Foundation model for health prediction using Apple Watch data (9to5mac.com)

A groundbreaking study by researchers from MIT and Empirical Health has led to the development of a foundation model called JETS, which utilizes 3 million days of Apple Watch data to predict various medical conditions with notable accuracy. This model applies the Joint-Embedding Predictive Architecture (JEPA) concept, initially proposed by Meta’s Chief AI Scientist Yann LeCun, to healthcare by effectively managing irregular multivariate time-series data. Through self-supervised learning, JETS leverages a vast dataset of physiological and behavioral metrics from over 16,500 individuals, converting incomplete observations into meaningful insights, and overcoming traditional supervised learning limitations. The significance of JETS lies in its ability to generate predictions even from sparse data where only 15% of participants had labeled medical histories. It achieved impressive AUROC scores—86.8% for high blood pressure and 81% for chronic fatigue syndrome—demonstrating its potential to prioritize health risks. This study underscores the transformative power of advanced AI techniques in healthcare, revealing that wearable technology data, typically dismissed as too incomplete, can yield critical health insights and contribute to better predictive models. As the AI/ML community continues to explore novel methodologies, JETS offers a compelling glimpse into the future of health data analytics and patient care.