EDRP — Diabetes Risk Prediction
Explainable stacking ensemble for diabetes risk prediction on the real-world DiaBD cohort (5,288 samples). Published at IEEE QPAIN 2026.
Overview
Explainable diabetes risk prediction on a Bangladeshi dataset of 5,288 samples (DiaBD cohort), published at IEEE QPAIN 2026 (DOI 10.1109/QPAIN69676.2026.11545949).
My role
Lead author
Problem
Early diabetes risk stratification from clinical records, where positive cases are under-represented.
Approach
Paper: categorical encoding, normalization, feature engineering and hybrid sampling feed a stacking classifier built on Logistic Regression, CatBoost, Gradient Boosting and XGBoost, trained on an 80/20 split and compared against popular ML models; SHAP and LIME explain individual predictions. The repository additionally contains a 1D CNN baseline trained with focal loss and SMOTE-ENN.
Results
- 95.78% accuracy, outperforming the other models tested.