Research
Applied deep learning research in remote sensing change detection and clinical diagnostic systems, based at the Visual Data Analysis Lab (VDAL), Green University of Bangladesh.
Explainable Ensemble Modeling for Diabetes Risk Prediction Using the Real-World DiaBD Cohort
Md. Ittesaf Hossain, Md. Rajibul Palas, Mozdaher Abdul Quader, Sabbir Hosen Mamun, Md. Sharifur Rahman
EDRP (Ensemble-based Diabetes Risk Prediction): an explainable stacking classifier built on Logistic Regression, CatBoost, Gradient Boosting and XGBoost, predicting diabetes on a Bangladeshi dataset of 5,288 samples. Uses categorical encoding, normalization, feature engineering and hybrid sampling, with SHAP and LIME for transparency.
- 95.78% accuracy, outperforming the other models tested.
- 80/20 train–test split on 5,288 samples from the DiaBD cohort.
- SHAP and LIME explanations for each prediction.
Explainable Ensemble Machine Learning Framework for Accurate Detection of Thyroid Disorders
Md. Ittesaf Hossain, Nur Hossain, Md. Mustakim Hossain, Md. Romzan Alom, Kabiratun Ummi Oyshe, Md. Ahsan Habib, Muhammad Aminur Rahaman, A B M Shawkat Ali
ETDD (Ensemble-based Thyroid Disorder Detection): an explainable stacking classifier built on Random Forest, CatBoost, LightGBM and XGBoost, detecting thyroid disorders from a 9,172-sample medical dataset. Includes missing-data handling, categorical encoding, normalization and feature engineering, with SHAP and LIME for transparent predictions. Joint work between Green University of Bangladesh and BUBT.
- 96.24% accuracy and 99.03% AUC, outperforming the other models tested.
- 80/20 train–test split on 9,172 samples, benchmarked against several leading ML models.
- SHAP and LIME explanations for each prediction.
RDWS-EKSD: ROI-Guided Dynamic Weighted Stacked Ensemble with Grad-CAM++ for Explainable Kidney Stone Detection
Md. Mustakim Hossain, Md. Ittesaf Hossain, Nur Hossain
An ROI-guided, dynamically weighted stacked ensemble for explainable kidney stone detection from coronal CT images. Four ImageNet-pretrained backbones (InceptionV3, ResNet50, MobileNetV2, EfficientNetB0) are combined through learnable dynamic weighting and a meta-learner, with kidney-specific ROI localization and Grad-CAM++ explanations.
- 99.71% accuracy on the held-out test set.
- Evaluated on a public dataset of 1,799 coronal CT images from 433 patients (790 kidney stone, 1,009 normal).
- ROI localization, enhancement, augmentation and weighted sampling to improve anatomical focus and reduce class imbalance.
- Grad-CAM++ visual explanations highlight the discriminative kidney regions behind each prediction.
An Adaptive and Intelligent Ensemble Framework for Efficient Sudoku Solving: Integrating Classical CSP Algorithms
Md. Ittesaf Hossain, Tanha Tabassum Tusmi, Mayeesha Farjana
IECSP: a constraint-satisfaction solver that picks a solving strategy per puzzle from real-time complexity analysis (empty-cell ratio and constraint density), combining backtracking, BFS, DFS, heuristic search, MRV, forward checking and arc consistency, with an AC-3 and naked/hidden-singles pre-processing pipeline.
- Solved all 70 puzzles across four difficulty levels (2,800 experiments) with a 100% success rate.
- Better computational efficiency and scalability than the single-strategy baseline.