Abstract: Objective    To establish a risk prediction model for tic disorder（TD） in children in order to provide a basis for clinical diagnosis and treatment. Methods    Initial screening of common risk factors for TD based on literature search and data mining was performed. Select 353 children aged 6-16 from the outpatient department of Liaoning University of Traditional Chinese Medicine Affiliated Hospital from December 2022 to July 2023，and divide them into two groups based on whether they had TD for risk factor investigation. Based on the findings，10 machine learning algorithms were used to build and compare TD occurrence risk prediction models：Decision Tree（DT），Linear Support Vector Machine（Linear SVC），Random Forest（RF），Linear Discriminant Analysis（LDA），Gradient Boosting，Bernoulli NB，Stochastic Gradient Descent（SGD），Ada Boost，XG Boost and Logical Regression（LR）. Results    There were statistically significant differences in risk factors of TD children compared to non-TD children（P<0.05），including gender, emotional state，degree of learning difficulty，friendship，dietary type，preference for food habits，daily time spent watching electronic products，sleeping situation，recurrent respiratory infections，allergic diseases，history of febrile convulsions，brain diseases or traumatic brain injury，mother's age at pregnancy（≥35），adverse life history during pregnancy，intrauterine development disorders，family harmony，education and upbringing methods, and family history of mental illness. The optimal TD risk prediction model was the SGD model，with accuracy of 0.87，area under curve（AUC） of 0.918，sensitivity of 0.814，and specificity of 0.886. The top five risk factors contributing to the model were：educational methods，gender，emotional status，family history of mental illness and daily time spent watching electronic products. Conclusion    The risk prediction model based on the SGD algorithm in this study is an optimal fitting prediction model，which has some predictive value and provides a basis for clinical diagnosis and treatment.

Key words:  , tic disorder, risk factors, characteristic analysis, prediction model

摘要： 目的    建立儿童抽动障碍（tic disorder，TD）风险预测模型，为临床诊疗提供依据。方法    基于文献检索和数据挖掘，初步筛选TD常见危险因素。选取2022年12月至2023年8月辽宁中医药大学附属医院门诊共353例儿童资料，年龄6～16岁，根据是否患有TD分为两组进行常见危险因素调查。基于调查结果，采用10种机器学习算法建立并比较TD发生风险预测模型：决策树（decision tree，DT）、线性支持向量机（linear SVC）、随机森林（random forest，RF）、线性判别分类（linear discriminant analysis，LDA）、梯度提升（gradient boosting）、伯努利朴素贝叶斯（bernoulli NB）、随机梯度下降（stochastic gradient descent，SGD）、Ada Boost、XG Boost、逻辑回归（Logistic regression，LR）。结果    TD儿童在性别、情绪、学习困难度、朋友关系、饮食类型、偏食习惯、每日看电子产品时长、入睡情况、反复呼吸道感染、过敏性疾病、高热惊厥史、脑部疾病或脑外伤、母亲妊娠年龄、孕期不良生活史、宫内发育不良、父母间关系、教育及抚养方式、家族精神疾病史等方面与非TD儿童差异有统计学意义（P<0.05）。最优TD风险预测模型为SGD模型，准确度0.85、受试者工作特征曲线（ROC）曲线下面积（AUC）为0.918、敏感度0.814、特异度0.886。对模型贡献度前五位的危险因素为：教育方式、性别、情绪状态、家族精神疾病史、每日看电子产品时长。结论    研究中基于SGD算法的风险预测模型为最优拟合预测模型，具有一定预测价值，为临床诊疗提供依据。

关键词: 抽动障碍, 危险因素, 特征分析, 预测模型