Scalable Multi-Cancer Early Detection Using EHR Foundation Model with Interactive Clinical Explanations - Recent advances in artificial intelligence have introduced a new paradigm for predictive modeling: foundation models. Unlike traditional models designed for single diseases, foundation models are pretrained on large-scale, longitudinal datasets using general-purpose learning objectives, enabling broad transferability across downstream tasks. In healthcare, electronic health records (EHRs) represent a rich source of longitudinal, patient-level data that may contain early signals of disease long before diagnosis. Leveraging EHR foundation models offers a compelling opportunity to improve early detection at scale. In this project, we propose to develop a modular, generalizable, and clinically usable AI framework for early disease detection, with multi-cancer risk prediction as a motivating exemplar. Many lethal cancers currently lack recommended screening guidelines, and nearly 50% of cancer diagnoses and 60% of projected cancer deaths in the U.S. occur in such cancers. EHR-based risk prediction can fill this gap by identifying high-risk patients earlier, enabling targeted follow-up such as triage screening and diagnostic testing for cancer. Leveraging the largest national-scale EHR database (Epic Cosmos), we will develop a generalizable EHR foundation model (Aim 1), a multi-cancer risk prediction pipeline (Aim 2), and an interactive explanation platform that supports clinician follow-up queries (Aim 3). While the Cosmos model must remain on-platform, we will develop a portable Cosmos-Proxy model by fine-tuning pretrained architectures (e.g., CEHR-GPT, MOTOR) from CUIMC across two additional independent Epic sites (MSHS and CSMC) and evaluate cross-system generalizability using data from the Veterans Affairs (VA) health system. The full pipeline, including model fine- tuning, risk score generation, threshold calibration, counterfactual simulation, and explanation, is designed to be modular, supporting seamless model updates and component-level reuse. These innovations will enable the development of AI tools that are scalable, interoperable across health systems, and interpretable by clinicians. While cancer is the focus of this proposal, the proposed framework is broadly extensible to other chronic and underdiagnosed conditions. This work will lay the technical foundation for real- world deployment of AI-driven early detection systems and future prospective clinical trials.
