🤖 machine learning

OpenMHC: Accelerating the Science of Wearable Foundation Models

This paper introduces OpenMHC, a comprehensive open-access dataset containing over 60 million hours of wearable health data from nearly 12,000 participants, alongside open-source implementations of foundation models and a unified benchmark, to democratize and accelerate research in wearable health AI.

Original authors: Narayan Schuetz, Yuze Bai, Lianggang Pan, Edgar Eggert, Favour Nerrise, Juan Delgado-SanMartin, Max Rosenblattl, Milana Gurbanova, Mohammad Asadi, Anders Johnson, Paul Schmiedmayer, Dennis Wang, Allan

Published 2026-07-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Narayan Schuetz, Yuze Bai, Lianggang Pan, Edgar Eggert, Favour Nerrise, Juan Delgado-SanMartin, Max Rosenblattl, Milana Gurbanova, Mohammad Asadi, Anders Johnson, Paul Schmiedmayer, Dennis Wang, Allan Lawrie, Daniel Seung Kim, Xin Liu, Akshay Paruchuri, Ehsan Adeli, Euan Ashley, Kelly W. Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your smartwatch is a tiny, tireless detective that never sleeps. Every minute of the day, it whispers secrets to your phone: how many steps you took, how fast your heart raced, how deep you slept, and even how many flights of stairs you conquered. For years, scientists have wanted to teach computers to listen to these whispers and understand our health better. They call these super-smart computers "foundation models." Think of them like a master chef who has tasted every dish in the world; once they've learned the flavors, they can invent new recipes or spot a bad ingredient instantly. But here's the problem: to become a master chef, you need a massive kitchen full of ingredients. In the world of health data, the biggest kitchens have been locked behind closed doors, owned by big tech companies, leaving researchers with only tiny, empty cupboards. Without enough data, these AI chefs can't learn the full story of how our bodies move and feel.

This is where a team of researchers from Stanford, Imperial College London, and others steps in with a massive key. They have unlocked the doors to the "My Heart Counts" study, a digital health experiment that has been running for over a decade. They are releasing a treasure trove called OpenMHC. This isn't just a small snack; it's a feast of over 67 million hours of wearable data from nearly 12,000 people. It includes everything from step counts and heart rates to sleep patterns and workout logs, collected from iPhones and Apple Watches. Along with this giant dataset, they are also sharing the "recipes" (code) and the "trained chefs" (AI models) they built to test it. Their goal is to let anyone, anywhere, try to cook up new ways to predict health issues, fill in missing data, or forecast how our bodies might behave tomorrow.

Important Note on the Ingredients: Before you start cooking, it's important to know that this kitchen has some specific quirks. The people who contributed data are mostly from the US, and the group is skewed toward white men in their late thirties. This means the "flavors" the AI learns might not taste exactly right for everyone else in the world. Also, because this was a digital study, all the health information (like whether someone has diabetes or how happy they feel) was self-reported by the users. This introduces some "noise" or uncertainty into the labels, which can make it harder for the AI to learn perfect patterns. Finally, the AI is designed to detect health traits based on the data collected at the time of the survey, not to predict future medical events like a heart attack happening next year.

The Big Reveal: What They Found

The researchers didn't just dump the data; they built a giant playground to see which AI models could actually make sense of it. They set up three main challenges:

Prediction: Can the AI guess your health traits (like if you have diabetes or how happy you are) just by looking at your watch data?
Imputation (Filling in the Blanks): Since people take their watches off to shower or charge them, the data has holes. Can the AI guess what happened during those missing minutes?
Forecasting: Can the AI predict what your activity will look like in the next 24 hours?

Here is what they discovered:

1. The "Specialized Chef" Wins, but the "Simple Cook" is Still Tough
When they tested different AI models, a model called LSM-2 (a type of foundation model trained specifically on wearable data) turned out to be the best overall chef. It was the most consistent at predicting health outcomes and filling in missing data. However, the researchers found something surprising: a much simpler, older-style model called XGBoost (which is like a very organized, rule-based calculator) came in a very close second. In fact, for some tasks, the simple model was just as good as the fancy, complex AI. This suggests that while big AI is powerful, we shouldn't assume that "more complex" always means "better." Sometimes, a well-tuned, simple tool works wonders.

