| Continuous physiological monitoring in outdoor environments is challenging due to unstable connectivity, latency variation and the need for timely clinical response. Cloud-centric architectures may experience delayed transmission and reduced reliability when communication conditions degrade. This work extends a previously evaluated fog-cloud health monitoring architecture by incorporating embedded machine learning, physician-supervised feedback and federated learning with model poisoning mitigation. Physiological signals collected from wearable devices are transmitted to smartphones, which act as fog nodes for preprocessing, feature extraction and alert screening using a lightweight logistic regression model. When an alert is detected, the smartphone sends it to the cloud through MQTT, where it is forwarded to a physician for validation. The physician's decision is later incorporated into the local dataset, establishing a human-in-the-loop learning cycle. Outdoor experiments provide initial evidence that smartphone-based fog inference can support alert screening under variable connectivity. Controlled federated learning experiments with 10%, 25% and 50% malicious-client proportions show that poisoned updates degrade FedAvg. In the 25% setting, the F1-score decreases from 0.960 to 0.214, while the coefficient-distance defense restores it to 0.960. The 50% setting exposes the limitation of the heuristic, indicating anomalous-update screening rather than complete Byzantine resilience. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.