Scientific Contributions
Exploring Machine Learning, Deep Learning, NLP, XAI, and Computer Vision.
Boosted Forest Soft Ensemble of XGBoost, Gradient Boosting, and Random Forest with Explainable AI for Thyroid Cancer Recurrence Prediction
Developed a Boosted Forest Soft Ensemble model with Explainable AI that achieved 98.70% accuracy in predicting thyroid cancer recurrence.
A Multi-Class Heart Disease and Stroke Risk Classification Framework for Combined Assessment: A Machine Learning Approach
Developed an integrated and explainable multi-disease ML system using LightGBM to jointly predict heart disease and stroke risks with 89% accuracy.
Sleep Disorder Classification Using HSE-NET Ensemble: A Robust Machine Learning Framework
Built HSE-NET, a stacked ensemble system for scalable sleep disorder classification, achieving 96% accuracy and outperforming prior models.
XSG-Ensemble: A Weighted Voting Ensemble Machine Learning Model for Enhanced Chronic Kidney Disease Prediction
Developed XSG-Ensemble, a weighted ML model for early CKD prediction, achieving ~98% accuracy and 99.85% AUC with clinical interpretability.
Obesity Detection using Traditional Machine Learning and Clinical Interpretation with Explainable AI
Built an interpretable multi-class ML system for obesity level prediction using lifestyle features, achieving up to 98% accuracy with XGBoost and LightGBM.
Privacy-Preserving Fetal Health Classification Using Federated Learning with XGBoost and Random Forest
Introduces a privacy-preserving federated ensemble learning framework for fetal health classification without sharing sensitive patient information.
Predicting Alzheimer’s Disease Using Hybrid Ensemble Learning and SHAP-Based Interpretability
Hybrid ensemble framework integrating stacking and bagging for early Alzheimer’s prediction, achieving 95.36% accuracy with SHAP interpretability.
Transformer-Based Multi-Class Categorization of Suicide Triggers in Bangla Headlines
Developed a fine-grained Bangla suicide news classification system using transformer models, achieving 92.92% accuracy and 0.92 F1-score.
Transformer-Assisted Categorization and Scope Detection of Bangla News Headlines
Developed a dual-task Bangla news classification system achieving up to 92% accuracy to categorize headlines by topic and geographic scope.
Automated Detection of Religious Aggression in Bengali Social Media: A Comparative Study
Comparative study of ML, DL, and Transformer approaches for detecting religious aggressiveness, enabling culturally sensitive content moderation.
Explainable AI-Driven Prediction of Colorectal Cancer Recurrence using Machine Learning and Deep Learning Models
Developed an explainable ML and DL system to predict colorectal cancer recurrence using structured clinical data. Achieved high performance with LightGBM (94.62% accuracy), XGBoost (97.50% recall), and SHAP-based feature interpretation highlighting tumor aggressiveness, stage, and age as key predictors.
AppleNet-B1: Explainable Auto-Reweighted EfficientNet for Apple Disease Detection
Developed Adaptive AppleNet-B1, an auto-reweighted EfficientNet-B1 model for detecting six apple diseases under real farm conditions. Implemented a two-stage dynamic fine-tuning strategy to handle class imbalance and visual variation, achieving 98.94% accuracy and 99.96% ROC-AUC. Explainability using Grad-CAM highlighted disease-affected regions consistent with field symptoms, supporting reliable and interpretable diagnosis for practical agricultural use.
Class-Weighted ResNet50 for Robust Guava Leaf Disease Classification under Real-Field Conditions
Evaluated deep learning models for guava leaf disease classification under real-field conditions using class-weighted training to address dataset imbalance. ResNet50 achieved the best performance, improving accuracy from 91.88% to 97.35%, with significant gains across EfficientNet variants and Vision Transformer. The approach supports reliable early disease detection for practical agricultural applications.
TomatoNet-R50: A ResNet50-based Deep Learning Architecture for Tomato Leaf Disease Detection
Proposed TomatoNet-R50, a modified ResNet50 with a lightweight regularized classification head for tomato leaf disease detection under small dataset conditions. Using controlled augmentation, dropout, and batch normalization, the model achieved 98.33% accuracy and 99.88% ROC-AUC, outperforming EfficientNet-B0, MobileNetV2, VGG16, SVM, Random Forest, and XGBoost while maintaining stable training with minimal overfitting.
BanglaBlend Based Evaluation of Bengali Formal and Colloquial Text Classification
Conducted a comparative study on Bengali stylistic text classification using the BanglaBlend dataset with ML, Deep Learning, and Transformer-based NLP models. XLM-RoBERTa achieved the best performance with 94% accuracy and 0.94 F1-score, followed by Bangla-BERT at 93.4%, highlighting the importance of contextual embeddings for handling formal (Sadhu) and colloquial (Cholito) Bangla text.
