Diabetes Prediction using Machine Learning and Deep Neural Models with Hybrid Resampling Techniques

Accurate diabetes prediction plays a vital role in early diagnosis and intervention, particularly in the context of growing global health concerns. However, medical datasets often suffer from class imbalance, where diabetic cases are significantly outnumbered by non-diabetic ones, limiting the effectiveness of traditional predictive models. This paper proposes a hybrid framework that integrates advanced resampling strategies with both machine learning and deep neural models to improve the detection of diabetes. The study explores three sampling methods SMOTE, ADASYN, and Borderline-SMOTE to address imbalance, and evaluates the performance of diverse classifiers including XGBoost, LightGBM, Naive Bayes, Random Forest, Gradient Boosting, 1D-CNN, and LSTM. The models were tested on original and balanced datasets, and assessed using accuracy, precision, recall, and F1-score. Results show that XGBoost consistently achieves the best predictive performance, attaining an F1-score of 85.0% and accuracy of 96.4% on the SMOTE-balanced dataset. Deep learning models, particularly 1D-CNN and LSTM, showed moderate but improved recall after resampling, confirming their applicability to structured medical data. Overall, the findings indicate that the proposed hybrid approach significantly enhances model sensitivity and offers a practical and scalable framework for early diabetes detection using imbalanced healthcare datasets.