基于KNN算法与logistic回归的代谢综合征风险预测模型构建与对比研究

Construction and comparative study of metabolic syndrome risk prediction models based on KNN algorithm and logistic regression

目的构建基于K最近邻(KNN)算法和logistic回归的代谢综合征预测模型并比较两种模型对代谢综合征的预测效能。方法纳入6793例研究对象进行数据分析,构建基于KNN算法和logistic回归的预测模型,对模型进行内部验证及外部验证,采用多维度指标对预测性能进行评估,对比两种预测模型的预测效能。结果基于KNN算法预测模型的内部验证曲线下面积(AUC)为0.776(95%CI:0.764~0.788)、校准截距为0.028(95%CI:-0.031~0.089)、校准斜率为1.181(95%CI:1.106~1.257)、布里尔分数为0.157;外部验证AUC为0.780(95%CI:0.768~0.791)、校准截距为0.262(95%CI:0.207~0.317)、校准斜率为1.053(95%CI:0.990~1.117)、布里尔分数为0.167。基于logistic回归预测模型内部验证AUC为0.783(95%CI:0.772~0.795)、校准截距为-0.008(95%CI:-0.088~0.073)、校准斜率为0.995(95%CI:0.934~1.058)、布里尔分数为0.156;外部验证AUC为0.782(95%CI:0.771~0.793)、校准截距为-0.045(95%CI:-0.113~0.022)、校准斜率为1.006(95%CI:-0.011~1.063)、布里尔分数为0.164。结论在代谢综合征的风险预测上,logistic回归预测模型表现优于基于KNN算法预测模型。

Objective To construct the metabolic syndrome prediction models based on the K nearest neighbor(KNN)algorithm and logistic regression,and to compare the predictive efficiencies between the two methods.Methods The included 6793 study subjects conducted the data analysis.The prediction models based on KNN algorithm and logistic regression were constructed.The models conducted the internal validation and external validation.The multiple dimensions indicators were adopted to evaluate their predictive performances and the predictive efficiencies were compared between the two predictive models.Results The area under internal validation curve(AUC)of the prediction model based on the KNN algorithm was 0.776(95%CI:0.764-0.788),the calibration intercept was 0.028(95%CI:-0.031-0.089),the calibration slop was 1.181(95%CI:1.106-1.257)and the Brier score was 0.157.In the external validation,AUC was 0.780(95%CI:0.768-0.791),the calibration intercept was 0.262(95%CI:0.207-0.317),the calibration slop was 1.053(95%CI:0.990-1.117)and the Brier score was 0.167.AUC of internal validation in the prediction model based on the logistic regression was 0.783(95%CI:0.772-0.795),the calibration intercept was-0.008(95%CI:-0.088-0.073),the calibration slop was 0.995(95%CI:0.934-1.058),the Brier score was 0.156,the external validation AUC was 0.782(95%CI:0.771-0.793),the calibration intercept was-0.045(95%CI:-0.113-0.022),the calibration slope was 1.006(95%CI:-0.011-1.063)and the Brier score was 0.164.Conclusion The logistic regression prediction model performance is better than the prediction model based on the KNN algorithm.