Title: Advancing General Intelligence and Interfaces for Wearable Health Data

Abstract:

Although wearable sensors are increasingly common and capable of gathering extensive behavioral and physiological data, translating these signals into individualized health insights remains a significant hurdle. This difficulty stems from the complexity of mapping low-level sensor readings to higher-level health states, a challenge exacerbated by substantial phenotypic diversity and variations in personal baselines, physiology, and lifestyle. Furthermore, acquiring wearable data alongside annotated health outcomes is both costly and labor-intensive, while retrospective annotation is largely impractical. Consequently, there is a notable shortage of high-quality, labeled datasets.

To address these challenges, we introduce a foundation model for wearable health that has been pretrained on over one trillion minutes of unlabeled sensor data from a cohort of five million participants. Our results indicate that simultaneously increasing model capacity and pretraining data volume yields consistent performance gains across 35 diverse health prediction tasks. These tasks cover cardiovascular and metabolic health, sleep quality, mental well-being, lifestyle habits, and demographic attributes. We observe that this large-scale population representation enables label-efficient few-shot learning and supports generative capabilities for reliable daily metric estimation.

To maximize the utility of these learned representations, we employed a suite of Large Language Model (LLM) agents to automatically explore the landscape of downstream predictive heads built upon the model embeddings. This approach resulted in broad performance enhancements that correlate with increased LLM capacity. Finally, we demonstrate how embedding these predictors into a Personal Health Agent facilitates responses that are more contextually aware, relevant, and safe. The efficacy of this system was validated through 1,860 ratings provided by a group of clinicians.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC