基于ERNIE分类模型的淋巴水肿智能辅助诊断研究  

Study on Intelligent Auxiliary Diagnosis of Lymphedema Based on ERNIE Classification Model

作　　者：汤昊宬 信建峰[2] 冀秀元 常鲲[2] 孙宇光[2] 夏松[2] 徐毅[1] 沈文彬[2] TANG Haocheng;XIN Jianfeng;JI Xiuyuan;CHANG Kun;SUN Yuguang;XIA Song;XU Yi;SHEN Wenbin(Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;Beijing Shijitan Hospital,Capital Medical University,Beijing 100038,China)

机构地区：[1]中国科学院自动化研究所,北京100190 [2]首都医科大学附属北京世纪坛医院,北京100038

出　　处：《医学信息学杂志》2023年第10期44-49,共6页Journal of Medical Informatics

基　　金：科技创新2030——“新一代人工智能”重大项目(项目编号:2020AAA0105003);科技创新2030——“新一代人工智能”重大项目(项目编号:2020AAA0105005);北京市科学技术委员会项目(项目编号:Z191100007619049)。

摘　　要：目的/意义 探讨人工智能辅助诊断技术应用于淋巴疾病的诊断效果,阐述样本稀缺背景下模型微调训练的解决方案,指出应用研究的难点和不足,并提出展望。方法/过程 从北京世纪坛医院淋巴外科既往收治的患者中选取755例为研究对象,基于患者电子病历非结构化文本数据,利用ERNIE 3.0分类模型进行淋巴水肿辅助诊断应用研究,通过两个层级分类任务,实现智能化诊断淋巴水肿疾病并区分原发性和继发性。结果/结论 ERNIE 3.0分类模型在淋巴水肿的判别任务中,准确率、精确率、召回率、平均F1值均超过0.95。在原发性和继发性淋巴水肿的区分上,模型的平均Macro F1值达到0.901。两个任务模型的曲线下面积分别达到0.97和0.865,表明模型的准确分类效果。模型分类结果具有较好的可解释性,填补了智能化方法在淋巴疾病领域的应用空白。Purpose/Significance The paper discusses the diagnostic effectiveness of artificial intelligence(AI)auxiliary diagnosis technology applied to lymphatic diseases,expounds the solution of model fine-tuning training under the background of scarce samples,points out the difficulties and shortcomings of applied research,and puts forward prospects.Method/Process Based on the unstructured text data of patients’electronic medical records(EMR),the ERNIE 3.0 classification model is used to conduct a study on the application of lymphedema auxiliary diagnosis.Through two levels of classification tasks,intelligent diagnosis of lymphedema disease and distinguish between primary lymphedema and secondary lymphedema are realized.Result/Conclusion The ERNIE 3.0 classification model shows accuracy,precision,recall,and mean F1 values over 0.95 in the discriminatory task of lymphedema.The mean Macro F1 value of the model reaches 0.901 in the differentiation of primary and secondary lymphedema.The AUC of the model reaches 0.97 and 0.865 in the two tasks,respectively,indicating the accurate classification effect of the model.In addition,the model classification results have good interpretability,and fill the application gap of intelligent methods in the field of lymphatic diseases.

关 键 词：原发性淋巴水肿 继发性淋巴水肿 辅助诊断 

分 类 号：R-058[医药卫生]