Improving Early Detection of Type 2 Diabetes from Primary Care Records with Sparse-Balanced SVM

Early detection of Type 2 Diabetes (T2D) from primary-care Electronic Health Records (EHRs) is challenged by high-dimensional features, class imbalance, and limited model interpretability. We introduce a Sparse-Balanced Support Vector Machine (SB-SVM) that combines sparsity-promoting regularization with class-dependent weighting to enhance detection of the minority class while maintaining clinical interpretability. Using the FIMMG primary-care EHR dataset from Italian general practitioners, we tested SB-SVM in three progressively complex scenarios. We compared it with linear/gaussian SVM, KNN, Decision Tree, Random Forest, and deep models (MLP, DBN). Performance was evaluated using stratified cross-validation, with AUC and recall reported. Sparsity was measured using the norm. Training, validation, and testing efficiency were analyzed. SB-SVM achieved mean AUCs of 0.91, 0.81, and 0.69 across the three cases, with higher recall than most baselines. Gains in recall and AUC were statistically significant compared to most competitors (p < 0.05), though differences with Decision Tree and Random Forest were not always significant. The model produced sparse, interpretable coefficients ( = 0.39, 0.91, 0.57), consistently highlighting clinically relevant predictors (e.g., HbA1c, age, renal function, hypertension, and antidiabetic prescriptions). SB-SVM also showed lower runtime than ensemble and deep models, supporting real-time applications. By combining class balancing and sparsity within a linear margin-based classifier, SB-SVM offers accurate, interpretable, and computationally efficient T2D risk prediction suitable for integration into Clinical Decision Support Systems in primary care.