Interpretable Deep Learning Models for Transparent Health Decision Support Systems

The increasing adoption of deep learning in healthcare has created significant opportunities for improving medical diagnosis, patient risk prediction, and clinical decision support. However, many deep learning models remain difficult to interpret, limiting their acceptance in health systems where transparency, accountability, and patient safety are essential. This study aims to develop and evaluate interpretable deep learning models for transparent health decision support systems. The proposed approach integrates convolutional neural networks, deep neural networks, attention mechanisms, SHAP values, and Grad-CAM to explain model predictions across medical imaging and clinical tabular datasets. Public healthcare datasets, including chest X-ray images, electronic health records, and diabetes patient records, are used to represent real clinical decision-making contexts. Model performance is evaluated using accuracy, precision, recall, F1-score, AUC, sensitivity, and specificity, while interpretability is assessed through explanation clarity, feature relevance, decision traceability, and clinical usefulness. The findings indicate that interpretable deep learning models can provide reliable predictive performance while offering clearer explanations for medical decisions. Visual heatmaps and feature attribution outputs help identify relevant disease indicators, patient risk factors, and clinical variables that influence model predictions. This study contributes to the development of trustworthy AI in healthcare by supporting transparent, ethical, and human-centered clinical decision-making. The research is also aligned with SDG 3 by promoting better health outcomes, patient safety, and responsible digital health innovation.