Network scaling in swarm learning across real hospitals of different sizes: a KPI-driven analysis on the eICU Collaborative Research Database

Swarm learning on real eICU hospitals shows AUROC rises monotonically with network size, reaching 0.804 (SD 0.002) at six hospitals across all joining orders.

• Why it matters: Prior scaling studies used artificial partitions, leaving uncertainty about how swarm benefits change with real inter-hospital differences in size, volume, and sepsis prevalence.
• What they did: Using a seven-KPI framework (including Swarm Gain and Relative Local Gain), the authors simulated progressive hospital joining on 25 eICU hospitals with FLARE federated averaging (30 rounds, five seeds) for whole-stay sepsis classification (AUROC).
• The result: Swarm Gain exceeded 1 at all nine transitions (mean 1.042 for two-to-four; 1.020 for four-to-six), while relative improvement over local training ranged from 4.1% to 29.5% and seed variability dropped from 2.2% to 0.2%.

AI summary of the abstract · may contain errors · not medical advice · read the paper.