Overcoming Data Scarcity in Pediatric Medicine: Federated Learning for Multi-institutional Diagnostic Models Without Data Migration

Abstract

Pediatric disease diagnosis is often constrained by limited sample sizes, fragmented data distribution, and strict privacy requirements across medical institutions. To address these challenges, this study proposes a dynamic weighted federated learning framework (DW-FL) for multi-institutional pediatric disease diagnosis, enabling collaborative model training without sharing raw patient data.

The proposed framework introduces a contribution-aware aggregation strategy that dynamically adjusts client weights based on model performance and data characteristics, and incorporates a weighted loss function to mitigate class imbalance commonly observed in pediatric datasets. Experiments conducted under both independent and non-independent data distribution settings demonstrate that the proposed approach achieves improved diagnostic performance and communication efficiency compared with conventional federated averaging methods.

These results indicate that dynamic weighting mechanisms can enhance the robustness of federated learning in heterogeneous pediatric scenarios, providing a feasible solution for privacy-preserving multi-center clinical collaboration.