Severity-aware hierarchical federated learning for privacy-preserving indoor emergency recognition in BLE-based medical IoT environments.

Indoor safety in GPS-denied environments such as hospitals remains critically hindered by the lack of data-local, context-aware emergency localization systems. This paper presents a complete three-layer IoT framework, namely Medical Indoor BLE Emergency System (MIBES), integrating BLE 4.2 and Federated Learning (FL) for real-time emergency detection, localization and severity weighted aggregation of manually triggered emergency alerts. Built upon an ESP32-based hardware architecture, the system employs BLE Safety Tags with long-press emergency triggers and Anchor nodes enhanced with Exponential Moving Average (EMA) smoothing and utilizing the User Datagram Protocol (UDP) for data encapsulation and transport. The key novelty of this work lies in the proposed Transformer-based Severity-Aware Hierarchical Federated Learning (SA-HFL) framework. This system jointly addresses class imbalance and statistical data heterogeneity in emergency localization through the integration of Focal Loss. This design lets important emergency classes get priority without ruining the overall model generalisation. The Transformer model achieves an accuracy of 99.52%, recall of 99.08% and a MCC of 0.9904 when evaluated on a real-world BLE dataset via 5-fold cross-validation and numerically outperforms the standard FedAvg and FedProx model baselines with performance gains that is supported by moderate to large effect sizes though not by statistically significant paired tests. Communication efficiency remains high with only 150,849 parameters exchanged in each aggregation round. Consequently, the results validate the proposed architecture as a scalable, privacy-preserving solution for reliable emergency alert localization and prioritization across healthcare critical infrastructure.