中国老年高血压患者抑郁发生风险的可解释机器学习预测模型构建:基于CHARLS数据的队列研究OA

Objective To construct prediction models for the risk of depression in elderly hypertensive patients in China by machine learning algorithms,and screen and explain the best machine learning model.Methods The data of 1 692 elderly patients with hypertension were from the China Health and Retirement Longitudinal Study(CHARLS)database in 2018 and 2020.In the 2020 database,the Center for Epidemiologic Studies Depression Scale-10(CESD-10)scores of patients were collected.Patients with a CESD-10 score≥10 were classified as having depression,and were divided into the depression group(n=515)and the non-depression group(n=1 177)according to whether depression occurred.In the 2018 CHARLS database,the socio-demographic characteristics,behavioral factors and clinical characteristics of patients were collected;LASSO regression analysis was used to screen the characteristic variables.A total of 1 692 elderly patients with hypertension were randomly divided into the training set(n=1 184)and the validation set(n=508)according to the ratio of 7∶3.In the training set,decision tree,random forest,extreme gradient boosting tree,K nearest neighbor algorithm,support vector machine,Logistic regression and naive Bayesian models were constructed respectively,and then the optimal machine learning model was selected in the validation set.The interpretability of the optimal machine learning model was analyzed by Shapley additive explanations(SHAP).Results The results of LASSO regression analysis showed that 9 characteristic variables were finally selected,namely,place of residence,education level,retirement,self-perceived health status,Activity of Daily Living(ADL)Scale score,lung disease,headache,low back pain,and life satisfaction.Based on the above nine characteristic variables,the decision tree,random forest,extreme gradient boosting tree,K-nearest neighbor algorithm,support vector machine,Logistic regression and naive Bayesian models were constructed in the training set.In the validation set,the accuracy,sensitivity,accuracy,AUC of the Logistic regression model were the highest,the specificity of the support vector machine model was the highest,the F1 score of the naive Bayesian model was the highest,and the calibration AUC of the K-nearest neighbor algorithm model was the highest.Comprehensive evaluation,the Logistic regression model was the optimal machine learning model.The results of SHAP analysis showed that in the Logistic regression model,the characteristic variables with SHAP values from high to low were self-perceived health status(-0.201),life satisfaction(-0.194),education level(-0.173),low back pain(0.150),place of residence(0.134),retirement(-0.119),headache(0.096),ADL Scale score(0.093),lung disease(0.076).Conclusion In this study,seven machine learning prediction models for the risk of depression in elderly patients with hypertension in China were constructed based on self-perceived health status,life satisfaction,education level,low back pain,place of residence,retirement,headache,ADL Scale score,and lung disease,among which the Logistic regression model is optimal.