Comparative Federated Analytics of Blood Transfused Patients in Five ICU Databases: Using Kullback-Leibler Divergence.

Abstract

This study assesses feature distribution differences across five intensive care databases using Kullback-Leibler Divergence (KLD). Analyzing bidirectional KLD patterns between individual databases and the composite of others, stratifying by transfusion status. Results reveal heterogeneity: HiRID shows highest divergence in both directions, particularly among transfusion cases; UKA exhibits moderate overall divergence but pronounced differences in transfusion scenarios; MIMIC-IV shows minimal divergence, indicating closest alignment with group distributions. Notably, transfusion cases consistently display higher divergence than non-transfusion cases across all databases, highlighting institution-specific practices. These findings stress the importance of assessing data heterogeneity before implementing federated learning models to understand generalization capabilities.