EviGen: Predictive Evidence Scaffolding for Verifiable Clinical Rationale Generation

TL;DR - EviGen is a three-layer framework that retrieves outcome-predictive evidence from longitudinal health records, uses it to scaffold clinical rationale generation, and verifies claims step by step. It improves predictive performance and rationale faithfulness while reducing omissions and hallucinations associated with full-context LLM and standard RAG approaches.

• A patient-conditioned retriever uses learnable queries to identify evidence predictive of clinical outcomes, ranking spans by attribution scores rather than textual relevance alone.
• An LLM generates clinical rationales grounded in the ranked evidence scaffold.
• A process-supervised verifier evaluates individual reasoning steps and flags unreliable claims.
• Across three medical prediction datasets, EviGen outperformed full-context LLM and RAG baselines and was preferred by clinical reviewers in a usability evaluation.