AI Health Agent Reads Wearable and Medical Record Data to Coach Users

The paper, by Heydari and colleagues, is a peer-reviewed development and evaluation study of a software system. It is not a trial that followed people to see whether using the tool made them healthier.

The team built what they call a personal health agent. It uses large language models, the same kind of technology behind popular chatbots. According to the authors, it reasons over "multimodal data" from consumer wellness devices and medical records. In plain terms, that means it can look at several kinds of information at once, such as wearable readings and clinical records.

To decide what the tool should do, the researchers analysed web search and health forum queries. They also ran user-centred research with end users and health experts. From this they identified four main ways people want help: general health knowledge, interpreting their personal data, wellness advice, and symptom assessment.

Those four needs shaped three specialist sub-agents. The first is a data science agent. It analyses long-term wearable and medical data and uses population-level statistics to put numbers in context. The second is a health domain expert agent. It combines personal and contextual data with health knowledge to produce personalised insights. The third is a health coach agent. It pulls the insights together, supports multi-turn conversation and goal setting, applies specified psychological strategies and tracks progress. The authors describe the full system as an open-source multiagent framework.

The evaluation was large. The authors tested the sub-agents and the full system across 11 benchmark tasks, using both automated and human evaluation. The work involved more than 7,000 annotations and 1,350 hours of effort from health experts and end users. The authors call it the most comprehensive evaluation of a personal health agent so far. That is their own description.

It reports no accuracy figures, no comparison scores and no information on error rates or safety problems. It does not show that the agent improves health outcomes. The study describes the work as a foundation for improving access to personalised health support, which is a statement of aim, not a demonstrated result.

The setting helps explain the interest. A Nature commentary on the study says an estimated one-third of US adults turn to chatbots for health questions. It also says it has been suggested that about 230 million people ask ChatGPT health and wellness questions every week. Those figures come from the commentary, not from the study.

A separate 2025 letter in npj Digital Medicine, by Agrawal and colleagues, offers a caution that applies across the field. It reports that only 5% of recent evaluations of language models in medicine used real electronic health record data. It also says automated scoring methods often track human judgement only loosely. The letter does not assess this agent. It argues that tests with real-world data and human review matter, and that evidence for health tools should be reported clearly.

For now, the study shows that a multi-part AI coach can be built and tested at scale. What to watch next is whether independent groups and real-world studies report how the agent performs and whether it changes health outcomes.

Home Light Therapy Beat Placebo on Sleep and Memory Scores in Mild Cognitive Impairment Exercising Outdoors in Nature Boosted Older Adults' Mental Health More Than Indoors AI Health Agent Reads Wearable and Medical Record Data to Coach Users HIIT Raised VO2 Max Most Consistently in Inactive People, Meta-Analysis Finds Too Much RNA Can Damage Mitochondria and Drain Cell Energy Ultra-Processed Calories Linked to Worse Daily Blood Sugar in Type 2 Diabetes Whole-Food Diet Changes Added About 12 Grams Daily Protein in Older Adults Breathing Practices May Calm Nerves and Brain, but Some People Need Caution Gut Microbe Transplants Eased Nicotine Withdrawal and Anxiety in Mice Vitamin B3 Linked to Younger Muscle Epigenetic Readings; Interval Training Mixed Sleeping an Hour Longer in Later Life Linked to Doubled Dementia Risk Cochlear Implant Users Had About 40% Lower Dementia Risk Over Time Only One in Five Women Start Hormone Therapy After Menopause Diagnosis Leafy Greens Linked to 11% Lower Heart Disease Risk, Meta-Analysis Finds Nudge-Based Activity Programs Linked to Better Frailty Scores and Mood in Older Adults New 'Pasta' Aging Clock Finds Drug Candidates That Rejuvenate Human Cells Recreational Runners Hit Higher VO2 Max With Less Work From Their Hearts Review Examines Gut Microbes in Preclinical Alzheimer's, With Timing Unresolved Not All Ultra-Processed Foods Are Equal, Study of 13,025 Products Finds Computer Model Finds No Single Cause of Aging Muscle's Protein Resistance Thinner Vagus Nerve Linked to Faster Heart Rate in Older Adults Five-Cohort Study Ties Consolidated Sleep and Brain Wave Coupling to Cognition Study of 18 Epigenetic Clocks Finds Most Give Unstable Readings Over Time Swinging Blood Pressure Readings Linked to Faster Cognitive Decline in Trials Aerobic Exercise Boosted Thinking Skills Most in Postmenopausal Women Protein MTR4 Falls in Aging Blood Vessel Lining Cells, Study Finds Gene Therapy Targeting a Mitochondrial Enzyme Slowed Ovarian Aging in Mice Review Proposes Gut-Brain-Fat Loop to Explain Ultra-Processed Food and Obesity Review Weighs Whether Years of Intense Endurance Training Harm Athletes' Arteries