Stanford Launches AI-Coached Heart Health Trial Targeting 15,000 Adults
A large-scale digital trial tests whether LLM-generated, personalized coaching prompts can boost daily physical activity better than generic messages.
Summary
Stanford researchers have launched a next-generation smartphone study called My Heart Counts, enrolling up to 15,000 adults in the US and UK. The study builds a digital biobank by collecting passive health data — steps, heart rate, sleep, ECGs — alongside electronic health records. Embedded within it is a randomized crossover trial testing whether AI-generated, personalized coaching messages based on the Transtheoretical Model of Change can increase daily step counts more effectively than generic prompts. The app runs on iOS now, with Android planned for 2027, and supports both English and Spanish. If LLM-driven coaching proves effective at scale, this model could reshape how preventive cardiovascular care is delivered — cheaply, automatically, and without needing human health coaches for every patient.
Detailed Summary
Cardiovascular disease remains the world's leading cause of death and disability, yet scalable, personalized behavioral interventions have been difficult to deliver without substantial human resources. Digital health tools offer a promising path forward, but earlier smartphone trials were constrained by platform exclusivity and reliance on human experts to craft individualized messages. The My Heart Counts 2.0 study was designed to address both limitations head-on.
This prospective, observational cohort study with an embedded randomized crossover trial will enroll up to 15,000 adults across the United States and United Kingdom via a redesigned smartphone app built on the open-source Stanford Spezi framework. The app collects a rich digital biobank: passive sensor data including steps, heart rate, sleep, and accelerometry; active fitness tasks such as the 6-minute walk test and Cooper 12-minute run; ECGs; longitudinal surveys; and electronic health record data integrated via HL7 FHIR protocols.
The core trial tests whether coaching prompts autonomously generated by a large language model (LLM) — grounded in the Transtheoretical Model of Change and personalized to each user — can increase daily step counts more than generic text prompts. The crossover design allows each participant to serve as their own control, increasing statistical power. Secondary endpoints include active minutes and calorie expenditure; exploratory endpoints cover cardiorespiratory fitness changes, resting heart rate, heart rate variability, and sleep quality.
For clinicians and health-conscious adults alike, the implications are significant. If LLM coaching outperforms generic messaging at scale, it would validate a low-cost, fully automated model for preventive cardiovascular care — one that could reach populations who lack access to health coaches or behavioral counselors.
Caveats remain. The study is still recruiting, so no outcome data yet exist. App-based trials face attrition and self-selection bias. The summary here is based on the published protocol abstract alone, and efficacy results are pending.
Key Findings
- Trial targets 15,000 adults across the US and UK using a fully digital, smartphone-based platform.
- LLM-generated personalized coaching prompts are tested head-to-head against generic activity messages.
- Primary endpoint is change in daily step count; secondary endpoints include active minutes and calorie burn.
- Digital biobank integrates ECG, accelerometry, fitness tests, surveys, and EHR data via FHIR protocols.
- App supports English and Spanish, broadening reach to underserved and Latino cardiovascular populations.
Methodology
Prospective observational cohort with an embedded randomized crossover trial; each participant experiences both LLM-driven and generic coaching arms. Primary analysis uses mixed-effects models to compare daily step counts between arms. Recruitment began March 2026 and is ongoing.
Study Limitations
No efficacy data are available yet — this paper describes study design and rationale only. App-based trials are prone to self-selection bias and high attrition rates. This summary is based on the published abstract only, as the full text was not accessible.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: