A Privacy‐Preserving and Interpretable Federated Learning Framework With Clinician Feedback for Multi‐Hospital Clinical Decision Support
Gaurav Raj, Gaurav Dubey, Anil Kumar Dubey, Kamal Jit, Akash Punhani, Nripendra Narayan Das · Concurrency and Computation Practice and Experience · 2025
ABSTRACT Federated learning has revolutionized collaborative clinical modeling in multi‐hospital networks by enabling privacy‐preserving analysis of distributed healthcare data. However, challenges such as Data Heterogeneity across Hospitals, Privacy Attacks via Model Updates, Interpretability and Clinician Trust hinder its practical deployment. The Proposed Privacy‐Respectful Interpretable Federated Learning with Intelligent Clinician Feedback Loop Collaboration Network (PRIFLIC‐Net) framework addresses these issues by integrating advanced federated learning with privacy‐preserving and interpretable artificial intelligence (AI) techniques tailored for clinical decision support systems. Each participating hospital processes its local Electronic Health Records (EHRs) using a Knowledge Abstraction and Filtering Module, which extracts structured and unstructured clinical features. These features feed into a Residual Deep Belief Network, which captures complex hierarchical patterns while mitigating vanishing gradient issues through residual connections. To ensure data privacy, a Privacy‐Preserving Transformation Layer applies differential privacy with Gaussian noise and homomorphic encryption, securing local model updates against privacy attacks. The Distributed Federated Aggregation Module employs FedNova and attention‐based weighting to fairly aggregate updates across heterogeneous data silos, enhancing robustness and scalability. A Global Model Update module synchronizes knowledge with adaptive learning rates, ensuring stable convergence. The Explainability and Clinician Feedback Layer integrates SHAP‐based feature importance visualizations and real‐time clinician input, fostering transparency and trust critical for clinical adoption. Evaluated on the MIMIC‐IV dataset, PRIFLIC‐Net achieves superior performance with 88.4% accuracy, 87.9% precision, 86.3% recall, 87.1% F1‐Score, 0.93 AUC‐ROC, 0.82 MCC, and a 0.118 Brier Score, outperforming state‐of‐the‐art models. By mitigating privacy risks and ensuring regulatory compliance, PRIFLIC‐Net delivers scalable, interpretable, and trustworthy clinical decision support across hospital‐wide, ICU, and emergency settings. The integration of advanced federated learning techniques with real‐time clinician feedback ensures not only enhanced diagnostic accuracy but also fosters practitioner trust. PRIFLIC‐Net stands poised to revolutionize privacy‐preserving AI deployment in healthcare, marking a significant step toward secure, transparent, and collaborative medical intelligence, ultimately contributing to safer and more effective patient care.