<![CDATA[Type 2 Diabetes Mellitus (T2DM) is a critical global health challenge, with 589 million adults currently diagnosed and projected to reach 853 million by 2050. Early detection is crucial, as it can reduce complication incidence by 30-40%, yet approximately 50% of cases remain undiagnosed. While machine learning approaches demonstrate promise for T2DM risk prediction, current systems face a fundamental accuracy-interpretability paradox: deep learning models achieve high accuracy (88-93%) but lack clinical transparency, while interpretable models sacrifice predictive performance. This study develops and validates a hybrid CNN-Fuzzy Logic system that directly addresses this paradox by combining high predictive accuracy with clinical interpretability. The system employs a Convolutional Neural Network component for non-linear feature abstraction combined with Mamdani Fuzzy Logic incorporating clinically derived weights aligned with ADA 2024 diagnostic criteria. Tested on the Pima Indian Diabetes dataset (n=154 test cases), the hybrid model achieved 92.5% accuracy (95% CI: 88.2-96.1%), 91.2% sensitivity, 93.1% specificity, and AUC-ROC 0.925, statistically superior to standalone CNN (88.9%, p=0.0037) and Fuzzy Logic (88.3%, p=0.0015) approaches. Interpretability scores reached 0.78-0.86, exceeding pure neural network baselines (0.32-0.42) and supporting clinician-understandable risk stratification. The system is operationalized as a web-based Clinical Decision Support System supporting both individual patient assessment and batch population screening. This hybrid architecture directly bridges the accuracy-interpretability paradox that has historically constrained ML adoption in clinical diabetes management.]]>
