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

Background. Distributed learning lets hospitals train prediction models jointly without exchanging patient records, but how the benefits of a swarm network change as it grows, and which hospitals gain most, have so far been studied with artificial partitions of single-center databases. We examined network scaling on real hospitals that differ in size, volume and sepsis prevalence. Methods. We applied the seven-KPI swarm-learning evaluation framework (Swarm Gain, Local Gain, Relative Local Gain, Model Stability Index, Jain's index, Data Utility Index and label entropy) to progressive network scaling on the eICU Collaborative Research Database. From 25 hospitals (71,887 ICU stays; sepsis prevalence 16.6%), six hospitals stratified by bed size (two small, two medium, two large; 14,776 stays) formed networks of two, four and six hospitals in all six category-joining orders; four further hospitals served as substitutes in robustness checks. Swarm learning was simulated with the client-controlled workflow of NVIDIA FLARE (federated averaging, 30 rounds, five random seeds). The task was whole-stay sepsis classification from discretized vital-sign, laboratory and medication features with a 1,313-parameter feedforward network, with AUROC as the metric. Results. Network AUROC increased with every addition in every ordering, from 0.734 (small-first), 0.755 (medium-first) and 0.781 (large-first) at two hospitals to 0.804 (SD 0.002) at six. Swarm Gain exceeded 1 at all nine unique transitions (mean 1.042 for two-to-four and 1.020 for four-to-six hospitals). Every hospital had Local Gain below 1; relative improvement over local training ranged from 4.1% for the strongest hospital (local AUROC 0.773) to 29.5% for the weakest (0.621), averaging 25.5% for small and 5.9% for large hospitals. Adding large hospitals gave existing members the largest gain (mean Relative Local Gain +6.2%); adding small hospitals never reduced it. Seed-to-seed variability fell from a coefficient of variation of 2.2% at two hospitals to 0.2% at six. Feature distributions were near-identical across hospitals (Data Utility Index 0.005-0.013), whereas the change in label entropy tracked Swarm Gain (r = 0.74). Substituting alternative medium or large hospitals left network AUROC between 0.790 and 0.805. Conclusions. On real multi-hospital data, swarm participation improved every hospital's model, most for hospitals with the weakest local models, with no transition in which an addition harmed the network. Because all orderings end in the same six-hospital network, joining order affected only intermediate performance. The results characterize the KPI framework under real inter-hospital variation in volume and label prevalence and support inclusive recruitment anchored by large hospitals. Per-hospital evaluation and warm-started sequential joining are needed to quantify member-level effects.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-12
DOI
https://doi.org/10.64898/2026.11.11.26367729
Primary Topic
Privacy-Preserving Technologies in Data
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

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

Dawit Gebremariam, Carlo Combi, Matteo Mantovani
medRxiv
Privacy-Preserving Technologies in Data
preprint

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

Dawit Gebremariam, Carlo Combi, Matteo Mantovani
preprint en

Abstract

Background. Distributed learning lets hospitals train prediction models jointly without exchanging patient records, but how the benefits of a swarm network change as it grows, and which hospitals gain most, have so far been studied with artificial partitions of single-center databases. We examined network scaling on real hospitals that differ in size, volume and sepsis prevalence. Methods. We applied the seven-KPI swarm-learning evaluation framework (Swarm Gain, Local Gain, Relative Local Gain, Model Stability Index, Jain's index, Data Utility Index and label entropy) to progressive network scaling on the eICU Collaborative Research Database. From 25 hospitals (71,887 ICU stays; sepsis prevalence 16.6%), six hospitals stratified by bed size (two small, two medium, two large; 14,776 stays) formed networks of two, four and six hospitals in all six category-joining orders; four further hospitals served as substitutes in robustness checks. Swarm learning was simulated with the client-controlled workflow of NVIDIA FLARE (federated averaging, 30 rounds, five random seeds). The task was whole-stay sepsis classification from discretized vital-sign, laboratory and medication features with a 1,313-parameter feedforward network, with AUROC as the metric. Results. Network AUROC increased with every addition in every ordering, from 0.734 (small-first), 0.755 (medium-first) and 0.781 (large-first) at two hospitals to 0.804 (SD 0.002) at six. Swarm Gain exceeded 1 at all nine unique transitions (mean 1.042 for two-to-four and 1.020 for four-to-six hospitals). Every hospital had Local Gain below 1; relative improvement over local training ranged from 4.1% for the strongest hospital (local AUROC 0.773) to 29.5% for the weakest (0.621), averaging 25.5% for small and 5.9% for large hospitals. Adding large hospitals gave existing members the largest gain (mean Relative Local Gain +6.2%); adding small hospitals never reduced it. Seed-to-seed variability fell from a coefficient of variation of 2.2% at two hospitals to 0.2% at six. Feature distributions were near-identical across hospitals (Data Utility Index 0.005-0.013), whereas the change in label entropy tracked Swarm Gain (r = 0.74). Substituting alternative medium or large hospitals left network AUROC between 0.790 and 0.805. Conclusions. On real multi-hospital data, swarm participation improved every hospital's model, most for hospitals with the weakest local models, with no transition in which an addition harmed the network. Because all orderings end in the same six-hospital network, joining order affected only intermediate performance. The results characterize the KPI framework under real inter-hospital variation in volume and label prevalence and support inclusive recruitment anchored by large hospitals. Per-hospital evaluation and warm-started sequential joining are needed to quantify member-level effects.

medRxiv
University of Verona (IT)
Good health and well-being
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon