Selected work

Representative Papers

StackLiverNet: A Novel Stacked Ensemble Model for Accurate and Interpretable Liver Disease Detection

Jul 31, 2025

Existing liver disease classification models suffer from high misclassification rates, poor interpretability, substantial computational overhead, and inadequate preprocessing. To address these challenges, this paper proposes StackLiverNet—a clinically oriented, interpretable stacked ensemble framework. It integrates random undersampling to mitigate class imbalance, recursive feature elimination for optimal feature subset selection, and a hyperparameter-optimized LightGBM meta-learner to aggregate multiple base classifiers. Innovatively, it incorporates LIME and SHAP for both local and global interpretability, and employs Morris sensitivity analysis to validate the clinical credibility of key biomarkers. Evaluated on a public dataset, StackLiverNet achieves 99.89% accuracy, a Cohen’s Kappa of 0.9974, and an AUC of 0.9993, with only five misclassifications. Training and inference times are merely 4.28 seconds and 0.11 seconds, respectively—demonstrating exceptional precision, robustness, and real-time deployability.

0 citationsRead paper

A Modified VGG19-Based Framework for Accurate and Interpretable Real-Time Bone Fracture Detection

Jul 31, 2025

To address the scarcity of radiological expertise, time-consuming and error-prone manual interpretation, and poor interpretability of existing deep learning models in X-ray fracture diagnosis, this study proposes an end-to-end interpretable fracture detection framework. Methodologically, we enhance the VGG19 architecture by integrating CLAHE-based contrast enhancement, Otsu thresholding, and Canny edge detection to improve feature robustness; additionally, we incorporate Grad-CAM to generate clinically comprehensible decision heatmaps. The system is deployed as a real-time web application with inference latency under 0.5 seconds. Evaluated on a public dataset, our model achieves 99.78% classification accuracy and an AUC of 1.00—significantly outperforming baseline methods. The core contribution lies in a practical, deployable interpretable AI solution that simultaneously ensures high diagnostic accuracy, low latency, and clinical trustworthiness, thereby offering a reliable decision-support tool for primary healthcare settings.

0 citationsRead paper

Improving Chronic Kidney Disease Detection Efficiency: Fine Tuned CatBoost and Nature-Inspired Algorithms with Explainable AI

Apr 05, 2025

To address the low early detection efficiency and poor model interpretability of chronic kidney disease (CKD) in resource-constrained settings, this study proposes an optimized CatBoost framework integrating simulated annealing (for feature selection), cuckoo search (for outlier correction), and SHAP-based interpretability analysis. The framework simultaneously enhances model robustness and clinical credibility without compromising predictive accuracy. Evaluated on a public CKD dataset, it achieves 98.75% accuracy, an AUC of 0.9993, and a Cohen’s kappa of 0.9735. SHAP analysis identifies urine specific gravity, serum creatinine, and albumin as the top three predictive biomarkers. To the best of our knowledge, this is the first work to synergistically embed three distinct metaheuristic optimization algorithms with explainable AI within a gradient-boosting framework. The resulting solution offers high performance, clinical verifiability, and deployment feasibility for primary-care CKD screening.

0 citationsRead paper

Recent publications

Latest Papers

StackLiverNet: A Novel Stacked Ensemble Model for Accurate and Interpretable Liver Disease Detection

Jul 31, 2025

Existing liver disease classification models suffer from high misclassification rates, poor interpretability, substantial computational overhead, and inadequate preprocessing. To address these challenges, this paper proposes StackLiverNet—a clinically oriented, interpretable stacked ensemble framework. It integrates random undersampling to mitigate class imbalance, recursive feature elimination for optimal feature subset selection, and a hyperparameter-optimized LightGBM meta-learner to aggregate multiple base classifiers. Innovatively, it incorporates LIME and SHAP for both local and global interpretability, and employs Morris sensitivity analysis to validate the clinical credibility of key biomarkers. Evaluated on a public dataset, StackLiverNet achieves 99.89% accuracy, a Cohen’s Kappa of 0.9974, and an AUC of 0.9993, with only five misclassifications. Training and inference times are merely 4.28 seconds and 0.11 seconds, respectively—demonstrating exceptional precision, robustness, and real-time deployability.

0 citationsRead paper

A Modified VGG19-Based Framework for Accurate and Interpretable Real-Time Bone Fracture Detection

Jul 31, 2025

To address the scarcity of radiological expertise, time-consuming and error-prone manual interpretation, and poor interpretability of existing deep learning models in X-ray fracture diagnosis, this study proposes an end-to-end interpretable fracture detection framework. Methodologically, we enhance the VGG19 architecture by integrating CLAHE-based contrast enhancement, Otsu thresholding, and Canny edge detection to improve feature robustness; additionally, we incorporate Grad-CAM to generate clinically comprehensible decision heatmaps. The system is deployed as a real-time web application with inference latency under 0.5 seconds. Evaluated on a public dataset, our model achieves 99.78% classification accuracy and an AUC of 1.00—significantly outperforming baseline methods. The core contribution lies in a practical, deployable interpretable AI solution that simultaneously ensures high diagnostic accuracy, low latency, and clinical trustworthiness, thereby offering a reliable decision-support tool for primary healthcare settings.

0 citationsRead paper

Improving Chronic Kidney Disease Detection Efficiency: Fine Tuned CatBoost and Nature-Inspired Algorithms with Explainable AI

Apr 05, 2025

To address the low early detection efficiency and poor model interpretability of chronic kidney disease (CKD) in resource-constrained settings, this study proposes an optimized CatBoost framework integrating simulated annealing (for feature selection), cuckoo search (for outlier correction), and SHAP-based interpretability analysis. The framework simultaneously enhances model robustness and clinical credibility without compromising predictive accuracy. Evaluated on a public CKD dataset, it achieves 98.75% accuracy, an AUC of 0.9993, and a Cohen’s kappa of 0.9735. SHAP analysis identifies urine specific gravity, serum creatinine, and albumin as the top three predictive biomarkers. To the best of our knowledge, this is the first work to synergistically embed three distinct metaheuristic optimization algorithms with explainable AI within a gradient-boosting framework. The resulting solution offers high performance, clinical verifiability, and deployment feasibility for primary-care CKD screening.

0 citationsRead paper