Abstract
INTRODUCTION: Timely and accurate diagnosis of diabetes mellitus remains a pending challenge due to the diversity of patient data and the limitations of traditional screening methods. OBJECTIVE: To propose a hybrid prediction framework incorporating Convolutional Neural Networks (CNNs) and Integrated Learning with a soft voting strategy to improve the accuracy, robustness and interpretability of diabetes diagnosis. METHODS: The model was evaluated on two publicly available datasets-the UCI Pima Indians Diabetes dataset (768 samples, 8 features), the same dataset used to describe the Pima Indians (2,000 samples, 8 features) and the Tianchi Medical dataset (5,642 samples, 41 features). After missing-value imputation, z-score standardization, and min-max normalization, CNNs were used for deep feature extraction, followed by integration with multiple classifiers-Logistic Regression (LR), Support Vector Machines (SVM), Random Forest, AdaBoost, XGBoost, LightGBM, and CatBoost-via a weighted soft voting scheme. Training and testing sets were split 75:25, and hyperparameters for each classifier were tuned through grid search. RESULTS: The proposed CNN-Voting integrated model consistently outperforms the individual models, achieving up to 98% accuracy, 0.99 F1 value and 99% recall on the largest dataset. Feature importance analysis revealed that blood glucose, body mass index (BMI), age, and urea were the features with the most predictive value, which was highly consistent with common knowledge in clinical medicine. CONCLUSION: This hybrid model not only improves predictive performance and generalisability, but also provides a scalable and interpretable solution for clinical decision support in diabetes management.