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

Area of Science:

• Public Health Surveillance
• Artificial Intelligence
• Data Science

Background:

• Modern public health surveillance relies on diverse data streams but often uses disconnected tools.
• This fragmentation hinders coordinated efforts in forecasting, outbreak detection, and fairness auditing.

Purpose of the Study:

• To develop an equity-aware multimodal copilot for digital public health surveillance.
• To unify advanced analytical tools within a single, analyst-facing framework.

Main Methods:

• Developed a framework integrating a graph-augmented Temporal Fusion Transformer, anomaly detection, subgroup fairness regularization, and retrieval-augmented large language model support.
• Evaluated the system using 260 weeks of surveillance data from Saudi Arabia, including syndrome counts, demographic, environmental, and digital signals.
• Employed multimodal feature construction for spatio-temporal modeling and fairness monitoring.

Main Results:

• Achieved strong forecasting performance with RMSE of 0.178 and MAPE of 10.6% for four-week-ahead predictions.
• Demonstrated high accuracy in outbreak detection with AUROC of 0.936 and F1 score of 0.832.
• Successfully reduced subgroup recall disparities and provided evidence-grounded explanations with an entity-level F1 of 0.89.

Conclusions:

• Integrating spatio-temporal modeling, fairness monitoring, and language assistance significantly strengthens public health decision support.
• The developed copilot enhances surveillance capabilities while maintaining crucial human oversight.
• The framework offers a unified approach to analyzing complex public health data for improved outcomes.