Class imbalance remains a critical challenge in medical data classification, particularly in diabetes prediction, as it significantly degrades minority-class sensitivity. This study proposes an Adaptive Feature-Aware Hybrid Resampling Strategy (AHRS) that dynamically integrates oversampling and undersampling based on Imbalance Ratio (IR) and Feature Importance (FI). Unlike conventional static resampling methods, AHRS iteratively adjusts class distribution while preserving informative feature structures. In addition, this study introduces the Integrated Balanced Index (IBI), a bounded composite metric integrating precision, recall, and specificity to provide a fairer evaluation of classification performance on imbalanced medical datasets. The proposed approach was evaluated using the Pima Indian Diabetes Dataset (768 instances) with K-Nearest Neighbor, Naïve Bayes, and Random Forest classifiers under 5-fold stratified cross-validation. Experimental results demonstrate that AHRS consistently outperforms SMOTE, Random Oversampling, and Tomek Links, achieving accuracy improvements of 5–7% and recall gains of up to 10%. Random Forest combined with AHRS achieved the highest IBI score of 0.90, indicating strong balance between sensitivity and specificity. The findings suggest that adaptive, feature-aware resampling combined with balanced evaluation metrics provides a reliable and interpretable framework for fair medical classification systems and Clinical Decision Support Systems (CDSS).
Jasmir et al. (Wed,) studied this question.