基于生成式AI的内分泌学临床思维动态评估系统的构建与应用

Construction and Application of a Dynamic Assessment System for Clinical Reasoning in Endocrinology Based on Generative AI

摘要: 目的：旨在突破传统教学的时空限制，构建融合生成式AI的动态评估系统以提高临床医学专业学位硕士研究生的内分泌学临床诊治思维。方法：通过GPT-4等大语言模型实现虚拟病例的实时生成与交互式诊断推理，结合眼动追踪技术量化分析学员注意力分布，形成“输入–生成–推理–反馈”的闭环学习模式。将生成式AI与经典PBL相结合，开发具备动态评估功能的AI内分泌教学系统。系统核心技术包括：基于LoRA的MedGPT-Endo模型微调、Symptom2Vec症状嵌入算法、Neo4j知识图谱动态推理及眼动数据实时反馈机制。结果：该系统显著提升学员鉴别诊断全面性(提升32.5%，p < 0.01)和决策逻辑性(提升28.7%，p < 0.01)，眼动追踪显示实验组对关键检查结果的注视时间占比提升28.7%，92%的学员认为AI反馈能精准识别思维盲区，满意度达4.6/5分。结论：生成式AI的临床思维动态评估系统为赋能医学教育提供了可复制的技术路径，可实现临床思维能力的动态评估与精准提升。

Abstract: Objective: To break through the spatiotemporal constraints of traditional teaching and establish a dynamic assessment system integrated with generative AI, aimed at improving the clinical diagnostic thinking in endocrinology for postgraduate students in clinical medicine master’s programs. Methods: Leveraging large language models such as GPT-4 to enable real-time generation of virtual cases and interactive diagnostic reasoning, combined with eye-tracking technology to quantitatively analyze trainees’ attention distribution, a closed-loop learning model of “Input-Generation-Reasoning-Feedback” is formed. By integrating generative AI with classical Problem-Based Learning (PBL), an AI-powered endocrinology teaching system with dynamic assessment functions was developed. Core technologies of the system include: LoRA-based fine-tuning of the MedGPT-Endo model, the Symptom2Vec symptom embedding algorithm, Neo4j knowledge graph-driven dynamic reasoning, and a real-time eye-tracking data feedback mechanism. Results: The system significantly enhanced the comprehensiveness of trainees’ differential diagnoses (improved by 32.5%, p < 0.01) and the logicality of their decision-making (improved by 28.7%, p < 0.01). Eye-tracking data indicated that the experimental group’s dwell time on key examination results increased by 28.7%. Additionally, 92% of trainees reported that AI feedback accurately identified their cognitive blind spots, with a satisfaction rating of 4.6 out of 5. Conclusion: The generative AI-driven dynamic assessment system for clinical thinking provides a replicable technical pathway for enhancing medical education, enabling dynamic assessment and precise improvement of clinical thinking abilities.