A federated attention-based stacked LSTM framework for interpretable malaria diagnosis under simulated non-IID...

Malaria remains a major global health burden, particularly in low-resource regions where microscopic diagnosis relies heavily on expert interpretation and is prone to variability. Although deep learning models have demonstrated strong classification performance for automated malaria detection, most existing approaches operate as black-box systems and assume centralized, independent and identically distributed (IID) data. These limitations restrict clinical trust and motivate the exploration of privacy-preserving learning paradigms for distributed healthcare environments. To address these challenges, this study presents FASL-Net, a Federated Attention-Based Stacked Long Short-Term Memory framework that integrates quantifiable interpretability with evaluation under simulated non-IID federated learning conditions. Microscopic blood smear images are transformed into patch-level sequential representations (100 × 75), enabling spatial dependency modeling through stacked LSTM layers. A temperature-controlled attention mechanism identifies diagnostically relevant regions, and interpretability is quantitatively assessed using attention entropy and causal deletion-insertion analysis. The proposed model achieved 94.39% accuracy in centralized training and 94.95% accuracy in a simulated three-client federated learning setting with heterogeneous data partitions. Attention entropy analysis yielded an average entropy of approximately 2.46, substantially lower than the theoretical maximum (≈ 4.60), indicating focused decision patterns. Deletion experiments resulted in an accuracy reduction of approximately 44% when highly attended patches were removed, whereas insertion experiments preserved approximately 95% performance using only the top 10% salient patches. McNemar's test between centralized and federated models produced a test statistic of 0.03 (p = 0.861), indicating no statistically significant difference under the current simulated setting. These results demonstrate the feasibility of combining federated learning, sequential attention modeling, and quantifiable interpretability for malaria diagnosis under controlled heterogeneous conditions. The study provides a foundational proof-of-concept for trustworthy federated malaria diagnosis and establishes a basis for future validation using real multi-center clinical datasets.