Developer-Oriented Classification of Mobile App Reviews Using a Hybrid BERT–XGBoost Ensemble
Proposed a hybrid BERT–XGBoost ensemble for developer-oriented classification of mobile app reviews into bug reports, feature requests, performance issues, and praise. By combining contextual embeddings with app metadata through adaptive weighted soft voting, the model achieved 79.17% accuracy and 0.945 macro ROC-AUC, improving automated review triage for software maintenance.
LeViT-Based Ocular Tumor Detection Using Ultra-Wide-Field Fundus Imaging
Proposed a hybrid LeViT (Lightweight Vision Transformer) framework for automated ocular tumor detection from ultra-wide-field fundus images. The model achieved 96% accuracy and 0.93 F1-score, outperforming MobileNetV2 and AlexNet. Grad-CAM visualizations highlighted hyper-reflective tumor regions, supporting interpretable and clinically viable retinal tumor screening.
A Comparative Deep Learning and Transformer-Based Approach to Contextual Emotion Classification in Bangla Lyrical Verse
Conducted a comparative study of classical ML, deep learning, and transformer models for emotion classification in Bangla lyrical verses across four categories: Happiness, Sadness, Anger, and Nostalgia. BanglaBERT achieved the best performance with 76.52% accuracy, establishing the first benchmark for Bangla lyric emotion classification and enabling future music recommendation and cultural analysis applications.
EffiViT-Hybrid: A CNN–Transformer Framework for Pancreatic Cancer Detection from CT Images
Proposed EffiViT-Hybrid, a CNN–Transformer architecture combining EfficientNet and Vision Transformer to detect pancreatic cancer from CT scans. The model achieved 97.74% accuracy and 99.90% ROC-AUC on internal data and maintained 95.54% accuracy on external validation, demonstrating strong generalization for assisting radiologists in early cancer diagnosis.
MangoNet-SA: A Lightweight MobileNetV2-Based Spatial Attention CNN for Robust Mango Leaf Disease Classification
Proposed MangoNet-SA, a lightweight MobileNetV2-based CNN enhanced with a spatial attention mechanism for mango leaf disease detection under real orchard conditions. The model achieved 92.04% accuracy while maintaining balanced precision, recall, and F1-score, outperforming MobileNetV2, MobileNetV3-Large, and EfficientNet variants. Designed for efficient deployment, it supports practical field-level disease monitoring on low-resource hardware.
Explainable Eggplant Leaf Disease Classification Using Vision Transformers and Convolutional Neural Networks
Developed an explainable framework combining Vision Transformers and CNNs for eggplant leaf disease detection across 11 classes. The model achieved 99.87% accuracy and strong class-wise precision, recall, and F1-scores. Grad-CAM visual explanations enable interpretable disease localization, supporting reliable large-scale monitoring in real agricultural conditions.
Robust AIDS Disease Classification Using a Threshold-Optimized Weighted Ensemble and Class-Balanced Machine Learning
Proposed a Threshold-Optimized Weighted Ensemble for AIDS risk prediction using Random Forest, Gradient Boosting, and XGBoost, with class imbalance handled via SMOTE. The ensemble achieved 90.19% accuracy, 79.61% F1-score, and 92.71% ROC-AUC, improving minority class prediction and supporting reliable early AIDS risk assessment.
Machine Learning and Deep Learning Pipelines for Environmental Monitoring and Sustainability
Presented a structured AI pipeline for environmental monitoring that integrates machine learning and deep learning techniques to analyze large-scale environmental data. The chapter highlights spatiotemporal modeling, reproducible workflows, and explainable AI to support reliable sustainability-driven decision making.
Energy Aware Machine Learning Frameworks for Sustainable Environmental Applications
Introduced Energy-Aware Machine Learning (EAML) frameworks that balance model performance with energy efficiency for sustainable AI systems. The chapter discusses Green AI principles, energy-performance metrics, and optimization techniques such as pruning, quantization, and knowledge distillation for resource-constrained environmental monitoring applications.
Green AI for Environmental Monitoring: Energy Efficient Models and Sustainable Computing Strategies
Explored Green AI principles for building environmentally responsible machine learning systems for continuous environmental monitoring. The chapter introduces sustainability metrics such as Energy per Task and Carbon Footprint, and discusses efficiency–performance trade-offs for developing energy-efficient AI models.
Carbon Conscious Artificial Intelligence: Environmental Impact, Ethical Concerns, and Sustainable AI Frameworks
Examined Carbon-Conscious AI frameworks that integrate sustainable model design with carbon-aware computing strategies. The chapter introduces auditing metrics and optimization techniques such as model compression, pruning, and quantization to balance AI performance with environmental responsibility and global climate goals.
Artificial Intelligence in Water Pollution Monitoring and Management: Methods, Challenges, and Future Directions
Investigated the role of artificial intelligence in improving water pollution monitoring and environmental management. The chapter explores machine learning and deep learning techniques for real-time water quality prediction, anomaly detection, and decision support systems for sustainable water resource management.