Knowledge-Infused Temporal GNNs for Proactive Rabies Hotspot Detection in Data-Sparse Regions

Dog rabies remains a major public health threat in low- and middle-income countries, where underreporting, limited surveillance, and socioeconomic disparities hinder early detection and control. Traditional models struggle to capture the spatial and relational dynamics of transmission, particularly in data-sparse settings. We propose a heterogeneous knowledge graph framework combined with a Graph Neural Network (GNN) to predict canine rabies outbreaks at the urban areas level in Arequipa, Peru. The graph integrates data on rabies cases, socioeconomic information, locations of health centers, and water channels ("torrenteras"). Outbreak prediction is formulated as a binary node classification task, with labels defined by future rabies incidence. Our custom model, KG-GNN, combines SAGE-based aggregation and multi-head graph attention, optimized with weighted loss functions and hyperparameter search to address class imbalance. Experiments on temporally disjoint datasets (2016-2021) show that our KG-GNN consistently out-performs baseline models (LSTM, Random Forest, XGBoost, and MLP). On the 2021 test set, it achieved the highest F1 score and recall, ensuring sensitivity to high-risk areas while maintaining reasonable precision. Moreover, node-level explainability (both local and global) identifies the main drivers of rabies risk, enabling public health authorities to interpret predictions and prioritize interventions. The results highlight the relevance of knowledge-infused graph-based modeling for outbreak prediction under limited supervision. The proposed framework is modular, scalable, and adaptable to other infectious diseases where spatial and relational structures shape transmission dynamics.