An equity-aware generative AI copilot for digital public health surveillance

Modern public health surveillance depends on multiple data streams, including routine case reporting, contextual regional indicators, environmental measurements, and digitally generated signals. In many operational settings, however, these inputs are analyzed through disconnected tools, leaving forecasting, outbreak flagging, fairness auditing, and interpretation weakly coordinated. To address this gap, this study develops an equity-aware multimodal copilot for digital public health surveillance that unifies a graph-augmented Temporal Fusion Transformer, anomaly detection, subgroup fairness regularization, and retrieval-augmented large language model support within one analystfacing framework. The empirical evaluation uses 260 weeks of surveillance data covering 9 administrative regions in Saudi Arabia. The data include weekly syndrome counts together with demographic context, environmental variables, and selected digital signals. Following preprocessing and multimodal feature construction, the predictive component learns temporal patterns and regional interaction, the anomaly module detects elevated-risk periods, the fairness term reduces disparity in true positive rates across predefined groups, and the copilot generates evidence-grounded narrative explanations for human review. On the held-out test set, the framework achieved an RMSE of 0.178 and a MAPE of 10.6% for four-week-ahead forecasting. For outbreak detection, it obtained an AUROC of 0.936 and an F1 score of 0.832. The fairness-aware configuration also narrowed subgroup recall gaps, and the retrieval-augmented copilot achieved an entity-level F1 of 0.89 with strong citation coverage. Overall, the results indicate that integrating spa-tio-temporal modelling, fairness monitoring, and grounded language assistance can strengthen public health decision support while preserving human oversight.