TRUSTWORTHY AND EXPLAINABLE DEEP LEARNING MODELS FOR MEDICAL DIAGNOSIS USING PAKISTANI CLINICAL AND IMAGING DATASETS

Abstract

The rapid adoption of deep learning (DL) in medical diagnosis has raised significant concerns regarding transparency, reliability, and clinical trust—particularly in developing healthcare ecosystems such as Pakistan. This study proposed a trustworthy and explainable deep learning framework for medical diagnosis using multi-institutional Pakistani clinical and imaging datasets. The dataset comprised 18,742 anonymized patient records collected from three tertiary-care hospitals in Sindh, Punjab, and Balochistan, including 12,560 radiological images (chest X-rays and CT scans) and 6,182 structured clinical profiles related to cardiopulmonary and metabolic disorders. A hybrid convolutional neural network–transformer (CNN-Transformer) architecture was developed for imaging analysis, while a gradient-boosted deep neural network was employed for structured clinical prediction. To enhance trustworthiness, the framework integrated uncertainty quantification using Monte Carlo dropout and Bayesian approximation, alongside post hoc explainability techniques including SHAP (Shapley Additive Explanations) and Grad-CAM visual saliency mapping. Experimental evaluation using five-fold cross-validation demonstrated superior diagnostic performance compared to baseline CNN and ResNet-50 models. The proposed model achieved an overall accuracy of 94.3%, sensitivity of 92.8%, specificity of 95.6%, and an AUC of 0.968 for disease classification. Calibration error was reduced to 2.7%, indicating improved prediction reliability. Explainability analysis showed that 87% of highlighted regions in Grad-CAM visualizations aligned with radiologist-annotated pathological areas, supporting clinical interpretability. Furthermore, uncertainty-aware predictions reduced false-positive alerts by 11.4% in borderline cases. The findings demonstrate that integrating explainability and uncertainty modeling significantly improves both diagnostic accuracy and clinician trust. This framework provides a scalable and ethically aligned solution for AI-driven medical decision support systems in Pakistan’s healthcare context, contributing toward safe, transparent, and accountable artificial intelligence deployment in resource-constrained clinical environments