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

Researchers developed OSF, a family of sleep foundation models trained on 166,500 hours of sleep data from nine public sources. The study reveals key insights about scaling and pre-training for sleep AI models, achieving state-of-the-art performance across nine datasets for sleep and disease prediction tasks.

• →OSF models were trained on a massive dataset of 166,500 hours of sleep recordings from nine public sources.
• →Existing foundation models fail to generalize when certain data channels are missing during inference.
• →Channel-invariant feature learning is essential for effective pre-training of sleep models.
• →Scaling sample size, model capacity, and multi-source data mixture consistently improves downstream performance.
• →OSF achieves state-of-the-art results across nine datasets for sleep and disease prediction tasks.