430-P: Development and Internal Validation of Machine-Learning Algorithms for Diabetic Kidney Disease Model of People with Type 2 Diabetes Mellitus

Objective: This study aims to establish DKD early screening model through different machine learning algorithms according to the clinical characteristics and biological indicators of type 2 diabetes patients, and compare its predictive value for DKD. Methods: We selected 74,982 type 2 diabetes patients hospitalized in Beijing Tongren Hospital from 2013 to 2022. The collected-data included common clinical characteristics and biological indicators. 625 DKD cases was diagnosed with the increase of urine microalbumin or the decrease of eGFR and diabetes retinopathy, and 1336 cases were diagnosed as Non-DKD. Establish DKD early screening models through 7 machine learning algorithms: Deep Neural Network(DNN), RF, SVM, XGBoost, GMB, DRF, Naive Bayes. Result: DNN were established for different feature dimensions: D0: blood routine, AUC=0.6911; D1: blood routine, DBP, SBP, WHR, AUC=0.7919; D2 : D1 dataset, biochemical indicators, AUC=0.8517. Evaluating the accuracy of the model, AUC of DNN is 0.902, while other algorithms have AUC between 0.81 and 0.88. Conclusion: The best prediction result is D2 dataset, which is the most effective combined with biochemical indicators to predict DKD; The single blood routine index also has important value in predicting the development of DKD. DNN is the the best model compared to other machine learning algorithms. L. Zhang: None.