S4Sleep: Elucidating the design space of deep-learning-based sleep stage classification models

Machine-learning-based automatic sleep stage scoring is a promising approach to enhance the time-consuming manual annotation process of polysomnography recordings. Although numerous algorithms have been proposed for this purpose, systematic exploration of architectural design decisions remains limited. This study conducts a comprehensive investigation into these design choices within the broad category of encoder-predictor architectures. The methodology identifies robust architectures applicable to both time series and spectrogram input representations, both of which leverage structured state space models as integral components. Without further hyperparameter adjustments, the proposed models S4Sleep(spec) and S4Sleep(ts) consistently surpass all existing approaches on the most commonly used benchmark datasets: Sleep EDF, the Montreal Archive of Sleep Studies, and, most notably, the extensive Sleep Heart Health Study dataset. The architectural insights derived from this research, along with the refined methodology for architecture search demonstrated herein, are expected to not only advance future research in sleep staging but also be beneficial for other time series annotation tasks.