基于机器学习的胃肠道疾病舌诊模型构建  

Construction of a machine learning-based tongue diagnosis model for gastrointestinal diseases

作　　者：张景慧 王娟 赵玉洁 段淼 刘毅然 林敏娟 谯旭 李真[1] 左秀丽[1] ZHANG Jinghui;WANG Juan;ZHAO Yujie;DUAN Miao;LIU Yiran;LIN Minjuan;QIAO Xu;LI Zhen;ZUO Xiuli(Department of Gastroenterology,Qilu Hospital of Shandong University,Jinan 250012,Shandong,China;Hospital Development Center of Qingdao Municipal Health Commission,Qingdao 266001,Shandong,China;Department of Gastroenterology,Affiliated Hospital of Shandong University of Traditional Chinese Medicine,Jinan 250011,Shandong,China;School of Control Science and Engineering,Shandong University,Jinan 250061,Shandong,China)

机构地区：[1]山东大学齐鲁医院消化内科,山东济南250012 [2]青岛市卫生健康委员会医院发展中心,山东青岛266001 [3]山东中医药大学附属医院消化内科,山东济南250011 [4]山东大学控制科学与工程学院,山东济南250061

出　　处：《山东大学学报（医学版）》2024年第1期38-47,70,共11页Journal of Shandong University:Health Sciences

基　　金：国家自然科学基金(82070551)。

摘　　要：目的构建基于机器学习的胃肠道疾病舌诊模型,以寻求更加方便、经济的方式实现对常见胃肠道疾病的非侵入性诊断。方法前瞻性收集接受电子内镜检查的948名受试者的舌象图片,经过质量筛选,最终获得符合应用标准的3140张图片构成本研究使用的舌象数据集。对原始舌象数据进行预处理、特征提取与模式识别,在传统机器学习方法的基础之上,提出一种从特征融合和决策融合两个方面实现信息融合的方法,以此构建以舌象特征为输入的胃肠道疾病舌诊模型。结果本研究构建的基于舌象的信息融合诊断模型的曲线下面积(area under the curve,AUC)为0.808,高于单一手工特征(AUC=0.769)和深度特征(AUC=0.779)模型;使用BSFCM混合采样方法进行样本增强提高了该模型对幽门螺杆菌(Helicobacter pylori,H.pylori)感染(AUC=0.816)、胆汁反流(AUC=0.829)、反流性食管炎(AUC=0.800)、胃糜烂(AUC=0.833)和十二指肠糜烂(AUC=0.818)的分类性能。结论本研究构建的基于机器学习的智能舌诊模型对多种胃肠道疾病具有较高的区分度,或为胃肠道疾病的诊断与筛查提供一种新的、有价值的思路与方法。Objective To construct a machine learning(ML)-based tongue diagnostic model for the diagnosis of gastrointestinal diseases so as to realize the non-invasive auxiliary diagnosis of common gastrointestinal diseases in a more convenient and faster way.Methods Tongue images of 948 subjects who underwent electronic endoscopy were prospectively collected.After quality screening,3,140 images that met the application criteria were finally obtained to constitute the tongue image data set,which underwent preprocessing,feature extraction and pattern recognition.On the basis of traditional machine learning methods,a method to realize information fusion in terms of feature fusion and decision fusion was proposed,and a tongue diagnosis model of gastrointestinal diseases was constructed.Results The area under the curve(AUC)of the model was 0.808,which was higher than that of the single handcrafted feature(AUC=0.769)and deep feature(AUC=0.779)models.Sample enhancement using the BSFCM hybrid sampling method improved the model's performance for Helicobacter pylori(H.pylori)infection(AUC=0.816),bile reflux(AUC=0.829),reflux esophagitis(AUC=0.800),gastric erosion(AUC=0.833)and duodenal erosion(AUC=0.818).Conclusion The intelligent tongue diagnostic model based on ML constructed in this study shows a high degree of differentiation for a variety of gastrointestinal diseases,and may provide a new and valuable idea and method for the diagnosis and screening of gastrointestinal diseases.

关 键 词：人工智能 舌象 胃肠道疾病 机器学习 舌诊模型 

分 类 号：R241.25[医药卫生—中医诊断学] R57[医药卫生—中医临床基础]