Researchers have created a “personal health agent” (PHE) powered by artificial intelligence that has the ability to give bespoke health advice.
The prototype combines personal health data from wearables, medical information, and broader clinical knowledge and is the most comprehensively evaluated system of its kind to date. The study was published in Nature this week.
It demonstrates how AI tools could one day be used to improve long-term health outcomes in a way tailored to each individual.
“These results provide a basis for further research of PHAs for daily health and well-being,” reported A. Ali Heydari, PhD, from Google Research in Mountain View, California, and co-workers.
Rapid advancements in large language models (LLMs) have driven transformative changes in health and a new generation of health agents.
Yet the ability of these LLMs to address the diverse needs of individuals in everyday health settings needs further exploration.
Initial experiments by the researchers confirmed that a monolithic LLM architecture struggled to balance the competing objectives of different categories of health and wellness needs.
The team therefore deployed a multi-agent architecture powered by state-of-the-art LLMs to create an agent that could address diverse consumer health needs.
The modular system comprises three specialized AI agents. The first functions as a data scientist to analyze medical data and information gleaned through wearables such as smart watches that record sleep activity and provide other longitudinal physiological data.
The second adds clinical knowledge and published scientific research to this, generating personalized insights and acting as a domain expert agent.
The third operates as a health coach to set goals, apply psychological strategies, and track progress.
A central AI “orchestrator” supervises the three agents, directing tasks to the appropriate specialist.
The framework is designed to be compatible with other LLMs.
The researchers evaluated the sub-agents and full system across 11 benchmark tasks using automated and human evaluation. Collectively, it involved more than 7000 annotations and 1350 hours of effort from health experts and end users.
The system generally performed well against a base model of the Gemini 2.0 family, and was significantly better in analysis plan quality and code generation pass rates.
However, the computer code written by the data-scientist agent to analyze health data was not perfect, and was unable to pass automated tests for its correctness at least a fifth of the time.
There may also be privacy and security concerns as the agents access sensitive, real-world health data, noted Emma Pierson, PhD, from the University of California, Berkeley, in an accompanying News & Views article.
“AI systems that provide health advice have great potential to improve the accessibility of health information, including for underserved populations who have historically lacked ready access to medical advice,” she continued.
“Rigorous assessments of short-term safety and effectiveness, such as those carried out by Heydari and colleagues, followed by RCTs [randomized controlled trials] to assess effects on long-term health outcomes, will measure whether AI systems can truly achieve their potential.”
