Localized Climate-Aware Causal AI for Predictive Epidemiology: A Unified Framework of Transferable Graph Neural Networks

Climate change has fundamentally disrupted the stability of infectious disease transmission, especially for vector-borne epidemics such as dengue, malaria, and Zika. Existing AI forecasting models often suffer from poor transferability, limited interpretability, and weak generalization under heterogeneous climatic conditions. We propose a unified Localized Climate-Aware Causal AI (LCCAI) framework that integrates structural causal modeling, hierarchical graph neural networks, and climate-sensitive transfer learning to address these challenges. The framework formalizes a novel Localized Causal Transferability Theorem, enabling robust generalization across distinct geographic regions while preserving mechanistic epidemiological validity. In multi-regional synthetic experiments simulating Vietnam’s diverse climatic zones, LCCAI reduced predictive instability by over 30% compared to non-causal deep learning models and outperformed ARIMA, LSTM, and GNN baselines in both accuracy and cross-regional stability. Beyond theoretical innovation, LCCAI offers immediate translational relevance for climate-adaptive public health policies and precision epidemic forecasting, providing a scalable foundation for real-time global health security in the Anthropocene.