Construction Of Risk Prediction Model Of Type 2 Diabetic Kidney Disease Based On Deep Learning

Objective:Based on the physical examination data set of Type 2 diabetes mellitus(T2DM)patients for several years,combined with the long short term memory(LSTM)neural network of deep learning technology,the prediction model of the future risk of diabetic kidney disease(DKD)is constructed,and the high risk groups of DKD are accurately screened before the kidney structural damage occurs,so as to provide a basis for early intervention for key groups.Methods:This study was a retrospective study.First of all,through the review of relevant literature at home and abroad and expert group interviews to determine the predictive risk factors of DKD,and then according to the standard based on the Lee’s joint clinic in Taiwan Province of China for 7 consecutive years of physical examination data collection,including:general data of patients,laboratory test results,urine routine examination,Hemoglobin A1c(HbA1c),systolic blood pressure(SBP)and Pulse Pressure(PP)variability during follow-up.After completing the data collection,the SPSS 22.0 software is used for correlation analysis,and the characteristic parameters that finally need to be included in the neural network are determined according to the correlation coefficient of each variable,and the invalid data is eliminated,and then the data that meets the requirements is converted to become the data that meets the input requirements of the neural network.Then,according to the DKD diagnosis standard,the data is divided into patients with T2DM without DKD data set and patients with T2DM with DKD data set.After data extraction and preprocessing by MATLAB software,training group and test group are established for model simulation training.The preprocessing steps include label definition,zero-score normalization analysis and standardized processing.Use pytorch language to design LSTM neural network,select Dropout algorithm to prevent over-fitting of the model,use mean square error function as loss function in the selection of network parameters,Adam algorithm to optimize the model and finally complete the construction of the network,and then use pytorch language to train the network and output the results.And compared with the prediction model constructed by SVM algorithm for precision,accuracy,recall rate and area under the curve(AUC)of receiver operating characteristic curve(ROC)model evaluation and comparative analysis,and based on LSTM to establish three models to explore the impact of HbA1c,SBP and PP variability on the overall performance of the model.Results:A total of 6040 patients with T2DM were included in the physical examination data set,including 4228 in the training set and 1812 in the test set of T2DM patients.In this study,a total of four DKD risk prediction models are constructed,in which the Loss curve of the LSTM model with all the variability parameters can converge quickly and is more stable compared with the curves of other models.It is the optimal prediction model in this study.The precision,accuracy,recall rate and AUC of the model are 77%,86%,76%,and 0.83,respectively.The precision,accuracy,recall rate and AUC of the model based on SVM algorithm are 63%,76%,52%,and 0.73,respectively.Compared between the two groups,the precision(P<0.001),accuracy(P<0.001),recall rate(P<0.001),and AUC(P=0.002)based on LSTM neural network were significantly higher than those of the prediction model based on SVM algorithm.In addition,the precision,accuracy,recall rate and AUC of the LSTM prediction model that did not include HbAlc variability were 64%,78%and 61%and 0.72,respectively.Compared with the optimal LSTM model with all variability parameters,the precision(P<0.001),accuracy(P<0.001),recall rate(P<0.001)and AUC(P<0.05)of the optimal LSTM model were significantly better than those of the LSTM prediction model without HbAlc variability.The precision,accuracy,recall rate and AUC of the LSTM prediction model without SBP variability were 65%,79%,65%and 0.75,respectively.Compared with the optimal LSTM model with all variability parameters,the precision(P<0.001),accuracy(P<0.001),recall rate(P<0.001)and AUC(P<0.05)of the optimal LSTM model were significantly better than those of the LSTM prediction model without SBP variability.The precision,accuracy,recall and AUC of the LSTM prediction model without PP variability were 70%,81%,67%and 0.77,respectively.Compared with the optimal LSTM model with all variability parameters,the precision(P<0.001),accuracy(P<0.001),recall rate(P<0.001)and AUC(P<0.05)of the optimal LSTM model were significantly better than those of the LSTM prediction model without PP variability.Conclusion:1.In this study,the LSTM neural network based on deep learning technology successfully constructed the risk prediction model of type 2 DKD,and its precision,accuracy,recall rate and AUC were significantly better than the prediction model constructed by SVM algorithm.2.The construction of DKD risk prediction model based on LSTM neural network needs to include HbAlc variability,SBP variability and PP variability as important characteristic parameters,which can further improve the overall performance of the model.3.The successful construction of DKD risk prediction model based on LSTM neural network provides a basis for early intervention for key groups,which is of great significance for improving the accuracy of early intervention,effectively preventing or delaying the occurrence of DKD,improving the quality of life of diabetic patients,and reducing medical expenses,but also provides a new perspective for other chronic disease prediction model construction methods.