Deep learning in diabetes prediction: the impact of data balancing techniques

This study addresses the data imbalance problem in diabetes prediction and systematically evaluates the impact of seven balancing techniques on the performance of deep learning models. The results show that the SMOTE+Tomek technique is the most effective and significantly improves the recognition rate and overall prediction performance of diabetic patients. The multilayer perceptron achieves the best balance between computational efficiency and prediction capability, while the hybrid CNN+LSTM model is slightly better in prediction accuracy. Feature significance analysis reveals that fasting glucose, glycated haemoglobin and body mass index are the key predictors, and the balancing technique enhances the interaction effect among features. The research results provide a more accurate methodological framework for clinical diabetes risk assessment, which is important for early intervention and screening.