FedPDM-Net: a federated prototype-guided disentangled deep learning framework for explainable malaria detection

Malaria remains a critical global health concern, particularly in resource-limited settings where accurate and timely diagnosis is essential. While deep learning methods have improved automated detection from blood smear images, most rely on centralized training, raising concerns related to privacy, scalability, and interpretability. To address these limitations, this study proposes FedPDM-Net, a federated prototype-guided disentangled deep learning framework for explainable malaria detection. The model integrates federated learning for privacy-preserving training, a prototype memory module for capturing representative infection patterns, and disentangled feature learning to separate morphological and infection-specific characteristics. An attention-based fusion mechanism enhances feature representation, and Grad-CAM provides visual interpretability. Experimental results show that the model achieves 97.3% accuracy and F1-score in the centralized setting and 96.6% accuracy and F1-score in the federated setting, with only a 0.7% performance drop. A 5-fold cross-validation yields 96.83% ± 0.20 accuracy, confirming stability. The model maintains > 95% performance under perturbations, demonstrating robustness. Statistical analysis shows significant improvement over baselines (p < 0.05), while calibration results (ECE = 0.032) indicate reliable confidence estimation. Additionally, efficient federated communication is achieved with 147 MB overhead over 5 rounds. Overall, FedPDM-Net provides a robust, interpretable, and privacy-preserving solution for automated malaria diagnosis.