2. Big Language Models (Like Chatbots) Struggled
The team also tried using massive language models (the kind that write essays and chat with you) to solve these health puzzles. They fed the models the data as text or pictures of graphs. The result? These giant chatbots generally performed worse than the specialized health models and even the simple calculators. It seems that for reading the specific "language" of heartbeats and steps, a specialized tool is much better than a general-purpose one.

3. Long-Term Memory Helps
When it came to filling in the missing data (imputation), the models that could look at a user's history over several days did much better than those that only looked at a single day. It's like trying to guess what someone is doing right now: if you only see them for one minute, you might be wrong. But if you know they usually go for a run at 6 PM, you can make a much smarter guess. The paper suggests that using a user's long-term history is a key to making these models smarter.

4. Fairness Matters
The researchers also checked if the models treated everyone equally. They found that while the fancy AI models were accurate, they sometimes made bigger mistakes for certain groups of people (like different age groups or genders) compared to the simpler models. This is a crucial reminder that as we build these health tools, we must ensure they work fairly for everyone, not just the majority.

Why This Changes the Game

Before this paper, if you wanted to build an AI to understand wearable health data, you were stuck. You couldn't get the data, and you couldn't see how the big companies trained their models. It was like trying to learn to fly a plane without ever seeing a cockpit.

OpenMHC changes that by giving everyone the cockpit, the flight manual, and the training simulator. By releasing 11,894 participants' data spanning 13 years, along with the code to run the models, the authors have democratized the field. They aren't claiming to have solved all health mysteries yet; in fact, they admit that for some conditions, the AI still struggles to beat a simple baseline due to the nature of the self-reported data. But they have provided the foundation for the entire scientific community to start cooking, experimenting, and hopefully, discovering new ways to keep us healthy. The kitchen is now open, and the ingredients are ready for anyone to use.

Technical Summary: OpenMHC: Accelerating the Science of Wearable Foundation Models

Problem Statement

While mobile and wearable devices offer unprecedented opportunities for continuous, passive health monitoring, the field of wearable health AI is hindered by a lack of open, large-scale data. Leading wearable foundation models are typically trained on proprietary datasets that are not publicly available, and their model weights and training code are rarely open-sourced. Existing public datasets are often small-scale, use specialized research-grade sensors, or lack the longitudinal depth and diversity required to train and evaluate foundation models. This data silo limits reproducibility, broader community participation, and the ability to establish standardized benchmarks for comparing models across diverse health domains.

Methodology

The authors address these gaps through the curation of OpenMHC (OpenMyHeartCounts), a comprehensive open-access dataset and benchmark derived from the My Heart Counts (MHC) study, a large-scale cardiovascular health study conducted at Stanford University.

1. The OpenMHC Dataset

Scale and Scope: The dataset comprises longitudinal records from 11,894 consenting participants collected over a 13-year calendar span (2012–2025). It includes 67 million hours of minute-level wearable data.
Data Modalities:
- Passive Sensor Data: 19 sensor channels at minute resolution, including step count, distance, flights climbed, heart rate, active energy, sleep (asleep/in-bed), and 10 workout types. Data is sourced from Apple HealthKit via iPhones and Apple Watches (and other compatible devices).
- Linked Variables: 169 self-reported and sparse HealthKit variables covering demographics, medical conditions, vitals, biomarkers, mental well-being, sleep, and lifestyle.
Demographics: Participants are primarily from the US (~90%), with smaller cohorts from the UK and Hong Kong. The cohort spans 92 distinct device models and includes users with up to 10.7 years of data history, though median participation is shorter due to typical mHealth attrition.
Preprocessing: The authors developed a rigorous pipeline to construct daily minute-level matrices ( $19 \times 1440$ ), handling device heterogeneity, missingness (non-wear time), and physiological outliers. They define specific heuristics for non-wear detection and missing data imputation.

