Privacy-Preserving Federated Learning for Stunting Risk Prediction across Indonesian Community Health Centers

Abstract

Stunting remains a serious public-health challenge in Indonesia, and data-driven prediction could help health workers identify at-risk children earlier and target limited resources more effectively. However, maternal and child health data are highly sensitive and are fragmented across many community health centers, and centralizing such records for model training raises privacy, legal, and infrastructural obstacles. This study designs and evaluates FedStunt, a privacy-preserving federated learning framework that trains a shared stunting-risk prediction model across distributed community health centers without moving raw records off site. The framework combines federated averaging with a differential-privacy mechanism based on gradient clipping and calibrated noise, secure aggregation so that the server observes only aggregated updates, and a proximal term to cope with statistically heterogeneous data across sites. Following a design science methodology, the artifact was compared against a centralized model that serves as a privacy-agnostic upper bound and against local-only models trained separately at each site. Evaluation spans discrimination and sensitivity, calibration, fairness across sites, the differential-privacy budget, and communication cost. The results are intended to show whether privacy-preserving federation can approach centralized predictive quality while keeping data local, and to provide a governance-friendly and reusable blueprint for collaborative health analytics in low-resource settings where data sharing is constrained.