The rapid progress of large language models (LLMs) and wearable device data is reshaping personal health management. Tailoring health support to individual needs is crucial, as queries like, “How many hours have I slept this month?” differ from open-ended questions, “How can I enhance my sleep quality?” To tackle these complexities, we introduce the Personal Health Agent (PHA), a human-centered framework that utilizes multimodal data for personalized health guidance. The PHA employs a multi-agent architecture focusing on three roles: data science, domain expertise, and health coaching, each managed by specialized sub-agents. We evaluated the system using real-world data from ~1,200 users who consented to share their Fitbit data, health questionnaires, and blood test results. Through extensive automated and human evaluations across 10 benchmark tasks, our research represents a comprehensive assessment of a health agent, laying the groundwork for accessible personal health solutions. This conceptual framework is for research purposes only and not a direct product.
Source link
