MoDE: A Dynamic Expert Aggregation Framework for Sleep Stage Classification with Contextual Factors

Sleep stage classification is essential for diagnosing sleep disorders, yet existing algorithms are limited by inter-subject variability and fail to leverage valuable clinical context. We propose Mixture of Domain Experts (MoDE), a novel generalizable sleep staging framework that moves beyond simple ID-based methods by using a memory bank of evolving subject prototypes and clinical records to approximate contextual factors. In MoDE, the Global-Local Expert Committee assigns an expert to each training subject, and a Dynamic Expert Aggregator combines their decisions based on the similarity between the new subject during testing and training subjects. Evaluated on the ISRUC-S1 dataset, MoDE achieves superior performance compared to state-of-the-art baselines, reaching an accuracy of 80.20%. An extensive ablation study verifies the contribution of each module. By imitating how human clinicians leverage clinical context and past case experiences to make decisions, MoDE improves generalization and brings AI-based sleep staging closer to clinical practice.