FedPACT: A Policy Calibration Framework for Federated Digital Twins in Healthcare Cyber-Physical Systems

Abstract Healthcare institutions increasingly use locally deployed predictive models, yet collaboration across institutions typically focuses on model training rather than the operational policies governing decisions based on model outputs. This paper introduces FedPACT, a federated policy-calibration framework that coordinates local decision thresholds across distributed healthcare systems without exchanging patient-level data or predictive-model parameters. Each hospital derives an initial threshold from its local predicted-risk distribution and simulated escalation capacity. FedPACT iteratively updates these thresholds using a weighted reference and a local-retention parameter, while preserving local data, models, and policy execution. The mechanism is analysed in terms of convergence, stability, threshold bounds, and computational and communication complexity. The framework is evaluated using precomputed 15- and 30-day readmission-risk estimates for Trauma and General Surgery, partitioned across five simulated hospital nodes with heterogeneous patient volumes, risk distributions, and escalation capacities. Experiments examine static calibration, sensitivity to the local-retention parameter, baseline comparisons, repeated dynamic-capacity simulations, and a virtualised communication microbenchmark. Across the evaluated configurations, local thresholds converged geometrically towards an invariant weighted reference, reducing inter-hospital threshold dispersion. This behaviour persisted under the tested dynamic-capacity conditions. However, threshold alignment did not equalise the number of patients eligible for escalation, which remained dependent on local patient volumes, risk distributions, and simulated capacities. The communication microbenchmark indicated a small application payload and sub-millisecond local processing overhead in the tested environment. These results demonstrate the feasibility of federated calibration of capacity-derived decision thresholds without exchanging patient-level records or predictive-model parameters. They do not establish clinical effectiveness, formal privacy guarantees, real-world network performance, or readiness for autonomous clinical decision-making.