Interpretable type 2 diabetes incidence prediction with AutoScore: A model based on standard clinical parameters

Accurate prediction of type 2 diabetes mellitus (T2DM) onset is critical to enable timely interventions and preventive strategies. Although machine learning (ML) approaches have shown promise in risk prediction, their complexity often limits clinical implementation. There is a need for interpretable, user-friendly models that retain predictive strength.

We studied 904 cardiovascular risk patients without T2DM at baseline, assessing 71 anthropometric, clinical, and laboratory variables. Over a four-year follow-up, 10 % developed T2DM. We applied AutoScore, an interpretable ML framework that generates parsimonious, point-based risk scores, and compared its performance with an optimized Support Vector Machine (SVM) with a linear kernel. The SVM was refined using feature selection, Tomek link removal, and up-sampling to address class imbalance.

Both approaches consistently identified fasting glucose, OGTT glucose, and the Matsuda index (reflecting glucose-insulin dynamics) as key predictors. The optimized SVM model achieved a higher balanced accuracy (75 % vs. 67 %), specificity (80 % vs. 77 %), and AUC (0.72 vs. 0.69) compared to AutoScore. However, AutoScore, other than the SVM model, relied exclusively on a small set of routinely available accessible parameters and thereby offered superior interpretability and ease of integration into clinical workflows. External validation in an independent cohort further confirmed the robustness of the AutoScore model.

Although black-box models such as SVM deliver slightly higher predictive accuracy, interpretable frameworks like AutoScore provide clinically actionable risk stratification based on standard data. Their transparency and simplicity make them particularly valuable for real-world decision support.