OSF: On Pre-training and Scaling of Sleep Foundation Models

Summary

OSF provides the first systematic open study of pretraining and scaling for sleep foundation models. The paper sweeps pretraining objectives, data compositions, model architectures, and scale using multi-cohort PSG data, establishing empirical scaling laws for sleep FM development. Key findings include the optimal pretraining objective mix, data curriculum effects, and how model size relates to downstream transfer across sleep staging, apnea detection, and disorder classification.