Leveraging COVID-19 Insights to Develop Novel Behavior-Informed Multi-Task Machine Learning Frameworks for Targeted Public Health Interventions in Rural America

The Sars-CoV-2 pandemic presented unprecedented challenges in many areas of healthcare. One important area was accurately forecasting the need for resources such as testing, vaccinations, and other critical resources that are linked to hospital care. In rural areas, these issues become even more complex due to increased isolation, health disparities, and social

determinants of health. Rural residents have different social and economic profiles and face different environmental factors than their urban counterparts. The data collected during the RADx-Up project provides a unique opportunity to leverage the Common Data Elements (CDE) to inform and develop the next generation of predictive modeling techniques. West Virginia

(WV), which hosted two unique and highly successful community-based testing, RADx-Up testing projects, provides a novel testing ground for this development. Furthermore, our RadX study has unique access to secondary data sources that supplement the two West Virginia RadX studies, providing the ability to leverage further the RadX Common Data Elements to

develop the next generation of forecasting tools that will be necessary during events like the Sars-CoV-2 pandemic. This study utilizes statewide Sars-CoV-2 testing, vaccination, hospitalization data, and RadX data from two state-wide West Virginia projects to develop novel behavior-informed multi- task machine learning frameworks to predict localities for targeted public health interventions.

In the first aim, we utilize the results of the RadX survey data in West Virginia in the context of the various waves of the pandemic. This will provide insights into behavior changes in how individuals in various communities sought testing and how it impacted other resources such as vaccinations. We also view this in the context of the disease severity with access to full-state

testing data and hospitalizations associated with those communities. This will create a behavioral profile to understand how these behaviors changed over various waves of the pandemic. In the second aim, we develop novel machine learning frameworks that allow public health officials and practitioners to include and exploit these insights to create better-targeted

public health interventions. This requires developing new computational and graph-constrained multi-task machine-learning paradigms that leverage the relationship between outcomes such as incidence, testing utilization, vaccine uptake, and hospitalization while also considering the relationships between locations. These methods will generalize beyond Sars-Cov-2 to other

emerging and re-emerging infectious diseases and apply broadly to other public health crises.