Post by Yuzhe Yang

Meet 𝗢𝗦𝗙 — a fully open benchmark and state-of-the-art family of sleep foundation models. 🌙 Recent sleep FMs have shown strong promise, but one basic question remains open: which 𝘱𝘳𝘦-𝘵𝘳𝘢𝘪𝘯𝘪𝘯𝘨 and 𝘴𝘤𝘢𝘭𝘪𝘯𝘨 choices really improve generalization in real-world settings? With 𝗢𝗦𝗙, we study this directly under 𝘤𝘰𝘩𝘰𝘳𝘵 𝘴𝘩𝘪𝘧𝘵 and 𝘮𝘪𝘴𝘴𝘪𝘯𝘨-𝘤𝘩𝘢𝘯𝘯𝘦𝘭 𝘪𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦, two major challenges for deployable sleep AI. 🏥 To make this possible, we built 𝗦𝗹𝗲𝗲𝗽𝗕𝗲𝗻𝗰𝗵, a fully open benchmark aggregated from public resources with: ⏱️ 166,500 hours of sleep recordings 🧑‍🤝‍🧑 21,000+ sleep studies 🌍 9 public datasets 💾 ~20M 30-second epochs. Across major self-supervised learning families, we identify the design choices that consistently matter for sleep FM pre-training. Our findings show that missing-channel inference can cause major drops for existing sleep FMs, but also that the right pre-training recipe can greatly improve robustness. We further find that scaling does help — in 📦 pre-training data, 🧠 model size, and 🌐 multi-source data mixture — but only when paired with the right SSL design. Guided by these insights, we build 𝗢𝗦𝗙, which beats state-of-the-arts across diverse downstream sleep and health tasks. 🚀 𝗢𝗦𝗙 offers a practical recipe for building more generalizable and deployable sleep AI! 👇 📄 Paper: https://lnkd.in/gnYewCn7 🌐 Website: https://lnkd.in/gC8rAtmA 💻 Code: https://lnkd.in/g4NZmDEf 🤗 Models: https://lnkd.in/gwatigQ7 Great work led by Zitao Shuai, Zongzhe Xu, David Yang, with collaborator Wei Wang! UCLA UCLA Computer Science Computational Medicine Department UCLA Henry Samueli School of Engineering and Applied Science #AI #Sleep #HealthAI #MultimodalAI #FoundationModels