Reconfigurable trustless digital twins for continuous patient monitoring: A federated learning-blockchain architecture in healthcare 5.0

Background: Synchronizing real-time Digital Twins (DT) for chronic disease management in Healthcare 5.0 introduces severe data heterogeneity and single-point-of-failure vulnerabilities. Furthermore, existing distributed machine learning frameworks lack the hot-swappable reconfigurability required for evolving, resource-constrained smart ward environments. Methods: We propose a Trustless Digital Twin Synchronization framework that extends the RECONFIGMED-CHAIN architecture. The system employs a Depthwise Separable Lightweight Convolutional Neural Network (L-CNN) for multi-modal sensor fusion at the bedside. Model updates are managed via an attention-gated federated aggregation layer under strict ( ϵ , δ ) -differential privacy. To prevent centralized data pooling, spatial gradients are anchored to an immutable global DT state via a PBFT-Lite consensus mechanism on a permissioned ledger. Results: Validated across 40 physical IoMT edge nodes mapped to 10 federated ICU clients, the proposed architecture achieved an AUROC of 0.913 ± 0.006 (AUPRC: 0.704 ± 0.011) for 24-hour Sepsis-3 prediction on MIMIC-IV, and an anomaly detection AUROC of 0.982 ± 0.003 on the WUSTL-EHMS wearable network dataset. The PBFT-Lite consensus protocol reduced commit latency to 384 ms ( p 50 at n = 16 validators), representing a 63.2% latency reduction over standard PBFT. Wearable node energy consumption was reduced by 38.5% per round compared to standard PBFT-FL baselines. Conclusion: By treating algorithmic reconfigurability as a first-class design dimension and anchoring updates to an immutable ledger, this framework guarantees verifiable, sub-second synchronization. It bridges the critical gap between secure federated AI and its operational, energy-efficient deployment in real clinical wards.