Persona-Conditioned Generation of Patient Self-Reports from EHRs

Abstract

Accurate diagnosis depends not only on clinical expertise but also on how patients describe their symptoms at first contact. Yet large English corpora of patient-authored self-reports are scarce, limiting advances in natural, context-aware narrative modeling. We address this gap by generating first-person self-reports from structured EHR content conditioned on persona attributes that capture social and clinical context. Reports are produced by two generators and scored by two independent graders using a rubric with four dimensions, complemented by a rubric-free preference test. Across 10k stratified cases, we compare two generators under a reliable evaluation protocol and select the higher-scoring one based primarily on Clinical Correctness and Faithfulness, yielding a dataset composed of narratives from the stronger system. Our contributions are threefold: (I) we developed and release a large, persona-conditioned dataset of patient-style self-reports grounded in patient-stated EHR facts, (II) we introduce a transparent evaluation framework that combines rubric-based scoring with rubric-free preference to mitigate grader bias and enable cross-validation, (III) we find that graders exhibit systematic stylistic preferences in rubric-free approach that influence scores independent of clinical content, and (IV) we study large language models for producing first-person self-reports from structured EHRs, highlighting where they succeed, where they fail, and how this affects use in telemedicine and triage.

Resources

Details