A Decentralized Federated Learning Framework for Privacy-Preserving Malnutrition Risk Prediction in Children

CSIR 4PI, Bangalore, Karnataka
CSIR 4PI, Bangalore, Karnataka
CSIR 4PI, Bangalore, Karnataka

Abstract

Malnutrition among children under five remains a critical public health challenge in low-income countries, including Ethiopia. The implementation of centralised machine learning methods faces difficulties because they violate patient privacy laws and existing health system limitations in resource-scarce environments. This paper presents a decentralised federated learning (DFL) framework using the ProxyFL gossip architecture, enabling collaborative malnutrition risk prediction across distributed nodes without any raw data sharing. Each node independently trains a Random Forest (RF) or XGBoost (XGB) classifier on a local partition of an Ethiopian pediatric dataset, exchanging only Fernet-encrypted feature importance vectors as proxy representations. The research examines the three-client federated system, which distributes its dataset among three separate health nodes that each contain approximately 534 training samples after removing the outliers. The study evaluates RF and XGBoost performance across all three client partitions by conducting a direct side-by-side assessment. The three clients show RF test AUC-ROC results of 0.721, 0.718 and 0.656, while XGBoost delivers results of 0.689, 0.674 and 0.539. The ProxyFL gossip protocol is validated as stable and algorithm-agnostic across all 20 communication rounds. SHAP analysis confirms BMI as the universally dominant predictor, with clinically coherent secondary features across all client partitions.