38-PUB: Light Gradient Boosting Machine Models Predict Retinopathy and Nephropathy in People with Type 2 Diabetes

Introduction and Objective: To examine the ability of light gradient boosting machine (GBM) models using machine learning for predicting retinopathy and nephropathy in people with type 2 diabetes. Methods: Data from diabetes specialists’ patient registries in Japan on complication of diabetes from 2009 to 2024 were used for 5,645 people who had laboratory data and medication history. Area under the receiver operation characteristic curves (AUCs) for retinopathy and nephropathy were calculated and comparing the ability of light GBM and logistic regression models by the Delong test for retinopathy and nephropathy. Results: During the study period 2,326 retinopathy and 3,226 nephropathy events occurred. The AUCs according to light GBM were significantly higher than those of the logistic regression model (retinopathy: AUC of light GBM: 0.71, AUC of logistic regression: 0.69, p<0.001; nephropathy: AUC of light GBM: 0.71, AUC of logistic regression: 0.68, p<0.001). Shapley additive explanations revealed that estimated glomerular filtration rate (eGFR), systolic blood pressure (SBP), and HbA1c were important for retinopathy, while duration of diabetes, eGFR, and Hba1c were important for nephropathy (Figure). Conclusion: Assistance by machine learning may be efficient and useful for predicting diabetic retinopathy and nephropathy. Disclosure K. Fujihara: None. S. Minusa: Employee; Current; Hitachi, Ltd. T. Osaki: Employee; Current; Hitachi, Ltd. H. Ban: Employee; Current; Hitachi, Ltd. H. Maegawa: None. H. Sone: None.