心血管病危险因素的风险特征分析及疾病预测模型研究

目的 探讨心血管病（cardiovascular disease, CVD）危险因素的风险特征重要度并优化CVD发生风险预测模型。方法 选取湖北省开展的“心血管病高危人群早期筛查与综合干预项目”中2015年10月—2020年11月纳入的初筛人群132 268例,采用随机梯度下降（SGD）、逻辑回归（LR）、随机森林（RF）、支持向量机（SVM）、K -近邻（KNN）和LightGBM等6种机器学习算法建立并优化CVD发生风险预测模型。结果 患有CVD人群的BMI（t = - 9.512,P＜0.001）、WC（t = - 11.476,P＜0.001）、SBP（t = - 38.533,P＜0.001）、DBP（t = - 21.571,P＜0.001）、脉压差（t = - 25.284,P＜0.001）、TC（t = - 6.616,P＜0.001）、LDL（t = - 7.374,P＜0.001）、TG（t = - 5.572,P＜0.001）、FBG（t = - 7.812,P＜0.001）等指标水平均高于健康对照人群。SGD、LR、RF、SVM、KNN和LightGBM 6种机器学习算法建立的CVD风险预测模型AUC值分别为0.835、0.828、0.851、0.852、0.836和0.871,且在LightGBM算法预测CVD风险模型中排名前七的危险因素依次为:SBP、脉压差、DBP、年龄、WC、LDL-C和PEF。在优化的CVD风险预测模型中,仅纳入年龄、SBP、DBP、WC和PEF这5个变量的AUC达到0.867,即可较准确地预测CVD发生的风险。结论 LightGBM机器学习算法是最优拟合预测模型,且仅纳入年龄、SBP、DBP、WC和PEF这5个变量CVD风险预测模型的预测性能良好。

Risk characterization of cardiovascular disease risk factors and disease prediction model study

Objective To explore the importance of risk characteristics of cardiovascular disease (CVD) risk factors and optimize the risk prediction model of CVD occurrence. Methods A total of 132 268 cases were selected from the initial screening population included in the Early Screening and Comprehensive Intervention Program for People at High Risk of Cardiovascular Disease in Hubei Province from October 2015 to November 2020, and stochastic gradient descent (SGD), logistic regression (LR), random forest (RF), support vector machine (SVM), K-nearest neighbor (KNN), and light gradient boosting machine (LightGBM) were used to build and optimize the risk prediction model for CVD occurrence. Results In CVD patients, the BMI (t=-9.512, P<0.001), WC (t=-11.476, P<0.001), SBP (t=-38.533, P<0.001), DBP (t=-21.571, P<0.001), pulse pressure difference (t=-25.284, P<0.001), TC (t= -6.616, P<0.001), LDL (t=-7.374, P<0.001), TG (t=-5.572, P<0.001), and FBG (t=-7.812, P<0.001) levels were higher than those of healthy controls. The AUC values of CVD risk prediction on models established by SGD, LR, RF, SVM, KNN and LightGBM were 0.835, 0.828, 0.851, 0.852, 0.836, and 0.871, respectively. The top seven risk factors in the prediction of CVD risk model by LightGBM algorithm were SBP, pulse pressure difference, DBP, age, WC, LDL-C, and PEF. In the optimized CVD risk prediction model with age, SBP, DBP, WC and PEF included had the AUC value of 0.867, which indicated accurate prediction of CVD risk. Conclusion The LightGBM machine learning algorithm is the best-fit prediction model, and the CVD risk prediction model incorporating only five variables, namely age, SBP, DBP, WC, and PEF shows good prediction performance.