Novel hierarchical support vector models with application to diabetes control

This study examines classification and prediction in non-separable, heterogeneous data settings, with emphasis on improving support vector–based methods. We consider support vector machines (SVM) and support vector regression (SVR) and demonstrate that classical kernel approaches can lose efficiency when population heterogeneity and imbalance are ignored. To address this limitation, we propose hierarchical stratified support vector models that incorporate subgroup structure through stratum-adaptive learning. Performance is evaluated using confusion matrix based indices together with novel weighted criteria for correct classification rate and mean squared error, ensuring alignment with stratified population targets. An application to diabetes control classification using glycated hemoglobin (HbA1c) data shows that the proposed hierarchical models consistently outperform classical SVM and SVR in accuracy and predictive error, while enabling scalable computation and meaningful stratum-specific inference.