New AI Model Scales Up Chronic Disease Prediction Across Hospital Networks

Identifying patients at high risk for chronic illnesses like heart disease and cancer remains a major challenge in healthcare, particularly when patient records are scattered across different hospital systems using incompatible data formats. A new electronic health record foundation model tackles this problem by learning from massive amounts of longitudinal medical data collected across multiple countries and healthcare networks.

Training at Scale Across Borders#

The model was pretrained on billions of individual medical events drawn from medical records of over 5 million patients spanning Taiwan and the United States. To make this cross-border training possible, the researchers developed a unified code alignment framework that translated between different coding systems used in each country's healthcare infrastructure—a critical step for harmonizing data from heterogeneous sources. The team scaled the model up to 2.4 billion parameters and used IsoFLOP analysis to identify compute-optimal configurations.

Outperforming Existing Approaches#

Tested across 11 chronic disease prediction tasks, the model demonstrated superior performance compared to tree-based methods and even general-purpose or biomedical language models. On U.S. patient cohorts, it achieved over 40% sensitivity at 99% specificity; on Taiwan cohorts, it reached above 70% sensitivity at the same specificity threshold. When evaluated on EHRShot—a standard benchmark for EHR models—the foundation model surpassed previously developed EHR models despite significant differences between training and test data distributions, showing strong few-shot generalization abilities.

Why Alignment Matters#

A key finding emerged from the researchers' comparison of different training approaches: data aligned across multiple healthcare systems provided a stronger learning signal than simply duplicating data from a single site, even when the total volume was comparable. This highlights that diversity in patient populations and medical contexts strengthens model robustness more than sheer volume alone.

Next Steps#

The team plans to open-source the code, potentially enabling other researchers and health systems to build on this work for population health management applications.