Physiological recordings — polysomnography, wrist accelerometry, and beyond — share underlying structure across cohorts, devices, and clinical tasks. This theme develops and evaluates self-supervised and foundation-model approaches that learn reusable representations across physiological modalities and downstream health outcomes, currently organized around two modalities.
Foundation models for polysomnography apply this at very large scale to multimodal sleep recordings, learning representations transferable across disease and mortality prediction tasks. Foundation models for wrist accelerometry take the same approach to wearable data, with representations that transfer across human activity recognition, sleep staging, apnea evaluation, and future health-risk prediction — see also Wearables for scalable health assessment.
