Q&A: Why COVID-19 Modeling is Vital to Kentucky

LEXINGTON, Ky. (May 27, 2020) — Teresa Waters is a professor and chair in the University of Kentucky College of Public Health’s Department of Health Management and Policy. Waters has been thoroughly involved in UK’s, and the state of Kentucky’s response to COVID-19.

UKNow recently spoke with Waters to learn more about the pandemic and how it is affecting Kentucky.

UKNow: What specifically have you been doing to help UK and the state respond to COVID-19?

Waters: In response to the COVID-19 pandemic, both UK Healthcare and Kentucky’s Cabinet for Health and Family Services (CHFS) formed working groups to model virus spread, hospitalizations, and deaths. There were many publicly available models offering a wide range of (mostly dire) predictions. I am not an epidemiologist but wanted to use my experience studying healthcare delivery to support these efforts. My primary role has been to advise the epidemiologists and other data scientists on strengths and weaknesses of models and appropriate data sources to use.

UKNow: How is this type of modeling used by local and state health departments, hospital administrators, and government agencies?

Waters: These models are useful for preparedness planning because they make predictions about the range of outcomes (infections, deaths, hospitalizations) that we can expect, depending on actions taken by individuals and organizations (e.g. social distancing; closing schools). Our mantra has been: hope for the best, but prepare for the worst, because this saves lives. UKHC and CHFS have used the COVID-19 models to “prepare for the worst,” setting up a field hospital for a potential surge in cases (UKHC) and allocating resources and attention to hotspots in Kentucky (CHFS).

UKNow: What data do you need to produce models of this pandemic?

Waters: Different models require slightly different inputs, but generally we need information on:

• Size and age of population at risk for exposure to COVID-19. Size because this will ultimately inform how many hospital and ICU beds we need, and age because we know that older individuals are more likely to be hospitalized and have poor health outcomes (ICU use, death).
• Initial reproductive rate (R0) and reproductive rate over time (Rt). R0 tells us how infectious COVID-19 is and how fast it will grow in the community if unchecked. Rt is the reproductive rate that we see after public health measures are put into place (e.g. social distancing, closing schools/businesses)
• Date of first (community acquired) COVID-19 hospitalization in a locality. This gives us an “anchor point” for modeling infections, hospitalizations and deaths over time. Data from around the world suggest that COVID-19 incubates for about 10-14 days. So while we can’t observe date of first community infection, we can observe first hospitalization and make an educated estimate of when infection started in a local area.
• Date of social distancing measures. While we did not know the exact impact of these measures, we did know that we would need at least 10-14 days to start seeing their impact, given COVID-19’s incubation time.
• COVID-19 hospitalization rates, percentage of cases in the ICU, average length(s) of stay. Preparedness planning focuses on how many people are in the hospital at the same time. That means we need to know how many people who are infected with COVID-19 will end up in the hospital, how many will end up in the ICU, how many will need ventilators, and how long they will stay (or need a ventilator). We have been using historical data (first from Wuhan and Italy, later from US data) to estimate hospital use and “census” (how many at the same time). “Peak census” tells us the maximum number of beds or ventilators we need in a particular area; we can compare that to other data we have on capacity (from the state).

UKNow: Over the course of this pandemic so far, we have seen the models shift a great deal. Why does that occur?

Waters: That is an excellent question, because many folks find this confusing. There are a few reasons that the models have been shifting. First, because we are getting better information about key model parameters. For example, our early models relied on data from Wuhan and Italy, because that was the only data available. We know that the US has a very different (diverse) population and our healthcare system is quite different. When we were able to get US-specific data, our models were definitely improved.

Another reason that models shifted is because we were seeing the impact of social distancing measures. Early on, we did not know how much social distancing would reduce the reproductive rate (Rt) of COVID-19. Our data now suggest that we are having great success in reducing that number to 1 or below. And we need that number to be below 1. Above 1 means that spread is expanding; below 1 means that it is contracting.

UKNow: So far, Kentucky as a state has not experienced as many hospitalizations or deaths as were once predicted. That’s good news but does that mean the models were wrong?

Waters: COVID-19 models are built to predict a range of outcomes, using the best available data, and depending on actions taken by individuals (e.g. social distancing measures).  The data have definitely gotten better, so that is part of the adjustment. But it is important to understand that models are not making hard and fast predictions; instead, they are offering us choices and paths.

One choice might be to do nothing (or very little) to prevent spread. That option has been modeled, and the predicted outcomes are bad. Our early models suggested that local hospitals and ICUs would have insufficient capacity to treat all the COVID-19 patients with no social distancing. A paper published by my colleagues in Gatton (ref) estimated that without social distancing, there would have been thousands more deaths in Kentucky. 

UKNow: What data do we need – or need more of – to continue to produce useful models of the pandemic?

Waters: Things have been relatively flat in Kentucky for several weeks, although some areas of the state are experiencing more infections than others. At this point, we need to continue to collect daily information on cases, hospitalizations and deaths, including geographic information (e.g. which county?) so that we can closely track the impact of relaxing social distancing measures. That will allow us to anticipate localized spikes and re-implement social distancing measures as needed to slow spread.

UKNow: Finally, when will we see a model or prediction cited in the news, what questions should they ask themselves as they interpret it?

Waters: I think readers can ask a few key questions:

1. Does the model use the best and most up-to-date data in making projections? As we’ve discussed, data quality can have a big impact on projections.
2. Who built the model and how many people are helping to refine it? The best models have been developed by world class data scientists and leverage a lot of input from a range of experts.
3. How do model predictions compare with actual outcomes over time? At first, we did not have historical, localized data, making it very hard to know how well models were performing. Now we have that data and can compare predicted trajectories (cases, hospitalizations, etc.) with actual experience.