Federated Learning and Explainable AI for Analyzing Heterogeneous Healthcare Data
Keabineh Kaleb Keba, Nagender Kumar Suryadevara, Atul Negi · 2025
This paper presents an interpretable Federated Learning (FL) framework for predicting medication order cancellations in Intensive Care Unit (ICU) settings, balancing predictive performance with stringent patient privacy requirements. Our approach addresses the urgent demand for explainable AI in healthcare by integrating privacy-preserving distributed training with robust model interpretability. Leveraging heterogeneous data sources including MIMIC-III tabular records, medical imaging, and PDF-based lab reports we developed and evaluated two machine learning algorithms. XGBoost achieved the highest performance, with 92.24% accuracy, 92.23% precision, and an F1-score of 90.92%. To enhance trust and clinical usability, we applied SHAP and LIME for both global and local interpretability, uncovering critical features such as medication amount, infusion rate, and temporal factors influencing cancellation decisions. Our federated architecture enables secure, collaborative training across institutions, crucial for sensitive healthcare domains. This work lays a foundation for scalable, privacy-conscious, and interpretable AI systems in critical care, and sets the stage for future research involving multimodal data, including audio-visual inputs and real-time analytics.