2. The OpenMHC Benchmark

To enable standardized evaluation, the authors established a unified benchmark with three tracks:

Track 1: Predictive Tasks: 32 supervised tasks across five domains (Demographics, Medical Conditions, Vitals & Biomarkers, Mental Well-being, Sleep & Lifestyle). Tasks include binary classification, ordinal classification, and regression.
Track 2A: Generative (Imputation): Evaluation of methods to reconstruct missing sensor data. This includes Single-day Imputation (masking within a 24-hour window) and Long-context Imputation (leveraging up to 7 days of history). Masking strategies simulate real-world failures: structural (random noise, temporal slices, signal loss) and semantic (sleep gaps, workout gaps, intensity failures).
Track 2B: Generative (Forecasting): Hourly resolution forecasting of the next 24 hours of wearable data based on historical trajectories.

3. Evaluation Framework

The benchmark employs Unified Evaluation Metrics to aggregate performance across heterogeneous tasks:

Skill Score ( $S$ ): Measures relative error reduction against a fixed reference model (e.g., Linear for prediction, LOCF for imputation, Seasonal Naive for forecasting).
Average Rank: Assesses consistency across tasks.
Fairness Skill Score: Evaluates performance parity across sensitive demographic subgroups (sex and age).

Key Contributions

Release of OpenMHC: The largest and most comprehensive open-access wearable health dataset to date, enabling large-scale pretraining and evaluation of foundation models on real-world consumer data.
First Large-Scale Public Benchmark: A standardized evaluation suite covering predictive, imputation, and forecasting tasks that reflects the complexity of real-world wearable data (missingness, irregular sampling, heterogeneity).
Open-Source Implementations: The authors provide the first open-source implementations of recent wearable foundation models, specifically Apple's Wearable Health Behavior model (WBM) and Google's Large Sensor Model (LSM-2), alongside classical baselines (XGBoost, MultiRocket, GRU-D) and time-series foundation models (Chronos-2, Toto).

Results

The authors evaluated a range of models on the OpenMHC benchmark:

Predictive Tasks (Track 1):
- LSM-2 (Google's masked autoencoder) achieved the best overall performance (Average Rank $R=2.20$ , Skill Score $S=+15.1\%$ ), outperforming other foundation models and classical methods.
- XGBoost was the closest competitor ( $R=3.52$ , $S=+11.6\%$ ), particularly strong in Demographics, highlighting that well-tuned simple models remain highly competitive.
- WBM (Apple's contrastive model) underperformed in this setting, likely due to its reliance on contiguous weekly segments which reduces usable data given the sparse nature of real-world wear.
- GRU-D achieved the highest fairness score, suggesting better performance parity across demographic subgroups.
- LLMs: Gemini-family LLMs evaluated in zero-shot and few-shot settings generally underperformed the Linear baseline, failing to extract sufficient signal from text or vision representations of the data.
Imputation (Track 2A):
- LSM-2-Sparse (7-day context) achieved the best results ( $S=+64.7\%$ ), significantly outperforming statistical baselines and other neural models like BRITS and DLinear.
- The results suggest that models trained specifically on realistic missingness patterns (like LSM-2) generalize better to the structured gaps found in wearable data than models trained on random masking.
Forecasting (Track 2B):
- Chronos-2 (fine-tuned) achieved the best overall performance ( $R=3.56$ , $S=+37.6\%$ ).
- However, foundation models showed lower fairness scores compared to DLinear, indicating that accuracy gains did not necessarily translate to equitable performance across demographic groups.

Significance and Claims

The paper claims that OpenMHC democratizes wearable health AI research by providing the necessary infrastructure (data, code, benchmarks) for reproducible science. Key insights derived from the study include:

Pretraining Objectives Matter: The choice of pretraining objective (e.g., masked reconstruction vs. contrastive learning) significantly impacts downstream performance, with masked reconstruction appearing better suited to the sparse and irregular nature of wearable data.
Simplicity vs. Complexity: Strong, well-tuned simple models (like XGBoost) remain highly competitive with complex deep learning architectures, cautioning against unnecessary complexity.
Longitudinal Context: Leveraging extended personal history (e.g., 7-day windows) provides clear benefits for imputation and forecasting, pointing to a promising direction for future research.
Data Access is the Bottleneck: The field is not limited by algorithmic constraints but by data access; OpenMHC aims to remove this barrier to foster broad community participation and accelerate progress in wearable health.

The authors explicitly note limitations, including demographic skew (predominantly white, male, US-based) and the fact that outcomes are self-reported rather than clinically adjudicated, meaning the tasks are framed as detection/characterization rather than prospective clinical event prediction.