摘要目的:应用机器学习中的随机森林算法探讨中国女性尿失禁（UI）发病的危险因素，并评价各危险因素对于UI发病的预测效果。方法:采用多阶段分层整群抽样，在全国调查55 477例成年女性UI情况；基线调查于2014年2月至2016年1月完成，2018年6月至12月电话随访；最终纳入基线无UI且随访UI诊断指标数据完整的对象。采用欠采样技术，按照1∶1的比例从随访时未发生UI的人群中随机抽取与随访对新发UI相等人数作为对照，将这些调查对象的研究数据按照7∶3的比例随机分成训练集和测试集。将单因素分析中 P<0.2的候选变量，带入训练集并采用随机森林算法建模，在训练集筛选UI发病的危险因素，根据重要性对危险因素排序，并在测试集中验证。 结果:共30 658例（55.26%，30 658/55 477）完成随访，中位随访时间3.7年。纳入本研究的24 985例基线无UI的对象中，随访调查UI发病人数为1 757例（7.03%，1 757/24 985），其中压力性UI 1 117例（4.47%，1 117/24 985），急迫性UI 243例（0.97%，243/24 985），混合性UI 397例（1.59%，397/24 985）。随机森林模型固定特征数量为2个、决策树数量为300棵时，平均袋外估计误差率最低，此时模型分类准确率为64.3%，敏感度为64.2%，特异度为64.4%。根据Gini系数平均下降量，得到预测UI发病的前10位影响因素依次为：年龄、分娩次数、分娩方式、体质指数（BMI）、绝经状态、糖尿病史、教育程度、盆腔手术史、城乡分布、婚姻状况。结论:应用机器学习中的随机森林算法，从复杂的多因素中识别出预测中国女性UI发病的前10位影响因素，依次为：年龄、分娩次数、分娩方式、BMI、绝经状态、糖尿病史、教育程度、盆腔手术史、城乡分布、婚姻状况。

abstractsObjective:To explore the risk factors of urinary incontinence (UI) in China by using random forest algorithm, and to evaluate the predictive effect of each risk factor on UI.Methods:A baseline survey with a multistage stratified cluster sampling design was conducted between February 2014 and January 2016, and followed up by telephone from June to December 2018. A total of 55 477 adult women from six provinces of China participated the survey. According to the ratio of 1:1, under sampling method was used to randomly select the same number of women as UI from the non UI women. The data were randomly divided into training set and verification set according to 7:3. The training set was used to establish the random forest model, which including the candidate variables with P<0.2 in univariate analysis, and the verification set was used to verify the predictive effects. Results:A total of 30 658 patients (55.26%, 30 658/55 477) completed the follow-up, the median follow-up time was 3.7 years. Among the 24 985 women without UI at baseline, 1 757 (7.03%, 1 757/24 985) had UI at followed up, including 1 117 (4.47%, 1 117/24 985) with stress UI, 243 (0.97%, 243/24 985) with urgency UI and 397 (1.59%, 397/24 985) with mixed UI. When fixed the number of features as 2 and the number of random trees as 300 in the random forest model, the out of bag error rate estimation was the lowest; with such parameter settings, the classification accuracy was 64.3%, the sensitivity was 64.2%, and the specificity was 64.4%. The top10 predictive UI factors that screening by the variable importance measure in random forest model were obtained as follows: age, parity, delivery pattern, body mass index (BMI), menopause, history of diabetes, education level, history of pelvic surgery, regions, and marital status.Conclusion:We identified the top10 predictive UI factors that screening by the variable importance in random forest model as follows: age, parity, delivery pattern, BMI, menopause, history of diabetes, education level, history of pelvic surgery, regions, and marital status.