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

Researchers propose CAFM, a Cohort-Anchored Foundation Model framework designed to improve interpretability and clinical reliability of AI systems for electronic health records by elevating patient cohorts to a primary learning object. The four-stage framework addresses limitations in existing EHR models through better data curation, cohort-conditioned training, multimodal alignment, and clinician feedback, with case studies demonstrating applications across kidney injury prediction, cardiovascular risk assessment, and imaging analysis.

The research addresses a critical gap in clinical AI deployment: existing foundation models achieve strong predictive performance but lack the interpretability and reasoning alignment that clinicians require for safe adoption. Rather than treating patient comparison as an emergent capability, CAFM explicitly structures the learning pipeline around clinically meaningful cohorts, fundamentally reorienting how these systems organize medical knowledge.

The healthcare AI sector has faced persistent challenges moving from research to clinical deployment. While large language models and foundation models excel at pattern recognition across diverse medical tasks, they remain black-box systems vulnerable to distribution shift and difficult to audit. This limits their utility in high-stakes clinical environments where explainability directly impacts liability and trust. CAFM's compositional design allows integration with existing models without architectural modification, lowering implementation barriers for healthcare institutions.

The framework's practical impact extends across multiple clinical domains. By organizing representations around patient cohorts rather than individual predictions, CAFM enables clinicians to understand decisions through peer comparison—a native form of clinical reasoning. This alignment with existing practice patterns could significantly accelerate adoption in hospitals and health systems skeptical of opaque AI systems.

The authors propose testable hypotheses and acknowledge open challenges in data quality, temporal irregularity, and multimodal integration. Success depends on whether cohort-anchoring genuinely improves both interpretability and predictive accuracy simultaneously. The framework represents a methodological shift toward clinician-centric AI rather than purely performance-optimized systems, potentially influencing how healthcare institutions evaluate and deploy foundation models going forward.

• →CAFM elevates patient cohorts from emergent properties to primary learning objects, improving model interpretability and clinical alignment
• →The four-stage framework addresses data quality, cohort structure, multimodal relationships, and clinician feedback without modifying underlying encoders
• →Compositional design enables integration with existing EHR foundation models, lowering adoption barriers for healthcare institutions
• →Case studies span acute kidney injury, cardiovascular risk, optic neuropathy, and report generation across imaging and temporal data
• →Framework prioritizes auditable clinical decision-making over pure predictive accuracy, addressing trust and liability concerns in healthcare deployment