Asynchronous proximal federated aggregation for heterogeneous healthcare networks

Area of Science:

• Medical Informatics
• Machine Learning
• Distributed Systems

Background:

• Federated Learning (FL) deployment in the Internet of Medical Things (IoMT) faces challenges from computational asymmetry and statistical heterogeneity.
• Synchronous aggregation protocols in FL are inefficient due to straggler effects from devices with varying computational capacities.

Purpose of the Study:

• To introduce the Asynchronous Proximal Federated Aggregation (APFA) framework.
• To address computational asymmetry and statistical heterogeneity in FL for IoMT.

Main Methods:

• APFA integrates a local proximal regularizer with a server-side staleness dampening penalty.
• It allows continuous, uncoordinated model updates from edge devices, enabling asynchronous learning.

Main Results:

• APFA achieved an 80% diagnostic viability threshold in 4.1 simulated hours on skewed datasets (CheXpert, MIMIC-IV).
• This represents a 71% reduction in total wait time compared to synchronous methods like FedProx.

Conclusions:

• APFA's mathematical integration mitigates weight divergence, ensuring robustness.
• The framework demonstrates the potential of asynchronous machine learning for scalable, privacy-preserving clinical diagnostics.