AI beyond LLMs: a wearable foundation model based on JEPA (www.empirical.health)

A new foundation model for health-related AI, named JETS (Joint Embedding for Time Series), was unveiled at the NeurIPS workshop, demonstrating its capability to analyze multivariate, irregularly sampled wearable data combined with blood-based biomarkers. This model builds on the Joint Embedding Predictive Architecture (JEPA) framework and has been pre-trained on a substantial dataset comprising 3 million de-identified person-days of data, achieving impressive accuracy in predicting various medical conditions and biomarker levels, surpassing existing baseline models. The significance of JETS lies in its demonstration that high-quality AI advancements can originate from small teams rather than only large organizations, challenging the notion that substantial funding and resources are essential for impactful contributions to AI/ML. Furthermore, JETS extends previous research by handling complex, irregularly sampled time-series data—an area often overlooked—suggesting a promising direction for future health AI, particularly in training models with physiological ground truth. This work marks a crucial step in leveraging wearable technology for enhanced healthcare analysis and opens avenues for future research in health-focused machine learning applications.