Design and Rationale of the My Heart Counts Cardiovascular Health Study: a Large-Scale, Fully Digital Biobank, and Randomized Trial of Large Language Model-Driven Coaching of Physical Activity

The My Heart Counts study is a large-scale, fully digital biobank and randomized trial utilizing a cross-platform smartphone application to evaluate the efficacy of large language model-driven, personalized coaching on physical activity compared to generic prompts, while simultaneously building a comprehensive dataset of multimodal health metrics for cardiovascular research.

The Gist

Imagine your heart is like a car engine. For years, we've known that if you don't drive it enough (exercise), it gets rusty and breaks down. But here's the problem: most people don't know their engine is sputtering until it's too late, and getting a personal mechanic to tell you how to fix it is expensive and hard to schedule.

This paper describes a new, high-tech project called "My Heart Counts" that tries to solve this problem using a smartphone app. Think of it as building a digital gym and a personal trainer that lives in your pocket, but with a few very special upgrades.

Here is the breakdown of what they are doing, using some simple analogies:

1. The Big Upgrade: From "One-Size-Fits-All" to "Custom Tailoring"

The original version of this app (launched in 2015) was like a generic pamphlet you pick up at a doctor's office. It gave good advice, but it only worked on iPhones, only spoke English, and the advice was written by humans who had to type it out one by one.

The new version is like a smart, self-driving car.

• It speaks your language: It works on both iPhones and Androids and speaks both English and Spanish.
• It remembers your history: It can look back at your phone's data from up to 10 years ago to see how your "engine" has been running before you even signed up.
• It's open source: The code is like a LEGO set that other scientists can take apart and rebuild to study different diseases (like diabetes or lung issues), not just heart problems.

2. The "Digital Biobank": A Time-Traveling Health Journal

Usually, when you go to the doctor, they only see you for 15 minutes. They get a snapshot of your health at that exact moment.

This app creates a continuous movie of your health instead of a snapshot. It acts as a "Digital Biobank" (a giant library of health data) by collecting:

• Passive Data: It quietly counts your steps, tracks your sleep, and measures your heart rate while you go about your day (like a silent observer).
• Active Data: It asks you to do specific challenges, like a 6-minute walk or a 12-minute run, to test your fitness.
• The "Bridge": It connects your phone data with your actual medical records (like your cholesterol levels from your doctor's lab). This is like merging your Fitbit data with your bank statement to get a full picture of your financial health.

3. The Main Experiment: The "AI Coach" vs. The "Generic Reminder"

The core of this study is a race between two types of coaches to see who can get you to walk more.

• Coach A (The Generic Reminder): This is like a sticky note on your fridge that says, "Walk 7,000 steps today!" It's the same message for everyone, every day.
• Coach B (The AI Coach): This is a super-smart personal trainer powered by Artificial Intelligence (specifically a Large Language Model, or LLM).
  • How it works: The AI doesn't just guess; it uses a psychological framework called the "Stages of Change." It figures out if you are in the "I'm thinking about starting" phase or the "I'm ready to go" phase.
  • The Magic: If you are struggling, it might say, "Hey, just 10 minutes is a great start!" If you are doing well, it might say, "You're crushing it! Let's try to beat your record." It tailors the message to you in real-time.

The Goal: They want to see if the AI coach can get people to walk more steps than the generic sticky note. If it works, it means we can give millions of people a free, personalized trainer without hiring millions of humans.

4. The "Heart Health" Dashboard: Your Personal Scoreboard

One of the coolest features is a dashboard that turns scary medical numbers into a simple score.

• The Problem: Doctors talk about "LDL cholesterol" or "systolic blood pressure," which feels abstract and boring.
• The Solution: The app takes all your data (steps, sleep, diet, blood pressure) and combines them into one "Heart Health Score."
• The Analogy: It's like a video game health bar. You can see in real-time how eating a salad or taking a walk fills up your health bar, while skipping sleep drains it. It makes the invisible risks of heart disease visible and manageable.

5. Why This Matters

The authors believe that heart disease is the biggest killer in the world, and lack of exercise is a huge reason why.

• Scalability: You can't hire a personal trainer for every person on Earth. But you can give an AI coach to everyone with a smartphone.
• Inclusion: By speaking Spanish and working on Android phones, they are trying to reach the people who are usually left out of medical studies (the "digital divide").
• The Future: If this works, it could change how we prevent heart disease. Instead of waiting until you get sick to go to the hospital, you could use your phone to keep your heart healthy every single day.

In short: This paper is about building a super-smart, inclusive, and free digital health companion that uses AI to talk to you in a way that actually motivates you to move, while quietly collecting data to help scientists cure heart disease for everyone.

Technical Summary

Based on the provided preprint, here is a detailed technical summary of the My Heart Counts Cardiovascular Health Study (Next-Generation).

1. Problem Statement

Cardiovascular disease (CVD) remains the leading cause of global mortality, with physical inactivity being a primary modifiable risk factor. Despite the availability of evidence-based interventions, traditional methods face significant barriers:

• Scalability: In-person coaching and group interventions are limited by provider time, geographic constraints, and high costs.
• Digital Divide: Previous digital health trials often exclude non-English speakers and non-iOS users, limiting generalizability and diversity.
• Static Interventions: Prior digital interventions relied on static, human-crafted text messages that lacked the personalization and adaptability required for sustained behavioral change.
• Data Fragmentation: There is a lack of integration between passive mobile health data (wearables), active patient-reported outcomes, and clinical Electronic Health Record (EHR) data.

