Transforming wearable data into personal health insights using large language model agents

Area of Science:

• Artificial Intelligence
• Digital Health
• Behavioral Science

Background:

• Standard Large Language Models (LLMs) struggle with complex numerical reasoning for personalized health insights from wearable devices.
• Tool-based approaches, particularly code generation, are needed for effective data analysis.

Purpose of the Study:

• Introduce the Personal Health Insights Agent (PHIA), an LLM agent designed for scalable analysis of behavioral health data.
• Evaluate PHIA's performance against a code generation baseline using novel benchmark datasets.

Main Methods:

• Developed PHIA, a system combining multistep reasoning, code generation, and information retrieval.
• Created two benchmark datasets with over 4000 health insights questions.
• Conducted a 650-hour human expert evaluation comparing PHIA to a baseline.

Main Results:

• PHIA achieved 84% accuracy on objective, numerical health questions.
• For open-ended questions, PHIA earned 83% favorable ratings and was twice as likely to receive the highest quality rating.
• PHIA significantly outperformed the code generation baseline.

Conclusions:

• PHIA demonstrates a powerful approach for analyzing complex behavioral health data from wearable trackers.
• This technology can empower individuals to understand their health data, promoting accessible, personalized wellness.
• The findings suggest a new era of data-driven personal health management.