated target arousal (non-apnea arousal) and non-arousal regions from simultaneously recorded multiple biomedical signals. Our contribution lies in a feature extraction algorithm that extracts generic and domain-specific features from different biomedical signals available in the challenge provided dataset to form a composite feature vector. 50 most significant features are selected based on Minimum Redundancy Maximum Relevance scores for final classification using multiple unbiased Random Forests. The approach is designed to produce a single label for a 20-second segment containing all channels, followed by smoothing the label time-series per subject. Our algorithm yields the median Area Under Precision-Recall Curve (AUPRC) as 0.29 on 5-fold cross-validation on the training dataset. The same value of AUPRC is maintained for the test dataset as well, thereby emphasizing the stability of the proposed algorithm. This method secured the global rank of 8 during the official phase of the challenge.