Using Data Science in the COVID-19 Pandemic in West Virginia

• Read original article here

As one of the team members charged with working with data during the COVID-19 pandemic in West Virginia, I noticed there has been an overall theme: Communicate and share the necessary information to keep the 1.79 million residents, health care facilities, and economy of West Virginia protected.

To do this, groups from all over the state—including the Department of Health and Human Resources, West Virginia Governor’s Joint Interagency Task Force on COVID-19, West Virginia National Guard, West Virginia Clinical and Translational Science Institute, and Data Driven West Virginia—partnered to share data and find solutions for West Virginians. The goal of every analysis was to provide pandemic leadership and policymakers with actionable insights based on relevant data.

While this may seem obvious, the ability to operationalize resources and the pandemic response because of insights or recommendations is key to gaining the trust of leadership and the public while working in dynamic situations.

Many methods and techniques have been used to deploy testing resources, provide vaccinations, and track vaccination metrics. We continue to develop new methods to adjust for various changes in the disease, policies, and public sentiment. Though these new developments are important, the most critical tool we have is the ability to collect and maintain visibility into the appropriate data for decision-making. Without the appropriate data, the software adage of “garbage in garbage out” will hold.

Many of the analyses performed throughout the pandemic required the number of daily confirmed Sars-CoV-2 cases and the number of Sars-CoV-2 PCR tests administered daily for each of the 55 counties in West Virginia. Also, once vaccines became available, vaccination data for each county stratified by target age groups was available. While information such as this could produce localized public health measures such as the real-time reproduction number (Rt), positivity, testing, and vaccination rates, it could not provide visibility into health care resources in this community.

To do this, the West Virginia Hospital Association provided a daily survey to understand how many patients in each health care facility around the state were COVID patients and how many were being treated with COVID protocols prior to being a confirmed COVID-positive patient. Furthermore, this survey collected information about the number of patients in intensive care units and on ventilators.

During the early phases of the pandemic, when concerns centered on personal protective equipment for frontline workers, a second survey by the West Virginia Hospital Association was built to include questions about amounts of PPE and expected PPE use. Collectively, this data painted an operational picture of resources used for COVID in West Virginia hospitals.

As the pandemic continued, the importance of data sharing became more apparent, especially as West Virginia’s leadership made the critical decision to forgo the federal pharmacy program and instead use local pharmacies to prioritize nursing homes and the 60+ population for vaccination. The Joint Interagency Taskforce developed a novel vaccination inventory management system based in Shiny to manage supply and demand to distribute vaccines. This system allowed the multiple entities in the taskforce to share requests for vaccines and see the allocation being distributed around the state in real time. As supply exceeded demand, the system allowed vaccine providers to directly request vaccines and monitor orders and scheduling of vaccinations in their communities.

It became critical to understand the number of vaccinations administered to West Virginians once vaccines were distributed. Data from West Virginia was combined with information from the Centers for Disease Control and Prevention to better understand the demographics of West Virginians who were vaccinated.

The last point we must make about data quality before we can get to the statistical analyses, models, or machine learning aspects of the tools we used is that the COVID-19 pandemic has been dynamic, to say the least. Not only have protocols and behavior changed, but the virus has evolved. The variants present challenges through changes in transmissibility, as well as reacting differently to treatments.

We have seen the virus be dynamic, but the data and policies around how data is collected have also become dynamic. We have moved from mandated testing to facing a negative public sentiment around testing, affecting our ability to understand case counts and positivity rates relative to other times in the pandemic.

The pandemic has created a public health data infrastructure investment that should be kept and maintained for all public health crises, because the first step to any critical response is having access to the right information. If we have learned anything since March 2020, it is that data infrastructure in public health is part of the national critical infrastructure.

The statistical tools that have been used during the pandemic response have also had to reflect the dynamic nature of the data. Additionally, a balance has had to be struck between what is scientifically interesting about COVID-19 and what is operationally actionable for decision-makers. While the former may produce publications, the latter provides the necessary analyses to deploy resources that directly affect the pandemic.

For instance, forecasting the number of COVID-19 cases is important to understand, but the question is what resources it affects. Do county-level case counts help deploy PPE to hospitals? Do these case count forecasts provide insights about the number of tests that need deployed to an area? Or do they provide guidance on the number of hospital beds or ventilators that must be available to maintain care for that population? This creates the distinction between the scientific question and the operational question.

We must also make sure behaviors and changes of behaviors throughout the pandemic are taken into consideration. For instance, how people seek care is a key component of describing and preparing for any hospitalization surges that will occur. In West Virginia, we found the way individuals seek care for COVID is similar to how they seek care for emergency room visits, thus we use that as a proxy for how each individual in the state will seek care during any type of surge.

In the early stages of the pandemic, this behavioral type analysis set the base for the suspectable population for the adjusted compartmentalized models we used to develop PPE forecasts. As more data became available and supply chains became more stable, these methods gave way to more traditional statistical and inventory control methods.

Most recently, we have been developing machine learning methods to inform targeted testing events. This tool has been used to deploy testing resources provided by the National Institutes of Health RADX-UP project to localities in West Virginia through community lead events.

COVID-19 continues to affect our communities. As data has become more abundant and the disease has become more dynamic, data analysis has become more difficult. One of the key components to any success thus far has been insights delivered from data. It is of the utmost importance that we continue to develop these methods and the infrastructure required to respond to health crises.