MedStructGen: A Two-Stage Method for Medical Record Generation Using Large Language Models

Abstract

Medical records are comprehensive repositories of patient health information and an essential tool for physicians to access medical histories. However, drafting medical records is a time-consuming process that contributes significantly to physician workload. Recent advances in Generative Artificial Intelligence (GAI) have shown strong potential in text summarization, yet most existing approaches rely on offline generation and singleturn interactions, failing to meet the real-time accuracy and user experience requirements of clinical practice. To address these limitations, we propose MedStructGen, a two-stage medical record generation framework that mirrors real-world clinical workflows. Stage 1 employs a simulation driven multiturn patient-physician dialogue model that adaptively refines its questioning strategy based on evolving symptom profiles, ensuring domain-complete diagnostic information capture. Stage 2 employs stage-specific fine-tuned LLMs with stepwise prompt engineering and entity alignment to generate EMRs that are compliant with standards and machine verifiable. This design not only improves contextual accuracy and completeness of information, but also achieves superior structural compliance, enabling seamless integration into Hospital Information Systems. Experiments on real-world hospital datasets demonstrate that our method achieves a 12.26% BLEU-4 improvement over one-stage baselines, with consistent gains in ROUGE and BERTScore. Located in a partner hospital, our system reduces physician documentation time by 1 to 1.5 hours per day, allowing more focus on patient care and personal well-being. The approach is generalizable and requires minimal customization for integration into other healthcare settings.

Ähnliche Arbeiten