Abstract
Diagnosing anemia accurately remains challenging, particularly in differentiating subtypes (e.g., Iron Deficiency vs. Thalassemia) in decentralized settings. Current Deep Learning (DL) models are limited by uni-modal data (only images), poor generalizability across diverse populations, and a lack of transparency regarding inherent biases. This paper proposes a Federated Explainable Multimodal Deep Learning (FEM-DL) framework for anaemia subtype prediction. We fuse non-invasive biometric images (e.g., conjunctiva, retina) with tabular clinical data (demographics, blood history) using a Transformer-based fusion network. Training is conducted via Federated Learning (FL) across multiple centers to ensure data privacy and enhance cross-domain robustness. Finally, we integrate XAI (SHAP and Grad-CAM) to audit model fairness across protected subgroups (e.g., ethnicity, gender) and provide interpretable feature attributions, establishing a new standard for ethical and globally scalable AI diagnostics.
