Logistic回归模型和反向传播神经网络模型在高尿酸血症影响因素分析中的应用

Application of logistic regression model and back propagation neural network model in the analysis of influencing factors for hyperuricemia

目的探讨logistic回归模型与反向传播神经网络模型在高尿酸血症(HUA)影响因素分析中的应用价值。方法采用logistic回归模型和反向传播神经网络模型分析2867例体检患者诊断HUA的影响因素,采用ROC曲线检验比较两种方法在HUA筛查中对HUA发生的诊断效能。结果Logistic回归模型和反向传播神经网络模型分析显示,影响HUA发生的前五位因素均为性别、SCr、BMI、TG和LDL-C。Logistic回归模型诊断HUA的AUC为0.741[95%CI(0.732~0.749)],0.76为最佳诊断界值,其相应的诊断灵敏度为70.35%,特异度为66.25%。反向传播神经网络模型诊断HUA的AUC为0.742[95%CI(0.733~0.750)],取最佳诊断界值为0.74时,对应的诊断灵敏度为69.45%,特异度为67.37%。结论Logistic回归模型与反向传播神经网络模型分析HUA发生影响因素时,其诊断HUA的灵敏度和特异度相仿,在实际应用中应根据具体数据的特征选用。

Objective To explore the application value of logistic regression model and back propagation neural network model in the analysis of influencing factors for hyperuricemia.MethodsThe influencing factors for hyperuricemia occurrence were analyzed using both models of logistic regression model and back propagation neural network model in 2867 patients with hyperuricemia diagnosed during physical examination.The ROC curve analysis was used to compare the diagnostic efficacy of the two models in the diagnosis of hyperuricemia during screening of HUA.Results The analysis of two models showed that the top five factors affecting hyperuricemia occurrence were gender,SCr,BMI,TG and LDL-C.For the diagnosis of hyperuricemia,AUC for logistic regression model was 0.741[95%CI(0.732-0.749)]and that for back propagation neural network model was 0.742[95%CI(0.733-0.750)].Taking 0.76 as the best cut-off value for the diagnosis of hyperuricemia using logistic regression model,the sensitivity was 70.35%and the specificity was 66.25%.Taking 0.74 as the best cut-off value for the diagnosis of hyperuricemia using back propagation neural model,the sensitivity was 69.45%and the specificity was 67.37%.ConclusionThe sensitivity and specificity of logistic regression model in the diagnosis of hyperuricemia are similar to those of back propagation neural model.In practice,a suitable method should be selected according to the characteristics of the data.