Researchers use Apple Watch to train a disease-detection AI-even when data is incomplete

Researchers from MIT and Empirical Health have developed an AI model called JETS, which predicts medical conditions using data collected from Apple Watch devices. The team trained this model on an extensive dataset spanning 3 million person-days of smartwatch and fitness tracker data, capturing real-world, irregular, and incomplete health information. Importantly, JETS is designed to handle such incomplete and irregular datasets effectively, a common challenge in wearable health monitoring.

JETS utilizes a machine learning architecture adapted from the Joint-Embedding Predictive Architecture (JEPA), enabling it to analyze multivariate time-series data from wearables. The model demonstrated strong predictive performance, surpassing earlier baseline methods. For example, it achieved an impressive area under the receiver operating characteristic curve (AUROC) of 86.8% for detecting high blood pressure and 70.5% for atrial fibrillation.

This innovation holds significant promise for advancing wearable health technology by making disease detection more accessible and continuous outside traditional clinical settings. By leveraging the vast amounts of data collected passively by consumer devices like Apple Watch, JETS offers a scalable approach to early disease prediction, potentially improving health outcomes through timely intervention. The model’s ability to work well with imperfect real-world data addresses a key bottleneck in applying AI to wearable health monitoring, marking a notable step forward in personalized health analytics.