HantaWatch Uses Federated Learning for Hantavirus Surveillance.

Genomic surveillance of Hantavirus faces significant challenges due to the distributed nature of sequence data, the heterogeneity of data sources, and limited expert review capacity. To overcome these hurdles, researchers have developed HantaWatch, a federated learning framework that allows multiple laboratories and surveillance sites to collaboratively train sequence-based models without needing to centralize or share sensitive raw data. HantaWatch is a comprehensive system that incorporates several key components. It begins with k-mer feature extraction to process genomic sequences, followed by the construction of source-aware federated clients. The framework employs adaptive DU-FedProx optimization for efficient and stable model training, alongside surveillance-specific model selection and prediction-only triage. Through extensive experiments on both binary and multi-class tasks, HantaWatch has demonstrated its capability to support high-risk screening, predict outbreak associations, classify clades, and categorize clinical syndromes. It achieves this while maintaining a balance between predictive performance, managing false-negative risks, and ensuring update stability. The framework translates model outputs into actionable risk scores, confidence estimates, uncertainty flags, and prioritized lists for expert review, offering a practical, privacy-preserving decision-support layer for decentralized Hantavirus surveillance.