大语言模型优化放射科住培医师乳腺亚专科胜任力评估
李志 , 史玉书 , 张瑞 , 赵艺蕾 , 黄强 , 楼海燕 , 肖文波
医学教育研究与实践 ›› 2026, Vol. 34 ›› Issue (2) : 327 -332.
大语言模型优化放射科住培医师乳腺亚专科胜任力评估
Optimizing Competency Assessment of Radiology Residents in Breast Imaging Subspecialty Using Large Language Models
目的 该研究聚焦于放射科住培医师亚专科报告书写岗位胜任力评估,旨在探究应用大语言模型进行自动化评价的可行性。 方法 以放射科乳腺亚专科住培医师报告为研究对象,模拟包含不同错误类型的100份乳腺 X 线摄影报告和100份乳腺 MRI 报告。通过乳腺组亚专科负责医师及教学主任协商制定岗位胜任力评价表,采用大语言模型思维链模式分步输入指令进行报告胜任力分级,并与人工分级结果对比。 结果 与人工相比,大语言模型(Large Language Model,LLM)判读时间仅0.13小时,效率优势显著,其分级准确性稍高于低年资带教师资,与高年资带教师资相近,且在不同级别岗位胜任力判定上表现稳定。在引入多种错误的报告中,LLM准确性达 91.3%,表明能够较好进行分级优先级划分,但对MRI报告书写评价的准确率低于MM,分别为89%和93%。 结论 LLM 用于放射科住培医师亚专科报告书写岗位胜任力评价具有准确性高、效率高、结果稳定等优势,可为住培教学评估提供有力支持,但LLM仍有处理及分析复杂文本时存在限制以及需要人工参与进行思维链提示输入等不足,未来需要进一步地探索与研究。
Objective This research focuses on the competency assessment of Radiology residents in sub-specialty report writing, aiming to explore the feasibility of using large language model (LLM) for automated evaluation. Methods Taking Radiology residents specializing in breast sub-specialty as the research objects, 100 mammography reports and 100 breast magnetic resonance imaging (MRI) reports which contain different types of errors were simulated. A competency evaluation form was developed through collaboration between the breast sub-specialty specialists and the teaching directors. The chain-of-thought model of LLM was applied to input instructions step by step to grade the competency report, and compare the results with the manual grading. Results Compared with human evaluators, LLM required only 0.13 hours for grading, demonstrating significant efficiency advantages. The accuracy of the LLM grading was slightly higher than that of junior faculty and similar to that of senior faculty, with stable performance in competency assessment across different levels. In reports containing multiple errors, the accuracy of LLM reached 91.3%, indicating its ability to effectively prioritize grading tasks. However, the accuracy for MRI report grading was lower than that for MM reports, with accuracy rates of 89% and 93%, respectively. Conclusion The application of LLM in evaluating the competency of Radiology residents in sub-specialty report writing offers advantages in terms of high accuracy, efficiency, and stable results, providing strong support for resident education assessment. However, LLM still face challenges in handling and analyzing complex texts, and requires human involvement for guiding input through the chain-of-thought model. Further exploration and research are required for future improvements.
| [1] |
中国医师协会. 住院医师规范化培训内容与标准 [S].北京:中国医师协会,2022. |
| [2] |
穆琳,王景宇,周宏伟, |
| [3] |
信斯言,沈子曰,吴红斌 .面向未来医学教育 全球健康与人工智能:2024AMEE年会综述2[J].中华医学教育杂志,2024,44(12):896-900. |
| [4] |
陈冬河,杨君,翁婉雯, |
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
江哲涵,奉世聪,王维民 .生成式大语言模型在医学考试题库建设中的实践探索[J].中华医学教育杂志,2024,44(8):561-569. |
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
熊绘,樊越,张迪, |
| [15] |
门茜儒,刘林 .人工智能赋能医学教育现代化的应用挑战与现实路径[J].医学教育研究与实践,2025,33(5):641-646. |
浙江大学医学院教育改革项目(jgzx2025003)
/
| 〈 |
|
〉 |