Why the SEIR COVID-19 Model Will Help Us Get Back to Work in the Fastest and Safest Way
Categories: Data, General, Tech & Data Insights
Grand Rounds is helping innovators and policy leaders across the country answer the COVID-19 multi-billion dollar question on everyone’s mind: when can we get back to a “new normal” in this post-novel coronavirus world and open up the economy?
Accurately predicting when the United States will be at peak COVID-19 infection through statistical models has been a huge focus among government and private sector leaders. If we could understand when, and by how much local capabilities to deal with the crisis will be exceeded, we can create strategies to safely balance mitigation and reopening. However, COVID-19 doesn’t affect each person and population in the same way, posing many challenges for data models. Moreover, policies are changing every week. The prediction from a model is only as good as the ability to measure and model policy shifts at different levels of effectiveness. No single model will be perfect, but there are models emerging that epidemiologists believe will have far more accurate predictions at the local level which can be used to provide targeted and granular scenario planning.
For a governor, mayor or business leader with hundreds of thousands of employees, we believe that you need as many inputs as possible to understand how the disease will likely affect your geography as well as updating tools to understand and plan in times of high uncertainty.
To date, several models have emerged to the forefront, but only one of them provides a local-level lens with the epidemiological rigor and flexibility to keep up with COVID-19, a SEIR model. SEIR is an acronym for susceptible (S), exposed (E), infected (I), and resistant (R).
At Grand Rounds, we believe there is immense value in SEIR modeling. Through our recent partnership with CovidActNow.org, Grand Rounds is making its SEIR model available for anyone to use.
What is a curve fitting model?
Curve fitting models, like the model made by the Institute for Health Metrics and Evaluation (IHME), look at how the disease progressed in other geographies, like China, Italy, and Spain and tries to extrapolate a prediction from there. The model tries to approximate the disease’s progression in the U.S. by fitting a best fit growth curve based on how the disease progressed in those other geographies who are farther along with their COVID-19 infections.
While this helps provide us with a ballpark prediction, the model would not be able to account for differences in how COVID-19 might spread in other geographies, differences in how long individuals might be infectious, changes in treatment patterns, and most importantly, changes to public policy that alter the trajectory of the disease spread. That is because this approach is not modeling the disease dynamics itself, as reflected in the huge fluctuations in IHME’s projections and the inability to explain this variability.
Our communities need more than ballpark predictions. Because the IHME model relies on total death counts and aggregates the data up to the national level, the IHME model gives the impression that the deaths in the U.S. peaked mid-April. This gives the false sense that a community may be ready to relax social distancing when in fact their community has not even been hit yet. At the time of this writing (late April), the country’s death toll has surpassed 50,000 with little sign of full containment as a high-likelihood scenario. For reference, IHME initially forecasted 60,000 deaths throughout the entire duration of COVID and continues to increase death projections on most updates.
What is a hierarchical model and how is that different from a curve fitting model?
The Grand Rounds SEIR model is a hierarchical model, allowing us to generate hyperlocal projections (ex. down to the zip code) in a statistically robust manner.
SEIR models are commonly used by epidemiologists to predict how a disease will be transmitted through a population by modeling out the disease pathway, taking into consideration how the underlying variables change in response to each other. The SEIR model, when applied to COVID-19, categorizes people into one of these states: susceptible to infection (S), exposed (E), infected (I), and recovered (R) or deceased. These can be richly extended to incorporate demographic features, specific transmission patterns such as workplaces or schools, as well us understanding impact on resources such as protective equipment or hospital systems.
By categorizing people as existing in one of the various states and then modeling how people progress through them, we can modify the individual parameters and incorporate different data that might or might not be present or statistically robust at the local level. For example, constraints on parameters such as rate of growth pre-infectious period, rate of hospitalization, and ICU case fatality rate can be inherited from the national level and interventions can then be modeled at state, county, or workplace specific levels.
SEIR thus offers leaders and policy makers the most flexibility and nimbleness when it comes to understanding the range of impact that COVID-19 might have. Again, COVID-19 infections are growing at an exponential rate so slight variations in the data will have profound effects on the projections and as a result policy decisions.
Projecting how COVID-19 will impact our communities, at the local level, is essential for an effective response to this disease in the absence of a vaccine.
We take a deeper dive into the various stages of the SEIR model and how that contrasts a curve fitting model, like IHME, on our tech blog.
Why is local modeling so important?
Healthcare in this nation is practiced locally. For better or worse, this decentralized approach to healthcare and public health policies means that different regions in the country have varying abilities to react to the pandemic. It’s why the predictions for California’s case peak differs from New York’s. It’s also why the predictions for San Diego’s peak are different from San Francisco’s.
Furthermore, local modeling will help us assess when it’s safe and appropriate to reopen a local geography’s economy, as well as mitigate that area’s healthcare system from being overburdened.
The data from Wuhan, China or Seoul, South Korea will not provide us with an accurate prediction of when Tallassee, Tennessee will reach its peak number of cases. And without an abundance of granular local data around this disease that humanity has never seen before, we know our communities need a model that can project down into lower level geographies in a statistically robust manner.
COVID-19 will be with us for some time and the effects will be staggering. Moreover, it is likely that we will face the secondary waves that Singapore and Japan are currently encountering. To that end, we believe we can improve our response to future waves by incorporating more data sources. Having a well-engineered software solution that automatically can return daily projections as soon as the earliest warning signs appear can save lives while also benefiting economies.
We believe our model can better identify the people that are unable to access routine care and are now most in need of additional help. We want to help those people fill the gaps in their care so they can avoid the emergency room or hospital for critical care. We also want to help those who do recover from COVID-19; individuals who received the most serious interventions (ex. ventilator breathing assistance)j during this time will need extensive care coordination for many years to come.
The Grand Rounds COVID-19 SEIR model is open source and you can view the code here.