2. Methodology

The study is a prospective, observational cohort study with an embedded randomized crossover trial, conducted entirely via a fully digital smartphone application.

A. Study Design & Platform

• Target Population: 15,000 adults (18+) from the US and UK.
• Platform: A redesigned app built on the Stanford Spezi open-source framework.
  • Cross-Platform: Initially iOS (March 2026), with Android planned for 2027.
  • Language: Available in English and Spanish (with Chinese and others in development).
• Recruitment: Fully virtual, utilizing social media and digital channels to reach diverse demographics, including those with cardiometabolic diseases (Type 2 Diabetes, Obesity) and rare conditions (Pulmonary Arterial Hypertension).

B. Data Collection Architecture

The study synthesizes three distinct data layers into a unified Digital Biobank using HL7 FHIR standards for interoperability:

1. Passive Sensor Data: High-frequency data from Apple HealthKit/SensorKit (iOS) and Google Health Connect (Android).
   • Metrics: Steps, flights climbed, heart rate (resting/active), HRV, sleep stages, VO2max estimates, raw triaxial accelerometry, PPG waveforms, and single-lead ECG.
   • Retrospective Capability: The app can ingest up to 10 years of historical health data prior to enrollment.
2. Active Assessments & Surveys:
   • Fitness Tests: 6-minute walk test (submaximal) and 12-minute Cooper run test (maximal).
   • Longitudinal Surveys: Disease-specific tools (Seattle Angina Questionnaire, KCCQ, EmPHasis-10) and general health metrics (WHO-5, GAD-7, diet, fatigue) administered at varying intervals.
3. EHR Integration:
   • US/UK Linkage: Participants can authorize access to clinical data (Lipids, HbA1c, Glucose, ICD-10 codes) via HL7 FHIR (Apple Health/Google Health Connect) or NHS number linkage.

C. The Embedded Randomized Crossover Trial

• Objective: To evaluate the efficacy of Large Language Model (LLM)-driven coaching versus generic digital nudges on physical activity.
• Intervention Arms:
  1. LLM Arm: Daily text-based prompts generated by an LLM fine-tuned on the Transtheoretical Model (Stages of Change). The AI adapts messaging based on the user's current behavioral stage.
  2. Control Arm: Generic, static reminders (e.g., "Reach 7,000 steps today").
• Protocol:
  • Participants undergo 1 week of baseline monitoring.
  • Crossover: 1:1 randomization to receive Intervention A for 1 week, followed by Intervention B for 1 week.
  • Opt-out: Participants can choose not to join the trial after onboarding.

D. Statistical Analysis

• Primary Endpoint: Change in mean daily step count between LLM and generic arms.
• Analysis Method: Linear mixed-effects models accounting for crossover design, participant random effects, and fixed effects for intervention, period, and carryover.
• Power: Target enrollment of 1,000 participants for the trial provides 90% power to detect a 500-step difference (assuming SD=3,000).

3. Key Contributions & Innovations

• Autonomous LLM Coaching: First large-scale trial utilizing an LLM grounded in behavioral psychology (Transtheoretical Model) to generate personalized, real-time health coaching at scale, replacing human-expert message crafting.
• Comprehensive Digital Biobank: Creation of a multi-modal dataset linking raw sensor data (accelerometry, PPG), patient-reported outcomes, and clinical EHR data, enabling the validation of digital biomarkers against clinical outcomes.
• Inclusive Design: Explicit focus on digital inclusion through bilingual support (English/Spanish) and cross-platform compatibility (iOS/Android), addressing the exclusion of non-English speakers in prior studies.
• Open-Source Modular Architecture: Built on the Stanford Spezi framework, the code is open-source, allowing other researchers to easily adapt the platform for different diseases (e.g., Parkinson's, Cancer) or interventions.
• Patient-Facing "Heart Health" Dashboard: A visualization tool that integrates real-time sensor data with clinical biomarkers to provide users with a dynamic, actionable score of their cardiovascular risk, demystifying "silent" risk factors.

4. Results

• Note: As this is a study design and rationale paper (preprint posted March 2026), no empirical results (e.g., specific step count increases, statistical significance of the LLM vs. control) are reported yet.
• Status: Recruitment commenced in March 2026 and is ongoing.
• Feasibility: The authors cite the success of the original My Heart Counts app (>100,000 users) as evidence that the recruitment targets (N=15,000 cohort, N=1,000 for the trial) are achievable.

5. Significance

• Scalable Preventive Care: Demonstrates a blueprint for delivering personalized, AI-driven behavioral interventions to millions of patients without the cost and logistical burden of human coaching.
• Precision Medicine: By integrating EHR data with continuous digital phenotyping, the study enables the discovery of novel digital biomarkers that predict clinical outcomes (e.g., heart failure exacerbations) earlier than traditional methods.
• Health Equity: The bilingual and cross-platform approach directly addresses the "digital divide," ensuring that cardiovascular research and interventions are accessible to underrepresented populations.
• Research Infrastructure: The open-source nature of the platform accelerates the field of digital health by providing a reusable, modular infrastructure for future randomized trials in various disease states.

In summary, the My Heart Counts Next-Generation study represents a paradigm shift from static, siloed digital health tools to a dynamic, AI-integrated, and inclusive ecosystem capable of conducting autonomous clinical trials and building a comprehensive digital biobank for cardiovascular research.