Federated modular clinical decision support networks for collaborative learning in resource-limited settings

Abstract Imperfect interoperability (IIO), where health facilities record different, often sparse subsets of clinical variables, remains a major barrier to deploying models trained with Federated Learning (FL) in global health settings. We introduce FedMoDN, a novel federated modular neural network architecture for collaborative learning across all features of an IIO distributed dataset, allowing healthcare facilities to use the full complement of their features without sharing, discarding, or imputing any data. We evaluate FedMoDN on a multi-site pediatric dataset comprising ~130,000 medical visits across 92 healthcare facilities in Tanzania and Rwanda. Across both internal and external validation health facilities, FedMoDN matches or surpasses models trained with centralized data sharing and competitive monolithic FL baselines, achieving a mean AUPRC of 0.80 versus 0.77 for the monolithic FL model on 18 external validation health facilities. Its relative advantage over a monolithic FL model rose from 4% (complete data) to 22% when 70% of test-time features were missing, and, unlike monolithic FL models, performance remained stable when health facilities contributed disjoint feature or label subsets. Furthermore, step-wise predictions provide clinically interpretable feature-attribution scores. By coupling IIO resilience with built-in interpretability, FedMoDN offers a promising decision support tool for resource-limited facilities sidelined by conventional FL.