Explainable AI in Healthcare- Integrating Grad-CAM and SHAP for Multimodal Diagnostic Systems

The adoption of deep learning in healthcare diagnostics has improved prediction accuracy across medical imaging, clinical text, and electronic health records. However, the "black-box" nature of these models poses a barrier to clinical trust, regulatory approval, and real-world deployment. This paper presents a systematic framework for explainable artificial intelligence (XAI) in healthcare, focusing on the integration of Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) within multimodal diagnostic pipelines. The approach is demonstrated using a multimodal architecture that fuses radiology images, structured tabular data (e.g., lab tests, vitals), and unstructured clinical notes. Grad-CAM provides localized heatmaps on medical images (e.g., CT, MRI, X-ray) to highlight salient regions driving predictions, while SHAP quantifies the contribution of tabular features (e.g., age, biomarkers) and textual tokens to model output. By combining these interpretability layers, clinicians can obtain modality-specific explanations alongside a holistic understanding of the decision-making process. Preliminary experiments on publicly available datasets (MIMIC-CXR, LIDC-IDRI, and MIMIC-IV clinical data) demonstrate that multimodal models with integrated Grad-CAM + SHAP achieve AUC improvements of up to 4% compared to unimodal models, while also providing actionable explanations that radiologists and clinicians rated as significantly more trustworthy in a user study. This work contributes by: (1) designing a unified framework for multimodal interpretability; (2) evaluating its effectiveness across imaging, tabular, and textual modalities; and (3) addressing the gap between high-performing models and clinically interpretable systems. Future work includes validating the approach in real-world