【作者】 张扬；

【导师】 李健；

【作者基本信息】 东南大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 随着人口老龄化趋势加剧和慢性病高发成为全球性公共卫生问题,慢性阻塞性肺疾病(Chronic Obstructive Pulmonary Disease,COPD)作为常见的慢性呼吸系统疾病之一,其长期管理与随访已成为基层医疗系统亟需解决的难题。目前传统的COPD随访方式依赖于面对面或电话交流,存在效率低、覆盖范围有限、信息采集质量不高等问题,难以满足个体化和连续性管理的实际需求。尽管人工智能技术已初步应用于临床随访场景,但多停留在单向信息推送层面,缺乏与患者的深度交互与动态适配能力。因此,探索更加智能化、个性化和高效的随访路径,对于提升COPD管理质量、减轻医护负担、促进患者参与具有重要意义。基于以上讨论,本研究提出了一种基于大型语言模型(Large Language Model,LLM)驱动的COPD智能随访方法,构建了一个贯穿数据采集、交互生成、模型优化与系统实现的完整随访解决方案。在数据层面,本研究联合东南大学附属中大医院呼吸与危重症医学科开展真实数据采集,设计符合临床随访流程的数据采集模板,涵盖人口学信息、生活方式评估、依从性行为、症状量表评分(m MRC、CAT、MNA-SF)及合并症信息等关键维度。通过结构化访谈共采集142名COPD患者的随访样本数据,构建“COPD随访信息种子数据集”,统计分析结果显示该数据集具有良好的代表性,充分反映了我国基层医疗体系中慢阻肺患者的主要特征,为后续建模提供了坚实的数据基础。针对缺乏高质量多轮随访对话数据的问题,本研究设计并实现基于LLM-Agents的模拟慢阻肺随访情景对话生成方法(Simulation of COPD Follow-up Dialogue Generation Based on LLM-Agents,SCFDG-LA)。该方法由患者模拟器、医护模拟器、监督以及信息记录四个智能体组成。结合基于大五人格特质的人格建模方式,塑造多样化的患者沟通风格和行为,定义智能体的动作空间和决策过程。针对传统的串行调用API生成效率低的问题,基于双线程池异步流水线实现了并发优化。实验表明,SCFDG-LA方法生成的COPD随访对话在结构完整性、句法复杂性、患者人格匹配度和医疗交互连贯性方面均优于基线方法。针对大语言模型在专业随访场景中的适配问题,本研究构建适应COPD随访任务的大语言模型能力增强机制,提出“两阶段能力增强微调”策略。第一阶段基于医学考试数据、专业知识图谱和临床对话语料进行指令微调,强化模型在医学问答和通识任务上的理解能力;第二阶段将随访任务拆解为多个子任务,基于“COPD随访信息种子数据集”和SCFDG-LA方法生成的COPD随访对话内容,构建统一格式的多任务数据集,引入损失归一化与动态任务加权机制,提升任务协同效率与泛化能力。实验证明,该策略在BLEU、BLEURT、GPT评分等主客观语言生成质量指标显著优于基线方法。在系统实现层面,本研究设计并实现了大语言模型驱动的COPD智能随访流程原型,构建可交互、可视化、流程闭环的端到端随访平台。系统整体由随访交互模块与健康管理模块组成,前者整合语音识别、COPD随访LLM、语音合成与数字人驱动,支持自然语音或文字的多轮交互与结构化信息记录;后者通过健康画像动态构建与知识库驱动,实现个性化健康教育、智能问答与随访后指导建议的自动生成。系统采用流处理与并行优化机制,保障低延迟、高流畅的用户体验,在界面交互与工程实现方面具备良好的部署基础。综上所述,本研究以大语言模型为核心技术支撑,系统性提出并实现了大语言模型驱动的COPD智能随访方法的完整解决方案,从真实数据采集、智能对话数据生成、模型能力优化到系统流程集成,全面实现了从理论研究到工程实践的闭环落地。更多还原

