AN OPTIMIZED RESIDUAL CONVOLUTIONAL NEURAL NETWORK FRAMEWORK FOR EARLY DIABETES PREDICTION USING CLINICAL PARAMETERS

Abstract

Diabetes mellitus is a chronic metabolic disorder with increasing global prevalence and serious health complications. Early prediction is important for prevention and effective management. This study developed an optimized Residual Convolutional Neural Network (RCNN) framework for diabetes prediction using clinical parameters from the Pima Indians Diabetes Dataset. Data preprocessing, normalization, and Synthetic Minority Over-sampling Technique (SMOTE) were applied to improve data quality and address class imbalance. The proposed RCNN model combines convolutional feature extraction with residual learning to identify complex nonlinear relationships among diabetes-related factors. Model performance was evaluated using accuracy, sensitivity, specificity, and Matthews Correlation Coefficient. The findings demonstrate that deep learning-based prediction can provide an effective computational approach for diabetes screening and clinical decision support. The proposed RCNN model achieved an accuracy of 96.41%, sensitivity of 96.17%, specificity of 97.53%, MCC of 0.94, and AUC of 0.99 on the testing dataset, outperforming conventional machine learning classifiers including Random Forest, Support Vector Machine, Decision Tree, and Extra Trees.