<![CDATA[Primary healthcare facilities often face significant technical challenges in the early detection of diabetes complications due to their heavy reliance on micro-datasets characterized by extreme class imbalances. Such data constraints typically make standard single-prediction models prone to overfitting and low sensitivity, which are critical issues in clinical decision-making. This study proposes a multi-level optimization framework to overcome these limitations by integrating Information Gain for dimensionality reduction, Synthetic Minority Over-sampling Technique (SMOTE) to handle class imbalance, and Adaptive Boosting (AdaBoost) with Categorical Naive Bayes as the base learner. To address the critical aspect of model validity on a highly limited dataset (n=20), this methodology employs a stratified cross-validation scheme to ensure reliable performance estimation and minimize prediction bias. The integration of Information Gain effectively optimized the model by reducing the predictor variables from seven to four optimal features, thereby decreasing computational complexity. Based on rigorous testing, the proposed ensemble architecture achieved a robust average accuracy of 87.00% and, more importantly, a high Recall (Sensitivity) of 90% in identifying positive complication cases. This research provides a significant scientific contribution to machine learning development by demonstrating that the systematic integration of SMOTE and AdaBoost offers a stable, precise, and computationally efficient solution for predictive modeling in data-constrained clinical environments. These findings suggest that the proposed model can serve as a reliable early warning system to improve patient outcomes in primary care settings.]]>
