Effectiveness of a Convolutional Neural Network in Sleep Arousal Classification Using Multiple Physiological Signals

Abstract

The 2018 PhysioNet Challenge utilizes 13 physiological signals collected during polysomnographic sleep studies to classify explicitly-defined arousal regions. The goal is to assign a probability of arousal at each sample for each test subject. Automatic detection of non-apnea arousals may help us better understand various causes of sleep disturbance and advance sleep arousal analysis. Neural networks possess powerful feature-learning abilities to gain insights from complex datasets. Our approach is based on a deep convolutional neural network (CNN), which we trained with normalization, pooling, activation and dropout techniques in Python using Keras on top of Tensorflow. The CNN was trained on 737 patients' sleep data and validated on 185 patients' sleep data. Our model obtained AUROC performance score of 0.514293+/-0.054509 and AUPRC performance score of 0.501947+/-0.063199. In this paper, we discuss the strengths and limitations of our CNN in sleep arousal classification using a variety of physiological signals. We also present some possible directions for future work.