Cohort-Anchored Foundation Models for Electronic Health Records: From Risk Scores to Auditable Peer Cohorts

The article introduces the Cohort-Anchored Foundation Model (CAFM) framework designed to enhance the deployment of AI in electronic health records (EHR) by treating patient cohorts as a primary focus in the learning process. CAFM consists of four key stages: deviation-aware data curation, cohort-conditioned pretraining, multimodal cohort alignment, and clinician-in-the-loop refinement, which collectively aim to improve data quality and facilitate auditable clinical decision-making. This framework can augment existing EHR models without altering their encoders and addresses critical challenges in clinical AI, such as interpretability and distribution shifts, making it a significant advancement for practitioners in the field.