A Secure Blockchain-Enabled Federated Learning Framework for Privacy-Preserving Public Health Surveillance and Pandemic Intelligence

Abstract

Public health monitoring demands the quick enable of multi-institutional intelligence – but not patient-sensitive data. The following paper introduces a federated learning system based on the blockchain for pandemic surveillance in 10 healthcare institutions while preserving user privacy. The framework is comprised of two-layer LSTMs based on 35 demographic, clinical, physiological, exposure, vaccination, and community-level variables with an observation window of 14 days and a prediction horizon of 7 days. Over 100 communication rounds, there are 5 local epochs in which institutions do not transfer patient level records. Model updates are protected by a trust threshold of 0.60 and differential privacy budgets ranging between ε=0.5 and ε=8, secure aggregation, hashing with SHA-256, and ECDSA-256 signatures. ROC-AUC of ≥ 0.85, sensitivity of ≥ 85%, poisoning-detection of ≥ 90%, invalid-update rejection of ≥ 95%, subgroup performance differences of ≤ 10 percentage points, and blockchain latency of ≤ 2 seconds are all among the evaluation targets. The framework generates individual risk probabilities, 7-day forecasts of outbreaks, alerts for outbreaks and clusters, explanations using SHAP and estimates of healthcare resources, setting measurable needs for future validation.