CP-FST: A Cognitive Privacy-Aware Federated Split Transformer for Secure Collaborative Healthcare Intelligence Systems

Secure and privacy-preserving collaboration remains a critical challenge in modern digital healthcare, where sensitive patient data cannot be centrally shared due to regulatory and security constraints. Addressing this need, this paper presents a Cognitive Privacy-Aware Federated Split Transformer (CP-FST) framework for secure and collaborative healthcare intelligence across distributed medical institutions. The approach integrates federated learning, split learning, and transformer-based representation to enable joint model training without sharing raw patient data, thereby addressing privacy, regulatory, and security concerns in digital healthcare environments. A cognitive privacy controller dynamically determines optimal model partitioning to balance privacy risk, communication overhead, and computational efficiency, while differential privacy-aware self-attention and trust-based secure aggregation enhance robustness against inference and adversarial attacks. Comprehensive preprocessing of electronic health records ensures reliable, sequence-aware clinical representations suitable for distributed learning. Experimental evaluation demonstrates superior predictive accuracy, faster convergence, lower training loss, and improved scalability compared with conventional centralized and federated methods, achieving approximately 98.7% accuracy. Overall, CP-FST provides a secure, efficient, and privacy-preserving solution for real-world collaborative healthcare analytics and supports future extensions toward larger multi-institutional deployments and stronger privacy guarantees.