A Federated Learning Approach to Secure AI-Based Patient Outcome Prediction Across Hospitals

The potential of artificial intelligence to transform healthcare is increasingly realized through patient outcome prediction models. However, traditional centralized training methods for such models pose significant privacy risks, particularly when sensitive patient data must be shared across institutions. This paper proposes a federated learning (FL) framework for developing robust and secure patient outcome prediction models across hospitals while ensuring data privacy and regulatory compliance. Using synthetic and real-world datasets such as MIMICIII and eICU, it creates a multi-hospital-based environment, where models are trained locally and then aggregated in a centralized manner without sharing raw patient data. The LSTMbased and transformer-based architectures are being applied in experiments to time-series health record data, and the accuracy of prediction is statistically significant in the outcomes of ICU mortality and readmission. The FL model achieves competitive performance compared to centralized training, with less than 3% performance degradation and full compliance with privacy-preserving standards. Differential privacy and secure aggregation enhancements was also explored to improve robustness against adversarial participants. Our findings indicate that federated learning presents a scalable, secure, and practical approach to collaborative AI in healthcare, bridging the gap between innovation and privacy protection.