Privacy-Preserving Federated Learning Framework for Disease Progression Prediction via Temporal-Aware Large Language Modeling

This study presents a novel federated learning approach for disease prediction by integrating the Chronos temporal large language model. The system addresses two critical challenges in public health management: data privacy protection and complex spatiotemporal modeling. Through a hierarchical federated architecture, the model enables collaborative learning across medical institutions without sharing raw data, ensuring compliance with privacy regulations such as GDPR. The Chronos model’s innovative temporal discretization and dual-stream attention mechanism effectively capture disease transmission patterns across different regions and time scales. Experimental results demonstrate that our system achieves a 21.4% improvement in prediction accuracy compared to traditional methods, while maintaining robust privacy protection with a reconstruction accuracy below 12.4% under various attack scenarios. The model’s practical value is validated through successful deployment in vaccine distribution optimization, achieving a 12.6% increase in coverage rate and reducing high-risk community response time by 3.2±0.8 days. This research provides a reliable technical foundation for public health decision-making while ensuring data security and privacy protection.