A closed-loop feedback framework based on large language models for iterative optimization in medical record generation

Abstract

Large Language Models (LLMs) have shown great potential for generating Electronic Medical Records (EMRs) in medical natural language processing. However, existing EMR generation systems often suffer from missing information, semantic inconsistencies, and logical errors, limiting their clinical reliability. To address these challenges, we propose a Closed-loop Optimization Framework (CLOF) for EMR generation, operating through five sequential stages: generation, extraction, comparison, feedback, and regeneration. By integrating entity recognition with semantic-difference-driven feedback, CLOF enables iterative self-correction and adaptive refinement, progressively improving output quality. The key contributions of this study are: (1) a semantic-difference-driven optimization mechanism to minimize information loss; (2) an entity-level set-difference strategy and feedback function for adaptive optimization; and (3) a quantitative evaluation system verifying stable convergence across ten medical dialogue scenarios. Experimental results demonstrate that CLOF significantly enhances information completeness and semantic consistency, with the average F1-scoreincreasing by approximately 20% compared to single-round generation. These findings indicate that CLOF provides a robust self-supervised pathway for improving medical text generation.

Ähnliche Arbeiten