随机森林算法在产后抑郁风险预测中的应用

目的： 探讨随机森林算法在产后抑郁影响因素的筛选和风险预测中的应用效果。方法： 选取2017 年6月至2018 年6 月在湖南省长沙市某三甲医院接受产前检查并在该医院分娩，符合纳入和排除标准的孕早期妇女为研究对象。入组时，使用自编的调查问卷、中文版爱丁堡产后抑郁量表(Edinburgh Postnatal Depression Scale，EPDS)调查研究对象的人口经济学、心理社会学、生物学和产科及其他特征；产后4~6 周内，采用中文版EPDS进行抑郁评分和自编的产后资料问卷收集分娩和产后资料。采用R软件在训练数据集上建立产后抑郁风险预测的随机森林模型，在验证数据集上采用预测准确率、灵敏度、特异度、阳性预测值、阴性预测值和曲线下面积(area under curve，AUC)评价模型的预测效果。结果： 共调查406 例研究对象，其中150 例的EPDS得分≥9，产后抑郁的发生率为36.9%。通过随机森林算法对训练集建立的模型在验证集上验证，得出预测准确度为80.10%，灵敏度为61.40%，特异度为89.10%，阳性预测值为73.00%，阴性预测值为82.80%，AUC值0.833。采用随机森林算法通过变量重要性评分对产后抑郁影响因素的重要程度进行排序，得出排名前10 位的重要预测变量为产前抑郁、产后经济担忧程度、产后工作担忧程度、孕早期血清游离三碘甲腺原氨酸、孕晚期高密度脂蛋白、向婴幼儿发脾气、孕早期血清总胆固醇、孕早期三酰甘油、孕晚期血细胞比容和三酰甘油。结论： 随机森林算法在产后抑郁的风险预测中具有较大优势，通过综合评价机制能从复杂的多因素中识别出产后抑郁的重要影响因素，并进行定量分析。这对识别产后抑郁关键因素，进行及时、有效干预具有重要意义。

Risk prediction for postpartum depression based onrandom forest

Objective: To explore the application of random forest algorithm in screening the riskfactors and predictive values for postpartum depression.Methods: We recruited the participants from a tertiary hospital between June 2017 andJune 2018 in Changsha City, and followed up from pregnancy up to 4−6 weeks postpartum.Demographic economics, psychosocial, biological, obstetric, and other factors wereassessed at first trimesters with self-designed obstetric information questionnaire and theChinese version of Edinburgh Postnatal Depression Scale (EPDS). During 4-6 weeks afterdelivery, the Chinese version of EPDS was used to score depression and self-designedquestionnaire to collect data of delivery and postpartum. The data of subjects wererandomly divided into the training data set and the verification data set according to theratio of 3 ? 1. The training data set was used to establish the random forest model ofpostpartum depression, and the verification data set was used to verify the predictive effectsvia the accuracy, sensitivity, specificity, positive predictive value, negative predictive value,and AUC index.Results: A total of 406 participants were in final analysis. Among them, 150 of whom hadEPDS score ≥9, and the incidence of postpartum depression was 36.9%. The predictiveeffects of random forest model in the verification data set were at accuracy of 80.10%,sensitivity of 61.40%, specificity of 89.10%, positive predictive value of 73.00%, negativepredictive value of 82.80%, and AUC index of 0.833. The top 10 predictive influentialfactors that screening by the variable importance measure in random forest model wasantenatal depression, economic worries after delivery, work worries after delivery, freetriiodothyronine in first trimesters, high-density lipoprotein in third trimester, ventingtemper to infants, total serum cholesterol and serum triglyceride in first trimester,hematocrit and serum triglyceride in third trimester.Conclusion: Random forest has a great advantage in risk prediction for postpartumdepression. Through comprehensive evaluation mechanism, it can identify the importantinfluential factors for postpartum depression from complex multi-factors and conductquantitative analysis, which is of great significance to identify the key factors forpostpartum depression and carry out timely and effective intervention.