Predicting the incidence of type 2 diabetes using FINDRISC and machine learning approaches: evidence from a population-based cohort

Introduction Early identification of type 2 diabetes is vital for prevention. While the FINDRISC questionnaire offers a non-invasive screening method, machine learning models may offer alternative ways to analyze risk patterns. This study compares the predictive performance of the aggregate FINDRISC score and its individual components using both conventional and machine learning approaches in a population-based cohort. Methods This prospective population-based cohort study evaluated the performance of the aggregate FINDRISC score and its individual items for predicting 6-year incident type 2 diabetes using logistic regression and four machine learning algorithms (decision tree, random forest, support vector machine, and artificial neural network). SMOTE was applied strictly within training folds during cross-validation. Model discrimination and calibration were evaluated using the AUC-ROC, decile-based calibration plots, and the Hosmer–Lemeshow test. Results Among 8,859 participants with complete 6-year follow-up, 525 incident cases occurred (6.0% incidence). The aggregate FINDRISC score yielded an AUC of 0.76 (95% CI: 0.74–0.79). In component-level models utilizing continuous anthropometric features, Random Forest achieved the highest observed discriminative performance with an AUC of 0.81 (95% CI: 0.79–0.83). The screening threshold of > 12 yielded the highest Youden index (0.41; sensitivity 73.0%, specificity 68.4%), while a threshold of > 11 offered higher sensitivity (78.9%) with lower specificity (61.1%). Conclusion The aggregate FINDRISC score demonstrated acceptable discrimination for predicting 6-year incident type 2 diabetes. While component-level machine learning models yielded higher observed AUCs in internal validation, the clinical significance of these modest differences remains uncertain. Due to the need for further assessment of calibration, clinical utility, and external validation, the standard FINDRISC remains the practical tool for primary care screening.