Apple Watch Data Fuels AI Disease Detection Research

Researchers from MIT and Empirical Health have utilized 3 million person-days of Apple Watch data to develop an AI model for predicting medical conditions. This study, titled JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare, was accepted to a workshop at NeurIPS.

It adapts the Joint-Embedding Predictive Architecture, or JEPA, which was proposed by Yann LeCun, to handle irregular multivariate time-series data. The dataset includes data from 16,522 individuals, with 63 distinct time series metrics recorded, spanning five physiological and behavioral domains.

Notably, only 15% of participants had labeled medical histories for evaluation, which is typically a limitation for traditional supervised learning approaches. Instead, the JETS model used self-supervised pre-training on the entire dataset and fine-tuned on the labeled subset.

The process involved creating triplets of data corresponding to day, value, and metric type, converting observations into tokens, and predicting missing data embeddings. JETS was evaluated against baseline models, achieving an AUROC of 86.8% for high blood pressure and 70.5% for atrial flutter, among other conditions.

The study highlights the potential of leveraging wearable technology data, such as that from the Apple Watch, even when measurements are inconsistently recorded.