Learning transferable human physiology from two million hours of sleep with SleepFM-2

Sleep provides a nightly window into health, capturing coordinated activity across the brain, heart, muscles and respiratory system. Here we introduce SleepFM-2, a next-generation sleep foundation model developed and evaluated on 282,511 nights of polysomnography (PSG) from 26 cohorts, including 235,865 recordings used for pretraining, together spanning over two million hours of brain activity, heart rhythm, muscle movement and breathing patterns. From a single night of sleep recording combined with age, sex and BMI, SleepFM-2 predicts 215 incident phenotypes across the phenome with a Harrell's C-index of at least 0.75 and Bonferroni-corrected P < 0.01 in two held-out test cohorts, including one health system unseen during pretraining. It outperforms both a demographics-only baseline and a strong baseline of 480 interpretable features derived from the same recordings. Beyond PSG, the SleepFM-2 frozen encoder transfers to wakeful EEG, headband and in-ear EEG, wrist PPG and even wrist accelerometry, where it outperforms end-to-end supervised training for sleep staging across six cohorts and performs comparably to foundation models pretrained directly on accelerometry for disease prediction in UK Biobank. Appendices report the remaining evaluations, including sleep scoring within the range of expert scorers and subjective sleep endpoints. Together, these results show that the rich multimodal physiology recorded during sleep can provide a transferable representation of human health across diseases, clinical tasks, sensing modalities and subjective experience.