Federated Learning for Fall Detection With Multimodal Residual Fusion and Pareto-Optimized Client Selection

With the increasing aging population and the prevalence of chronic diseases, fall detection has become a critical component in elderly healthcare monitoring. However, challenges such as multimodal data integration and joint analysis in Internet of Medical Things (IoMT) environments and data heterogeneity across sources hinder efficient and accurate fall detection. This paper proposes a Federated Learning-based framework with Multimodal Residual Fusion and Pareto-optimized Client Selection (FLPCS-MRF). Firstly, the framework incorporates a multimodal feature fusion network with a residual mechanism, which adaptively learns the optimal fusion scheme through residual connections, dynamically suppressing noise interference from redundant modalities. Secondly, to address variations in data modalities, distributions, and quality across clients, by considering all client factors rather than treating clients as independent, five innovative evaluation metrics are designed to assess the convergence and generalization performance of the local models. Finally, a Pareto-optimized client selection method is introduced to efficiently select reliable clients for global aggregation, ensuring both the stability and robustness of the global model. Extensive experiments on the UP Fall dataset demonstrate the effectiveness of the proposed approach, achieving 95.27% accuracy and 95.42% F1-score, outperforming existing methods. Additionally, it demonstrates strong robustness in complex scenarios involving imbalanced data distributions and missing modalities.