基于机器学习的儿童啃咬行为的预测  

Prediction of children’s object mouthing behaviors based on machine learning

作　　者：王晓燕[1] 李晓萌 陈虎 张扬鸽 高捷 张庆丽 彭燕梅[5] 穆立娟 孟召学 王琳 陈芳芳[1] 涂燕晖 WANG Xiaoyan;LI Xiaomeng;CHEN Hu;ZHANG Yangge;GAO Jie;ZHANG Qingli;PENG Yanmei;MU Lijuan;MENG Zhaoxue;WANG Lin;CHEN Fangfang;TU Yanhui(Department of Children Health Care,Children’s Hospital Affiliated to Capital Institute of Pediatrics,Beijing 100020,China;Nanjing Future Network Industry Innovation Co.,Ltd.,Nanjing Jiangsu 210000,China;Department of Pediatrics,Chui Yang Liu Hospital Affiliated to Tsinghua University,Beijing 100022,China;Department of Children Health Care,Shijing Shan District Maternal and Child Health Hospital,Beijing 100040,China;Department of Child Health Care,Beijing Huairou Maternal and Child Health Hospital,Beijing 101400,China;Department of Child Health Care,Beijing Fangshan Maternal and Child Health Hospital,Beijing 102400,China;Department of Child Health Care,Tongzhou Maternal and Child Health Center of Beijing,Beijing 101100,China)

机构地区：[1]首都儿科研究所附属儿童医院保健科,北京100020 [2]南京未来网络产业创新有限公司,江苏南京210000 [3]清华大学附属北京市垂杨柳医院儿科,北京100022 [4]北京市石景山区妇幼保健院儿童保健科,北京100040 [5]北京市怀柔区妇幼保健院儿童保健科,北京101400 [6]北京市房山区妇幼保健院儿童保健科,北京102400 [7]北京市通州区妇幼保健院儿童保健科,北京101100

出　　处：《中国妇幼健康研究》2022年第3期15-22,共8页Chinese Journal of Woman and Child Health Research

基　　金：国家重点研发计划国家质量基础的共性技术研究与应用(2018YFF0214801)。

摘　　要：目的构建基于机器学习的儿童使用消费品时发生啃咬行为的预测模型。方法对北京市6所医疗机构就诊的1803例儿童进行问卷调查,依据使用消费品时是否发生啃咬行为分为有啃咬行为组(n=617)与无啃咬行为组(n=1186)。随机抽取1442例儿童的临床资料作为训练集构建预测模型,其余作为测试集进行内部验证。采用单因素分析筛选的指标应用基于机器学习的XGBoost、随机森林、Logistic回归、决策树、贝叶斯网络和SVM算法构建预测模型。完成机器学习算法对特征重要性的排序,比较6种方法构建的模型对儿童使用消费品时是否发生啃咬行为的预测价值。结果XGBoost、随机森林、逻辑回归、决策树、贝叶斯网络和SVM模型的曲线下面积(AUC)分别为0.939、0.935、0.921、0.911、0.893、0.772,灵敏度分别为0.771、0.833、0.879、0.838、0.870、0.233,特异度分别为0.928、0.898、0.847、0.874、0.751、0.969。上述6种机器学习算法优劣排序为:XGBoost>随机森林>Logistic回归>决策树>贝叶斯网络>SVM。儿童年龄(OR=0.721,95%CI=0.683~0.761)、儿童受教育水平(小学:OR=0.244,95%CI=0.170~0.352)、主要照顾者对化学物质知识了解程度(了解一点:OR=0.679,95%CI=0.466~0.990)、主要照顾者陪伴程度(经常陪伴:OR=0.471,95%CI=0.347~0.639)、母亲职业(商业服务人员:OR=0.479,95%CI=0.234~0.980)是儿童啃咬行为主要的影响因素(P<0.05)。结论基于机器学习算法建立的儿童使用消费品发生啃咬行为的预测模型具有较高的应用价值,其中XGBoost的预测效能优于随机森林、Logistic回归、决策树、贝叶斯网络和SVM。Objective To construct a prediction model based on machine learning for children’s object mouthing behaviors when using consumer products.Methods A questionnaire survey was conducted among 1803 children from 6 medical institutions in Beijing.The children were divided into object mouthing behaviors group(n=617)and non-object mouthing behaviors group(n=1186)according to whether object mouthing behaviors occurred when using consumer products.The data of 1442 collected from 1803 children was randomly chosen as the training set for establishing the prediction model and the rest was used for internal verification.Single factor analysis regression was used to filter input indicators.XGBoost,random forest,Logistic regression,decision tree,Bayesian network and support vector machine(SVM)algorithm based on machine learning were used to construct a diagnostic predictive model.To complete sorting the importance of features by the machine learning algorithm.The models constructed by 6 methods were compared for predictive and diagnostic value of children’s object mouthing behaviors when using consumer products.Results The area under the curve(AUC)of XGBoost,random forest,Logistic regression,decision tree,Bayesian network and SVM were 0.939,0.935,0.921,0.911,0.893 and 0.772.The sensitivities of the 6 models were 0.771,0.833,0.879,0.838,0.870 and 0.233.The specificities were 0.928,0.898,0.847,0.874,0.751 and 0.969.The above six machine learning algorithms were ranked as follows:XGBoost>random forest>Logistic regression>decision tree>Bayesian network>SVM.The age of the child(OR=0.721,95%CI=0.683-0.761),the education level of the child(primary school OR=0.244,95%CI=0.170-0.352),primary caregiver’s degree of knowledge of chemical substances(a little OR=0.679,95%CI=0.466-0.990),primary caregiver’s companionship degree when using consumer products(often OR=0.471,95%CI=0.347-0.639),and the mother’s occupation(business serviceOR=0.479,95%CI=0.234-0.980)were the main influencing factors to the occurrence of children object mou

关 键 词：啃咬行为 机器学习 预测模型 儿童 消费品 

分 类 号：R179[医药卫生—妇幼卫生保健]