Scalability and Saturation in Swarm Learning: A KPI-Driven Analysis on Real-World Clinical Data

The adoption of distributed machine learning in healthcare is rapidly expanding to leverage multicentric clinical data while adhering to strict privacy regulations. While Swarm Learning (SL) offers a fully decentralized alternative to a standard centralized Federated Learning setting, the scalability of such peer-to-peer networks remains underexplored. Existing evaluations predominantly rely on synthetic datasets and standard utility metrics, failing to capture the complex distributions of real-world clinical data and struggling to identify saturation regimes where adding participants offers negligible benefits. In this paper, we present a comprehensive scalability analysis of a SL network using real-world intensive care datasets (MIMIC and eICU). To systematically evaluate the network’s behavior, we introduce three novel Key Performance Indicators (KPIs) designed to quantify the learning outcome at the node, patient, and swarm levels. Our findings provide a robust evaluative framework to inform the optimal design, sizing, and deployment of future decentralized clinical collaborations.