Reinforcement Learning Based Prediction and Classification of Sleep Disorder

The application of machine learning techniques for the classification of sleep stages based on electroencephalogram (EEG) data has experienced substantial advancements in recent years (SSC). Nonetheless, the efficacy of these models is largely contingent upon the availability of extensive labeled datasets for training, which limits their applicability in real-world scenarios. In such contexts, sleep research facilities can generate significant volumes of data; however, the process of labeling this data can be both resource-intensive and time-consuming. A prominent strategy to mitigate the challenge of insufficient labels is the adoption of the Reinforcement Learning (RIL) paradigm, which has seen a surge in interest recently. This research aims to evaluate the extent to which RIL can enhance the performance of existing SSC models within a limited-label environment. Conduct a comprehensive analysis across three SSC datasets and find that competitive results can be attained by optimizing pretrained SSC models with merely 5% of labeled data, in contrast to traditional supervised training utilizing complete labeling. Furthermore, self-supervised pretraining enhances the robustness of SSC models against challenges associated with domain shifts and data imbalance. In the context of limited labeled data, this study explores the potential of self-supervised pretraining to improve the performance of contemporary sleep stage classification models.