A new Apple-backed AI model trained on Apple Watch behavioral data can now predict a wide range of health conditions more accurately than traditional sensor-based approaches, according to a recently published study.
The research paper, titled "Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions," introduces a machine learning model that analyzes user behavior to flag potential health issues. Unlike earlier methods that focus on real-time sensor outputs like heart rate or blood oxygen, the new model identifies patterns in how people move, sleep, and exercise over time.
At the center of the study is a foundation model that the researchers call the Wearable Behavior Model (WBM). It analyzes high-level behavioral metrics such as step count, sleep duration, heart rate variability, and mobility, which are all calculated by the Apple Watch using on-device algorithms.
The researchers found that this approach allows the AI model to detect certain health conditions more effectively than models based solely on direct biometric data. The WBM showed particularly strong performance in identifying what the researchers called static health states, such as whether a person takes beta blockers, and transient health conditions like sleep quality or respiratory infection. For pregnancy detection, the model achieved up to 92% accuracy when combined with traditional biometric data in a hybrid approach.
Apple collected data for the model through the Heart and Movement Study, which involves more than 160,000 participants who voluntarily share data via the Apple Watch and iPhone. The foundation model was trained on over 2.5 billion hours of data and evaluated on 57 different health-related prediction tasks. It uses a time-series machine learning architecture designed to identify changes in behavior over days or weeks, allowing it to identify health conditions that unfold over time rather than instantaneously.
The researchers argue that wearable devices have now evolved to the point where they can support this kind of AI-powered analysis at scale. Whether such a model will be integrated into a user-facing feature in the future is unknown, but it goes to show that current Apple Watch hardware can go much further in terms of accurate and intelligent health analysis.
