Federated Learning Meets Digital Twins: Adaptive Multimodal Health Monitoring with Entropy-Based Client Selection

Today, digital twin (DT) is a hot topic in both academia and industry. It is considered a valuable solution for different sectors and applications, and its full deployment is expected with the release of sixth-generation (6G) networks, which will offer useful cutting-edge technology tools that can meet the very stringent requirements of the DT paradigm. Digital health (eHealth) is an outstanding candidate for the fruitful use of DT, which could effectively enable innovative services such as real-time health monitoring and predictive diagnostics. In this study, we introduce a novel multimodal DT framework that uses clinical data on electrodermal activity (EDA), temperature, photoplethysmography (PPG), and electrocardiogram (ECG). The proposed framework exploits federated learning (FL) to facilitate collaborative training among distributed clients (which consist of monitoring devices) and protect data privacy. Specifically, collaborative clients are selected through an entropy-based strategy, which dynamically ranks participants based on the uncertainty of their local predictions. Other parameters collected in real time and considered to select the most effective clients relate to device capabilities, such as battery level, processing power, and network stability. The experimental results presented in this paper demonstrate that the proposed approach improves generalization and convergence speed compared to standard FL methods without client selection. These results suggest a promising path for the implementation of intelligent and resource-aware DTs in large-scale, privacy-sensitive eHealth environments.