Leveraging Ensemble Learning for Multiclass Diabetes Prediction from Health Indicators

Diabetes is a primary global health concern that can lead to severe long-term complications without timely detection and management. In this study, a three-class classification model (non-diabetic, prediabetic, diabetic) is proposed to enhance early risk prediction and intervention. A large dataset encompassing demographic, clinical, and lifestyle indicators was utilized, with categorical features one-hot encoded and numeric attributes scaled via Min-Max normalization. The Synthetic Minority Over-Sampling Technique (SMOTE) was employed to address class imbalance, ensuring adequate representation of minority classes. Multiple machine learning algorithms, including ensemble methods, were evaluated through rigorous training and cross-validation. The Extra Trees classifier achieved the highest performance, demonstrating 93.96% accuracy and an AUC of approximately 98.62%. These findings underscore the effectiveness of robust preprocessing and ensemble-based strategies in improving diabetes classification. Future work involves exploring enhanced interpretability tools and validating the model in broader clinical environments to ensure sustained accuracy and real-world applicability.