MMSleepNet: A Μultimodal Model for Automatic Sleep Staging

Abstract

Automatic sleep stage classification is crucial for assessing sleep quality and diagnosing sleep disorders. While numerous approaches have been developed for this purpose, many rely solely on single-channel electroencephalogram (EEG) signals or consider inputs from a single domain. Polysomnography (PSG) offers a more comprehensive solution by providing multiple channels of signal recording, enabling models to extract and integrate information from diverse channels for enhanced classification performance. Furthermore, valuable insights can be derived from both the time and frequency domain representations of PSG signals. This study introduces MMSleepNet, a deep learning model that adeptly extracts and fuses information from multiple PSG channels across both the time and frequency domains to facilitate automatic sleep stage scoring. Recognizing the potential of sleep labs to generate vast amounts of unlabeled data for sleep scoring and the inherent difficulty of labeling this data, the study explores the efficacy of various self-supervised learning algorithms to address label scarcity and enhance the proposed model's performance.