机器学习对H.pylori感染患者的特征变量及预测模型研究  

Machine learning study of characteristic variables and predictive models for patients with H.pylori infection

作　　者：袁一鸣 杜结玲[2] 洪慧斯 韦翠花 卢苑香 YUAN Yiming;DU Jieling;HONG Huisi;WEI Cuihua;LU Yuanxiang(Department of Pharmacy,Zhongshan Hospital of Traditional Chinese Medicine,Zhongshan 528401;Department of Endoscopy and Treatment,Zhongshan Hospital of Traditional Chinese Medicine;Department of Science and Education,Zhongshan Hospital of Traditional Chinese Medicine;Department of Endoscopy and Treatment,Zhongshan South District Hospital;Department of Endoscopy and Treatment,Dongfeng People′s Hospital of Zhongshan,China)

机构地区：[1]中山市中医院药学部,广东中山528401 [2]中山市中医院内镜中心 [3]中山市中医院科教科 [4]中山市南区医院内镜室 [5]中山市东凤人民医院内镜室

出　　处：《胃肠病学和肝病学杂志》2024年第8期958-965,共8页Chinese Journal of Gastroenterology and Hepatology

基　　金：中山市科学技术局第一批社会公益(医疗卫生一般项目)(2021SYF01)。

摘　　要：目的 分析H.pylori感染患者感染的危险因素,建立H.pylori感染患者预测模型,为防治H.pylori感染提供参考。方法 选取2021年7至2022年5月在中山市中医院、中山市东凤人民医院、中山市南区医院共1 477例接受H.pylori检测者为研究对象,依据胃镜和~(14)C、~(13)C呼气试验的检测结果,将H.pylori受检人群分为感染组和无感染组,分别进行问卷调查,调查内容包括受检者基本情况、临床表征、慢性基础病、生活和饮食习惯等,共计63个变量。采用单因素及机器学习中的Logistic回归、决策树分析及添加交互项的Logistic回归对H.pylori感染进行多因素分析,并比较3个模型的ROC曲线下的面积、灵敏度、特异度,验证模型的准确性,建立H.pylori感染预测模型,将特征和危险因素建立森林图。结果 Logistic回归分析的AUC为0.7361,灵敏度为0.7615,特异度为0.6034。决策树分析的AUC为0.6528,灵敏度为0.6801,特异度为0.5773。添加交互项后的Logistic回归分析的AUC为0.7388,灵敏度为0.7588,特异度为0.6034。添加交互项的多因素Logistic回归结果显示,有胃胀,口气、口臭,在家煮食午餐,在家无而外出有使用公筷习惯,同居家人有感染,疫情后才使用公筷,居住4~10层楼,同时有胃胀及口气、口臭为模型的显著性变量。结论 胃胀,有口气、口臭,同时有胃胀及口气、口臭,在家煮食午餐,居住的楼层数,外出居家是否使用公筷,是否有使用公筷习惯,家人是否感染H.pylori是感染H.pylori的特征因素,用Logistic回归模型作为主模型进行变量筛选,添加交互后的模型,AUC有所提升,交互项的预测模型对H.pylori感染者预判能力好,运算容易,使用经济、便利,适合区域性推广。Objective To analyze the risk factors of infection in patients with H.pylori infection,establish a prediction model for patients with H.pylori infection,and provide reference for the prevention and treatment of H.pylori infection.Methods A total of 1477 patients tested for H.pylori in Zhongshan Hospital of Traditional Chinese Medicine,Dongfeng People′s Hospital of Zhongshan and Zhongshan South District Hospital from Jul.2021 to May 2022 were selected as study subjects.The results of gastroscopy,14 C and 13 C breath tests were used to divide the H.pylori tested population into infected and non-infected groups,and questionnaires were administered to investigate the contents of the survey,including the basic a total of 63 variables were included in the survey,including the basic conditions,clinical features,chronic underlying diseases,life and dietary habits.The results were subjected to multifactorial analysis of H.pylori infection using single factor and Logistic regression in machine learning,decision tree analysis,and Logistic regression with added interaction terms,and the area under the ROC curve,sensitivity and specificity of the 3 models were compared to verify the accuracy of the models,and a prediction model of H.pylori infection was established,and the characteristics and risk factors were established as a forest plot.Results The AUC of Logistic regression was 0.7361,sensitivity was 0.7615,and specificity was 0.6034.The AUC of decision tree was 0.6528,sensitivity was 0.6801,and specificity was 0.5773.The AUC of Logistic with the addition of interaction term was 0.7388,sensitivity was 0.7588,and specificity was 0.6034.The multifactorial Logistic regression with the addition of interaction terms showed that having stomach bloating,bad breath and halitasis,cooking lunch at home,not having it at home but having the habit of using communal chopsticks when going out,having infected family members living together use public chopsticks only after the epidemic,lived in 4 to 10 floors,and also had stomach bloatin

关 键 词：幽门螺杆菌 二元Logistic回归模型 决策树 森林图 交互项 

分 类 号：R378[医药卫生—病原生物学]