Personalized Meta-Federated Learning for IoT-Enabled Health Monitoring.

Area of Science:

• * Computational health informatics and machine learning.
• * Internet of Things (IoT) for remote health monitoring.
• * Signal processing and pattern recognition for biosignals.

Background:

• * Federated learning (FL) is utilized for privacy-preserving health monitoring using biosignals.
• * Standard FL models exhibit uneven performance across subjects due to complex temporal dynamics and inter/intra-subject variability in biosignals.
• * Existing methods struggle to balance global model generalization with individual subject needs.

Purpose of the Study:

• * To introduce the Personalized Meta-Federated learning (PMFed) framework for personalized IoT-enabled health monitoring.
• * To enhance the performance and efficiency of federated learning in analyzing subject-specific biosignals.
• * To address the challenges of inter-subject and intra-subject variability in biosignal data.

Main Methods:

• * Implemented a meta-federated learning paradigm with a novel momentum-based model aggregation strategy based on domain similarity.
• * Developed an adaptive model personalization mechanism to tailor global models to subject-specific biosignal features.
• * Evaluated the framework using an IoT-enabled computing system on three real-world health monitoring tasks.

Main Results:

• * PMFed demonstrated superior detection performance, improving F1 and accuracy by up to 9.4% and 8.7% respectively.
• * Achieved significant reductions in training overhead (up to 56.3%) and throughput (up to 63.4%) compared to state-of-the-art federated learning algorithms.
• * Showcased effective personalization by adapting the global model to individual subject biosignal characteristics.

Conclusions:

• * The PMFed framework effectively addresses biosignal variability in IoT health monitoring, outperforming existing federated learning approaches.
• * PMFed offers a robust solution for personalized health monitoring, balancing privacy preservation with individual-specific model adaptation.
• * The proposed methods significantly improve detection accuracy and reduce computational costs, paving the way for efficient real-world applications.