In short :

This study introduces a multimodal foundation model trained on large-scale polysomnography (PSG) data, designed to jointly leverage multiple physiological signals (EEG, EOG, EMG, respiratory signals) to learn rich representations of sleep. The model is pretrained in a self-supervised manner, enabling it to capture underlying physiological patterns without relying on extensive manual annotations. It can then be adapted to a wide range of downstream clinical tasks, including sleep staging and disease prediction. Bioserenity contributed U.S.-based PSG datasets that played a key role in the pretraining phase. In addition, carefully curated subsets built from detailed patient questionnaires by Umaer Hanif (co-authors) enabled targeted fine-tuning and evaluation across specific clinical endpoints. Results show that this foundation model approach outperforms task-specific baselines across multiple settings, particularly in low-label regimes. More broadly, the study demonstrates how multimodal sleep data can be leveraged to uncover clinically meaningful signatures, paving the way for more scalable and generalizable diagnostic tools.