Graph-Based Prediction of Spatio-Temporal Vaccine Hesitancy From Insurance Claims Data.

Area of Science:

• Computational epidemiology
• Public health informatics
• Machine learning for healthcare

Background:

• Declining immunization rates due to vaccine hesitancy pose a significant public health threat.
• Understanding the spatio-temporal spread of vaccine hesitancy is critical for targeted interventions.
• Predicting vaccine hesitancy at high spatial resolutions, like ZIP Codes, is challenging due to data limitations.

Purpose of the Study:

• To develop and evaluate a novel framework, VaxHesSTL, for predicting vaccine hesitancy at the ZIP Code level.
• To assess the impact of different network structures (contact-based vs. spatial proximity) on prediction accuracy.
• To integrate active learning to optimize data requirements for high-resolution hesitancy prediction.

Main Methods:

• Developed a hybrid Graph Neural Network (GNN) and Recurrent Neural Network (RNN) framework (VaxHesSTL).
• Utilized a large All-Payer Claims Databases (APCD) dataset from Virginia (over 5 million individuals, 6 years).
• Incorporated an active learning strategy to identify optimal training data subsets.

Main Results:

• VaxHesSTL significantly outperforms baseline models that do not account for spatial relationships.
• Contact network-based graphs improve prediction performance compared to solely spatial proximity graphs.
• Active learning enabled accurate hesitancy forecasting for all ZIP Codes using data from only 18% of them.

Conclusions:

• The VaxHesSTL framework effectively predicts vaccine hesitancy at a granular ZIP Code level.
• Contact network structure is a vital component for accurate spatio-temporal hesitancy modeling.
• Active learning offers an efficient approach to overcome data scarcity in high-resolution public health prediction.