DHT-Stack: A Deep-Hybrid Transformer Stacking Framework for Uncertainty-Aware and Cost-Sensitive Dengue Screening

Abstract

Contributed to the development of DHT-Stack, a Deep-Hybrid Transformer Stacking framework designed for early and reliable dengue screening using routine hematological and clinical blood parameters. The research addresses critical challenges in traditional machine learning-based dengue screening, including class imbalance, false negative predictions, model interpretability, and prediction uncertainty. Key Research Contributions: Developed a hybrid stacking ensemble architecture integrating TabPFN, CatBoost, LightGBM, and XGBoost as Level-0 base learners. Applied SMOTEENN for handling class imbalance and improving the quality of training data. Utilized Boruta feature selection to identify 17 significant hematological biomarkers from 18 clinical parameters. Implemented cost-sensitive learning with a false-negative penalty of 5.0 to prioritize patient safety and reduce missed dengue-positive cases. Designed a Level-1 Ridge Logistic Regression meta-learner using out-of-fold predictions and stratified 10-fold cross-validation. Incorporated model disagreement-based uncertainty quantification to identify predictions requiring additional clinical review. Applied SHAP for global feature importance analysis and LIME for local prediction explanations. Evaluated model performance against standalone machine learning models using accuracy, precision, recall, and AUC-ROC metrics.

Key contributions

1. 1Experimental Results:
2. 2Achieved 96.92% classification accuracy.
3. 3Obtained 98.54% recall for identifying dengue-positive cases.
4. 4Achieved an AUC-ROC score of 0.9948.
5. 5Reduced false-negative cases to 1.48% in the reported confusion matrix.

Cite this paper

Jihadul Islam, Tanvir Anjum Rahat, Mohammed Sakib Hasan, Md Abdul Masud, "DHT-Stack: A Deep-Hybrid Transformer Stacking Framework for Uncertainty-Aware and Cost-Sensitive Dengue Screening," in 2026 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON), 2026.