DECENTRALIZED DEEP LEARNING IN HEALTHCARE: ADDRESSING DATA PRIVACY WITH FEDERATED LEARNING
Godfrey Perfectson Oise, Chioma Julia Onwuzo, Mary Fole, Samuel Abiodun Oyedotun, Joy Akpowehbve Odimayomi, Nkem Belinda Unuigbokhai, Prosper Otega Ejenarhome, Babalola Eyitemi Akilo · FUDMA Journal of Sciences · 2025
This study presents a privacy-preserving federated learning framework combining recurrent neural networks for healthcare applications, balancing data privacy with clinical utility. The decentralized system enables multi-institutional collaboration without centralized data collection, complying with HIPAA/GDPR through two technical safeguards: differential privacy via DP-SGD during local training and secure aggregation of model updates. Using LSTM/GRU architectures optimized for sequential medical data, the framework achieves an F1 Score of 67% with precision (60%) and recall (75%) suitable for clinical deployment, validated by Cohen's Kappa (40%) and Matthews Correlation Coefficient (40%). Experimental results using real-world datasets demonstrate the system's effectiveness in processing temporal patient records while maintaining data locality. The model reaches 07% of centralized accuracy despite privacy constraints, proving federated learning can deliver medically relevant performance without raw data sharing. The F1 Score above 0.75 with differential privacy confirms that rigorous privacy protections need not compromise predictive utility, while MCC values exceeding 0.4 indicate clinically meaningful performance for applications like readmission risk stratification. The work makes three primary contributions to medical AI: a functional FL-RNN implementation for sensitive health data, quantitative evidence of the privacy-utility tradeoff in clinical settings, and benchmarks for communication-efficient training across non-identical hospital datasets. These outcomes provide healthcare organizations with a practical template for developing collaborative AI that meets both clinical requirements and regulatory standards, particularly for time-sensitive applications involving electronic health records and vital sign monitoring. The framework's balanced performance across all evaluated metrics positions federated learning as a viable alternative to centralized approaches in privacy-sensitive healthcare environments.