Genetic Variant-based Machine Learning Prediction of Diabetic Nephropathy: A Pilot Study in Brazilian Patients

Abstract Purpose Diabetic nephropathy (DN) is a major microvascular complication of type 2 diabetes, where genetic predisposition significantly influences susceptibility. Genetic variants that exacerbate oxidative stress and endothelial injury can disrupt glomerular microvascular integrity and permeability. Together, these molecular alterations may contribute to albuminuria and progressive deterioration of renal function, as reflected by declining estimated glomerular filtration rate (eGFR). Methods This pilot case-control study investigated the association of two genetic variants, MTHFR rs1801133 and VEGF-A rs28357093, with DN in 345 individuals from Central Brazil. Genotyping was performed using PCR-RFLP. Clinical and genetic data were analyzed using logistic regression under different inheritance models and five supervised machine learning (SML) algorithms (CART, RF, KNN, SVM, and LR). Model performance was assessed using accuracy, precision, sensitivity, specificity, ROC curve analysis, bootstrap resampling, and a sensitivity analysis excluding fasting blood glucose, serum creatinine, and eGFR. Results The MTHFR (rs1801133) and VEGF-A (rs28357093) variants were associated with DN risk ( p = 0.05 and p = 0.02, respectively). ML analyses demonstrated high predictive performance, with RF and CART achieving the highest discriminative ability among the evaluated algorithms. KNN and LR showed high sensitivity, reaching recall values of up to 1.00. eGFR and glycated hemoglobin emerged as the most influential predictors, whereas the genetic variants provided complementary predictive information. These findings highlight the potential of ML approaches for integrating clinical and genetic data in DN risk prediction. Conclusion ML models exhibited satisfactory predictive performance driven mainly by clinical parameters. The observed results support the potential utility of integrating genetic data into predictive workflows while highlighting the importance of larger cohorts for refining precision risk stratification in DN.