Optimizing the Performance of Diabetes Risk Prediction Using Ensemble Learning Techniques

Early detection is a critical strategy for reducing diabetes prevalence across all age groups. However, given the likelihood of delays in accessing medical services for various reasons, the development of user-friendly machine-learning models becomes essential. The study aims to develop a diabetes risk prediction optimized model in individuals using modifiable behavioral risk factors. Advanced data pre-processing approaches were used to increase the data quality and the prediction performance of the trained models was improved using several ensemble methods to combine best-performing ML models. Outlier detection using the iForest approach, imbalance multi-class handling, and selection of relevant behavioral risk factors were achieved at the pre-processing stage. The development stage of the model starts with training some supervised classifier models, followed by applying different ensemble methods for model performance optimization. Out of which the stacking ensemble approach performs best with a K Fold score of 97.87%, average accuracy of 99.40%, ROC AUC score of 99.70%, and F1-score of 98.41% and the optimized model was deployed on a web-based interface for real-time diabetes risk assessment. This study contributes a powerful prognostic framework, which significantly advances diabetes risk prediction and proactive healthcare management.