Domain-Adapted Large Language Models for Schema-Consistent Medical Record Generation from Doctor–Patient Dialogues

Automatically generating structured electronic medical records (EMRs) from multi-turn doctor–patient dialogues can reduce documentation burden and help standardize clinical data. We formulate medical record generation (MRG) as an instruction-following text generation task and present a large language models(LLMs) based approach combining domain continued pre-training and instruction tuning with LoRA. Unlike prior works, our framework explicitly enforces schema consistency through constrained decoding and sentinel-based field alignment. Concretely, we adapt Qwen LLMs on a medical corpus and fine-tune them to generate EMRs that conform to a target schema (chief complaint, present illness, past history, etc.) given raw dialogues. We evaluate on IMCS-V2-MRG using ROUGE-1/2/L averaged (ROUGE-AVG). Our best system, Qwen3-32B + LoRA + Prompt, achieves 68.2% ROUGE-AVG, outperforming strong baselines. An ablation on Qwen3-32B shows that instruction tuning accounts for the majority of gains (53.1% \(\rightarrow \) 67.3%, +14.2), while domain continued pre-training provides a smaller but consistent improvement (67.3% \(\rightarrow \) 68.2%, +0.9). These results indicate clear benefits from LLM scaling and task-specific alignment for MRG. Finally, we develop a gastric-cancer–specific prototype that integrates dialogue intake and EMR generation with human-in-the-loop review. The study suggests that domain-adapted LLMs are a promising direction for reliable, schema-consistent medical record generation.