Objective To investigate the risk factors of early diabetic nephropathy and construct a risk prediction model based on BP neural network algorithm.Methods A total of 1 048 diabetic patients admitted to Yongkang Hospital of Traditional Chinese Medicine from January 2020 to December 2022 were retrospectively analyzed,including 115 diabetic nephropathy pa-tients(10.97％),and were divided into the DKD group(115 diabetic nephropathy patients)and the DM group(933 diabetic patients).The relevant data of patients were collected and matched according to 1∶1 nearest proximity method after the confoun-ding factors were excluded by propensity score matching(PSM).The prediction model was built based on the correlation factors by using the statistically significant indicators in the single factor analysis and BP neural network algorithm.Mean absolute error(MAE)was used to analyze the model efficacy,the predictive value of the risk prediction model was evaluated by receiver operat-ing characteristic curve(ROC),and external validation was performed.The model consistency was evaluated by calibration curve.Results The confounding factors were gender,hypertension,fasting blood glucose and uric acid.After the modeling set was 1∶1 and PSM was performed by the nearest method,the confounding factors were excluded:95 cases in DKD group and 95 cases in DM group.Univariate results indicated that there were significant differences in age,type 2 diabetes,total cholesterol,urinary protein excretion rate,diabetes course,and cystatin C(CysC)between groups(P＜0.05).The prediction accuracy was BP neural network algorithm,decision tree,support vector machine and logistic regression in the descending order.The results of BP neural network showed that the top 4 factors affecting the occurrence of early diabetic nephropathy were proteinuria excretion rate,age,diabetes course and cystatin C(CysC)in order.The AUC of the prediction model was 0.959(95％CI:0.917-1.000),the Yoden index was 0.867,and the corresponding sensitivity and specificity were 0.867 and 1.000,respectively.The external validation AUC was 0.958(95％CI:0.922-0.995),and its sensitivity and specificity were 0.804 and 1.000,respec-tively.The calibration curve in the calibration diagram was close to the standard curve.Conclusion The BP neural network al-gorithm model based on machine learning,which takes age,disease course,urinary protein excretion rate,TC,CysC and type 2 diabetes as predictive features,has good predictive value for early diabetic nephropathy,and can be clinically applied to the management and identification of high-risk population.
back propagation(BP)neural networkdiabetic nephropathyearly kidney diseasepredictive modelinflu-encing factor
NETL
NSTL
维普
万方数据
