【摘要】 目的 针对通用大语言模型在专病方向深度不足的问题,探索大模型在医疗专病领域的系统性建设路径与实践方法,构建专病模型从训练、评估到应用的完整闭环体系。方法 在构建集成大模型能力的智能应用整体架构基础上,以骨科颈椎病为研究切入点,设计“全参数预训练—监督微调—GRPO强化学习”三阶段训练策略,探索并验证专病模型全流程建设路径。结果 成功构建初步可用的骨科颈椎病专病模型;该模型知识准确率(91.7%)显著优于基准模型DS-R1 (64.3%),临床医生使用满意度100%。结论 本研究创新性提出“整体规划、专科突破、场景落地”实施策略,为大模型在医疗专病领域的应用提供了可复制的实践范式,证明相较于通用模型,专病模型在对应临床场景中表现更为精准。更多还原

【Abstract】 Objective In response to the issue of the general large language model’s insufficient depth in specific disease areas,we explore the systematic construction path and practical methods for large models in the field of medical specific diseases,and establish a complete closed-loop system for the training,evaluation,and application of specific disease models.Methods Based on the construction of the overall architecture of intelligent applications that incorporate the capabilities of large models,taking cervical spondylosis in orthopedics as the research entry point,we designed a three-stage training strategy of "fullparameter pre-training-supervised fine-tuning--GRPO reinforcement learning",to explore and verify the entire construction path of the specialized disease model.Results A preliminarily usable specific disease model for cervical spondylosis in orthopedics is successfully built.The knowledge accuracy rate of this model(91.7%) is significantly better than that of the benchmark model DS-R1(64.3%),and the satisfaction rate of clinicians is 100%.Conclusion This study innovatively proposes an implementation strategy of "holistic planning,specialized breakthrough,and scenario-based application," offering a replicable practical paradigm for the application of large models in the field of medical specialized diseases.It demonstrates that,compared to general-purpose models,specialized disease models exhibit superior precision in corresponding clinical scenarios.更多还原