AI Research

Towards a General Intelligence and Interface for We... | AI Research

Key Takeaways

Wearable devices collect vast amounts of physiological and behavioral data, but turning these raw signals into meaningful health insights remains a significa...
While ubiquitous wearable sensors capture a wealth of behavioral and physiological information, effectively transforming these signals into personalized health insights is challenging.
To overcome these limitations, we propose a foundation model for wearable health that is pretrained on more than one trillion minutes of unlabeled sensor signals drawn from a large cohort of five million participants.
We find that this population scale representation unlocks label-efficient few-shot learning and generative capabilities for robust daily metric estimation.
Wearable devices collect vast amounts of physiological and behavioral data, but turning these raw signals into meaningful health insights remains a significant challenge.

Paper AbstractExpand

While ubiquitous wearable sensors capture a wealth of behavioral and physiological information, effectively transforming these signals into personalized health insights is challenging. Specifically, converting low-level sensor data into representations capable of characterizing higher-level states is difficult due to high phenotypic diversity and variation in individual baseline health, physiology, and lifestyle factors. Moreover, collecting wearable data paired with health outcome annotations is laborious and expensive, and retrospective annotation remains practically unfeasible, contributing to a scarcity of data with high-quality labels. To overcome these limitations, we propose a foundation model for wearable health that is pretrained on more than one trillion minutes of unlabeled sensor signals drawn from a large cohort of five million participants. We demonstrate that the joint scaling of model capacity and pretraining data volume leads to systematic improvements in performance, as evaluated on a diverse set of 35 health prediction tasks, spanning cardiovascular, metabolic, sleep, and mental health, as well as lifestyle choices and demographic factors. We find that this population scale representation unlocks label-efficient few-shot learning and generative capabilities for robust daily metric estimation. To further leverage this learned representation, we deploy a classroom of LLM agents to autonomously search the space of downstream predictive heads built on the model embeddings, showing broad performance improvements that increase with LLM model capacity. Finally, we show how integrating these downstream predictors into a Personal Health Agent can support model responses that are more relevant, contextually aware, and safe, and we validate this via 1,860 ratings from a cohort of clinicians.

Wearable devices collect vast amounts of physiological and behavioral data, but turning these raw signals into meaningful health insights remains a significant challenge. This paper introduces a foundation model designed to bridge the gap between low-level sensor data and high-level health understanding. By training on an unprecedented scale of data, the researchers aim to create a more versatile and intelligent interface for personal health monitoring.

The Challenge of Wearable Data

Translating sensor data into health insights is difficult because individual health, physiology, and lifestyle factors vary significantly from person to person. Furthermore, training models to recognize specific health outcomes is hindered by a lack of high-quality, labeled data. Collecting these labels is expensive and time-consuming, making it difficult to build accurate models that work across diverse populations.

A Massive Foundation Model

To address these hurdles, the researchers developed a foundation model pretrained on over one trillion minutes of unlabeled sensor data from five million participants. By scaling both the model's capacity and the volume of training data, the team achieved consistent performance improvements across 35 different health tasks. These tasks cover a wide range of areas, including cardiovascular health, metabolic function, sleep quality, mental health, and lifestyle habits.

Autonomous Optimization with LLM Agents

A unique aspect of this research is the use of a "classroom" of Large Language Model (LLM) agents. These agents are tasked with autonomously searching for the best ways to interpret the model’s internal representations for specific health predictions. This approach allows the system to improve its performance on downstream tasks automatically, with the researchers noting that performance gains increase as the capacity of the LLM agents grows.

Enhancing Personal Health Agents

The researchers integrated these predictive capabilities into a Personal Health Agent to provide more contextually aware and relevant health information. By grounding the agent's responses in the model's data-driven insights, the system can offer safer and more personalized guidance. This integration was validated through 1,860 evaluations conducted by a cohort of clinicians, confirming the practical utility of the approach in a real-world health context.

Comments (0)

No comments yet

Be the first to share your thoughts!