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

Area of Science:

• Cardiovascular Medicine
• Biostatistics
• Machine Learning in Healthcare

Background:

• Accurate prediction of type 2 diabetes mellitus (T2DM) onset is crucial for effective intervention.
• Existing machine learning (ML) models often lack the interpretability needed for clinical implementation.
• There is a need for user-friendly, interpretable ML tools that maintain predictive accuracy.

Purpose of the Study:

• To evaluate the performance of AutoScore, an interpretable ML framework, for predicting T2DM onset.
• To compare AutoScore with an optimized Support Vector Machine (SVM) using clinical and laboratory variables.
• To assess the clinical utility and interpretability of ML models for T2DM risk stratification.

Main Methods:

• A cohort of 904 cardiovascular risk patients without T2DM at baseline was studied.
• 71 anthropometric, clinical, and laboratory variables were assessed over a four-year follow-up.
• AutoScore was applied and compared against an optimized linear kernel SVM, with SVM refined for class imbalance.

Main Results:

• Fasting glucose, OGTT glucose, and Matsuda index were key predictors for both models.
• The optimized SVM showed higher balanced accuracy (75% vs. 67%) and AUC (0.72 vs. 0.69).
• AutoScore offered superior interpretability and used routinely available parameters, with robust external validation.

Conclusions:

• Interpretable ML frameworks like AutoScore provide clinically actionable T2DM risk stratification.
• AutoScore's transparency and simplicity are valuable for real-world clinical decision support.
• While SVMs offer slightly higher accuracy, AutoScore's interpretability facilitates easier clinical integration.