【Abstract】 As the trend of population aging intensifies and the high incidence of chronic diseases become global public health issues,chronic obstructive pulmonary disease(COPD)is one of the common chronic respiratory diseases,and its long-term management and follow-up has become a difficult problem that the primary medical system urgently needs to solve.At present,the traditional COPD follow-up method relies on face-to-face or telephone communication,which has problems such as low efficiency,limited coverage,and low quality of information collection,and it is difficult to meet the actual needs of individualized and continuous management.Although artificial intelligence technology has been initially applied to clinical follow-up scenarios,it mostly stays at the level of one-way information push,lacking deep interaction and dynamic adaptation capabilities with patients.Therefore,exploring more intelligent,personalized and efficient follow-up paths is of great significance to improving the quality of COPD management,reducing the burden of medical care,and promoting patient participation.Based on the above discussion,this study proposed a COPD intelligent follow-up method driven by a large language model(LLM),and constructed a complete follow-up solution that runs through data collection,interaction generation,model optimization and system implementation.At the data level,this study conducted real data collection in collaboration with the Department of Respiratory and Critical Care Medicine of Zhongda Hospital Affiliated to Southeast University,and designed a data collection template that conforms to the clinical follow-up process,covering key dimensions such as demographic information,lifestyle assessment,compliance behavior,symptom scale scores(m MRC,CAT,MNA-SF)and comorbidity information.Through structured interviews,a total of 142 COPD patients’follow-up sample data were collected to construct a COPD follow-up information seed data set.The statistical analysis results showed that the data set has good representativeness and fully reflects the main characteristics of COPD patients in my country’s primary medical system,providing a solid data foundation for subsequent modeling.To address the issue of insufficient high-quality multi-round follow-up dialogue data,this study designed and implemented a simulation method called SCFDG-LA(Simulation of COPD Follow-up Dialogue Generation Based on LLM-Agents).The method consists of four agents:patient simulator,medical simulator,supervision,and information recording.Combined with the personality modeling method based on the Big Five personality traits,it shapes diverse patient communication styles and behaviors,and defines the action space and decision-making process of the agent.In order to solve the problem of low efficiency of traditional serial call API generation,concurrent optimization is implemented based on dual-thread pool asynchronous pipeline.Experiments show that the COPD follow-up dialogue generated by the SCFDG-LA method is superior to the baseline method in terms of structural integrity,syntactic complexity,patient personality matching and medical interaction coherence.In order to solve the adaptation problem of large language models in professional follow-up scenarios,this study constructs a large language model capability enhancement mechanism adapted to COPD follow-up tasks and proposes a"two-stage capability enhancement"strategy.In the first stage,instructions are fine-tuned based on medical examination data,professional knowledge maps and clinical dialogue materials to enhance the model’s understanding ability in medical question-answering and general knowledge tasks;in the second stage,the follow-up task is decomposed into multiple subtasks,and a multi-task dataset in a unified format is constructed based on the COPD follow-up information seed dataset and the COPD follow-up dialogue content generated by the SCFDG-LA method.The loss normalization and dynamic task weighting mechanism are introduced to improve the task coordination efficiency and generalization ability.Experiments have shown that this strategy is significantly superior to the baseline method in subjective and objective language generation quality indicators such as BLEU,BLEURT,and GPT scores.At the system implementation level,this study designed and implemented a prototype of the COPD intelligent follow-up process driven by a large language model,and built an interactive,visual,and closed-loop end-to-end follow-up platform.The system as a whole consists of a follow-up interaction module and a health management module.The former integrates speech recognition,COPD follow-up LLM,speech synthesis,and digital human drive,and supports multiple rounds of interaction and structured information recording of natural speech or text;the latter realizes personalized health education,intelligent question and answer,and automatic generation of post-follow-up guidance suggestions through dynamic construction of health portraits and knowledge base drive.The system adopts stream processing and parallel optimization mechanisms to ensure low-latency and high-smooth user experience,and has a good deployment foundation in interface interaction and engineering implementation.In summary,this study takes the large language model as the core technical support,systematically proposes and implements a complete solution for the COPD intelligent follow-up method driven by a large language model,from real data collection,intelligent dialogue data generation,model capability optimization to system process integration,and fully realizes the closed-loop implementation from theoretical research to engineering practice.更多还原