Deep learning for EEG-based sleep stage classification: a review.

Deep learning architectures are now widely applied in sleep electroencephalogram (EEG) analysis. These developments have significantly advanced EEG-based sleep stage classification. This progress offers considerable benefits for sleep research and the diagnosis of neurological disorders. This work provides a systematic review of deep learning models for sleep staging. It analyzes the evolution of their architectures and input representation strategies. Additionally, it summarizes the current usage of preprocessing strategies, public datasets, and evaluation metrics. The cross-dataset accuracy comparison and stage-wise F1 distribution analysis indicate that the performance of existing models is highly dependent on the dataset, and that the evaluation system overly relies on overall accuracy. As a result, it fails to reveal performance weaknesses on minority stages and pathological data. Furthermore, this review discusses the core challenges current models face in generalization, interpretability, and clinical applicability. To address these challenges, the construction of diverse datasets, optimization of architecture design, enhancement of interpretability mechanisms, and promotion of clinical validation are also proposed as future directions. This review aims to guide the transition from algorithmic innovation to clinically reliable sleep staging research.