Long-term type 2 diabetes risk prediction using machine learning with childhood, adolescent, and genetic factors

Abstract Late diagnosis of type 2 diabetes increases patient morbidity and healthcare burden, yet current prediction models use adult risk factors to estimate 5–10-year risk. In this manuscript, we evaluate long-term prediction of type 2 diabetes over 38 years using childhood factors and polygenic risk scores (PRSs). Data from the longitudinal Cardiovascular Risk in Young Finns Study were analyzed. Childhood features—including anthropometric, demographic, lifestyle, parental characteristics, blood biomarkers, and PRSs—were used to train machine learning models to predict type 2 diabetes. Multiple feature selection, imputation, and classification algorithms were applied. Results were validated using nested cross-validation and assessed by the area under the receiver operating characteristic curve (AUROC). Added value of genetic data was evaluated using category-free net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Feature importance was assessed with Shapley additive explanations (SHAP). Diabetes status was available for 2,144 of 3,596 participants; 249 (11.6%) developed type 2 diabetes, with a mean onset age of 44.3 years. Random forest achieved the highest performance in distinguishing high-risk individuals (AUROC 0.737 [95% CI, 0.706–0.766]). Including PRSs improved prediction (NRI 0.201 [95% CI, 0.121–0.280]; IDI 0.014 [95% CI, 0.002–0.025]). The top five predictors were the PRS for type 2 diabetes, maternal body mass index (BMI), PRS for BMI, subscapular skinfold thickness, and C-reactive protein. Combining childhood, adolescent, and genetic data with machine learning, particularly tree-based ensemble methods such as random forest, enables long-term prediction of type 2 diabetes risk and improves early identification of high-risk individuals to institute lifelong risk modification strategies.