Reproduce “OpenMHC: Accelerating the Science of Wearable Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Narayan Schuetz, Yuze Bai, Lianggang Pan, Edgar Eggert, Favour Nerrise, Juan Delgado-SanMartin, Max Rosenblattl, Milana Gurbanova and 11 more
The paper introduces OpenMHC, a shared benchmark for wearable models. It combines a HealthKit-derived dataset for qualified researchers and model implementations with 11,894 participants, 2.82 million participant-days, 67 million wear-hours, and 32 prediction tasks. On self-reported outcomes, LSM-2 achieves +15.1% skill versus Linear, compared with XGBoost’s +11.6%. The results do not establish clinical validity and are limited by cohort bias.
Key insights
- Self-supervised pretraining for wearable data is sensitive to design choices, with minute-level input resolution and masked reconstruction objectives (like LSM-2) proving more effective than hourly data or contrastive learning (like WBM).
- Strong classical models, such as XGBoost, remain highly competitive in predictive tasks, often outperforming more complex neural architectures.
- Leveraging extended longitudinal personal history significantly enhances performance in generative tasks like data imputation, highlighting the value of multi-day context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.