Researchers trained foundation model on 3M days of Apple Watch data
a day ago
- #AI in healthcare
- #wearable technology
- #machine learning
- Researchers from MIT and Empirical Health used 3 million person-days of Apple Watch data to develop a foundation model for predicting medical conditions.
- The model is based on Yann LeCun's Joint-Embedding Predictive Architecture (JEPA), which focuses on inferring missing data rather than reconstructing exact values.
- The study, titled 'JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare,' was accepted to a NeurIPS workshop.
- JETS adapts JEPA for irregular multivariate time-series data, such as wearable data with gaps in heart rate, sleep, and activity measurements.
- The dataset included 16,522 individuals, with 63 distinct time-series metrics across five domains: cardiovascular health, respiratory health, sleep, physical activity, and general statistics.
- Only 15% of participants had labeled medical histories, making 85% of the data unusable for traditional supervised learning. JETS used self-supervised pre-training followed by fine-tuning.
- JETS achieved high AUROC scores for conditions like high blood pressure (86.8%), atrial flutter (70.5%), chronic fatigue syndrome (81%), and sick sinus syndrome (86.8%).
- The study highlights the potential of novel models to extract insights from incomplete or irregular wearable data, even when devices aren't worn continuously.