BIDS Special Seminar

Traditional clinical prediction models are narrowly focused on outcomes like length of stay, readmission, or disease progression. Building and maintaining these single-purpose models demands substantial data and feature engineering, limiting scalability and generalization across health systems.

Developed by Epic in collaboration with researchers at Yale University and Microsoft Research, Curiosity is a family of foundation models for longitudinal health records designed to support a diverse set of clinical tasks within a single generative modeling approach. Trained on de-identified data from 118 million patients and 115 billion clinical events from hundreds of healthcare systems in Cosmos, Curiosity models patient timelines as sequences of medical events, enabling prediction and simulation of clinical trajectories.

Davis White will outline the methodological framework underlying the first version of Curiosity, including the establishment of scaling laws that relate data volume, model capacity, and performance in medical event modeling. Results across 78 real-world clinical and operational tasks demonstrate that Curiosity can match or exceed specialized models without task-specific tuning. These findings suggest that foundation models can support more generalizable, scalable, and data-driven approaches to clinical prediction and research.