CONVOLUTIONAL NEURAL NETWORKS FOR SLEEP STAGES CLASSIFICATION USING MULTIPLE PSG SIGNALS

Classifying sleep stages serves as an indicator for improving health, such as aiding in sleep disorder diagnosis and enhancing sleep quality. The polysomnography (PSG) is a comprehensive recording technique that captures various physiological signals during human sleep. Historically, the collection of substantial PSG datasets has been challenging. The Sleep Heart Health Study (SHHS) dataset provides a large collection of PSG data with various physiological signals. The data of sleep stage is usually imbalanced, and each physiological signal is recorded at several sampling frequencies. In this paper, we propose a convolutional neural network architecture that can accommodate variable-length inputs and facilitate the comparison of different data types. We classify the sleep stage for both balanced and imbalanced datasets. Our results demonstrate that the balanced dataset yields better performance in sleep stage classification. Furthermore, we investigate how interpolated and non-interpolated signals influence classification performance. We find no significant differences in classification accuracy between those signal types. Moreover, regarding the computational cost, the non-interpolation case is performed at a lower with less computational cost than the interpolation case.