A Federated Learning and Digital Twin Framework for Real-time Healthcare Analytics: Addressing Privacy, Scalability, and Communication Overhead

The rapid expension of Healthcare Internet of Things (IoT) devices, generates over 2.3TB of data daily, creating communication bottlenecks and hindering timely decision-making in response to rapidly increasing data volumes. This paper presents a scalable architectural framework that synergises Federated Learning (FL) with Digital Twin (DT) technology, leveraging Eclipse Ditto to facilitate real-time data visualisation and management. Our experimental framework utilises multiple simulated clients, each representing a unique DT, to perform local model training on partitioned datasets. These clients periodically transmit model updates to a central aggregator, which consolidates a global model using the FedAvg algorithm. Eclipse Ditto ensures seamless synchronisation between physical devices and their virtual counterparts, enabling real-time monitoring and control. Through comprehensive simulation-based performance evoluations using wearable patient datasets, notably Wearable Stress and Affect Detection (WESAD), we validate the efficacy of the framework in addressing key challenges prevalent in such IoT environments. Our analysis demonstrates efficient asynchronous handling of client updates, resilience to network variability, and inherent scalability in resource-constrained settings. Furthermore, the results confirm that the system can mitigate issues related to non Independent and Identically Distributed (non-IID) data distributions and communication overhead, all whilst maintaining high model accuracy and enhancing data privacy. This work underscores the transformative potential of combining FL with DT platforms to advance real-time analytics in complex, distributed IoT environments.