It’s another morning in Albany and Gov. Andrew M. Cuomo is back before the cameras, his hand on the clicker of a PowerPoint presentation packed with rows of figures, graphs and computer projections, some of them dire.
“We have a model, a projection, and that’s what we follow,” Cuomo said from the wood-paneled room one day last week.
As the unfolding coronavirus pandemic crests a plateau in the state, epidemiological computer modeling has been thrust into a front-and-center role in the public response like never before. Compared with other Cuomo-managed disasters marshaled from snowbanks, beside floodwaters or highway reststops, the most consequential work in observing and planning for response to the coronavirus is happening largely behind the scenes, guided by legions of computers at leading universities engaged in predictive modeling.
Since the first case was confirmed in the state March 1, these computer models have refined their projections with millions of new data points to provide the most educated guesses about how COVID-19 will spread or recede. They’ve been the basis for desperate pleas for more hospital beds and ventilators, how to deploy increased staffing levels, expectations for the next hot spots, and estimates for how many are likely to die.
"The point is to provide guideposts, to show what has been going on and are we on a trajectory that makes sense for reducing this," said Jeffrey Shaman, director of the Climate and Health Program at Columbia University, which has modeled the virus with specific impacts to New York City. He stressed that the models "aren't forecasts, they're projections. We're dealing with a situation where we can affect the outcome by what we do today."
For that reason, the projections in the models have shifted over time. Cuomo last Thursday showed the wide variation of conclusions in models, including Columbia's "nightmare" projection that 136,000 people could be hospitalized in New York City alone. As of Thursday, the number hospitalized statewide stood at 18,000.
Cautioned Shaman, "When we make a projection there’s a big uncertainty range around the scenario."
The mathematics also have been used to justify unprecedented business shutdowns and persuade people of the need for social distancing, information that is fed into the models to graphically demonstrate how effective quarantines can be.
No modern-day catastrophe has relied so heavily or publicly on mathematical modeling to anticipate impacts and plan for worst-case scenarios than the coronavirus pandemic, the modelers say. The most sophisticated models have been around for a decade, and they’ve grown in complexity and accuracy with the advent of supercomputers and the urgency of health crises, including the SARS, N1H1, HIV/AIDS and Ebola epidemics.
“The models are relatively well established,” said Steven Skiena, director of Stony Brook University’s Center for Artificial Intelligence-Driven Innovation and Discovery, who has worked on predictive models outside the field of epidemiology. “How you fit in parameters and how people respond to different events, that is a challenge that requires care and judgment,” making each new model unique.
One of the most widely used epidemiological models during this crisis is called the compartmental model, which attempts to develop insights into the potential spread of a virus by defining and compartmentalizing segments of the population into basic categories: the susceptible (or uninfected), the infected, and the recovered. As more data on the shift of people from one compartment to the next becomes available, researchers, feeding the data into mathematical computer models, can begin to develop a profile of the virus itself, and begin to make educated projections about needs and outcomes.
The models grow in complexity with the more data that’s added and as a pandemic's timeline progresses. They factor in compartments with more granularity: the infected but asymptomatic, the infected and symptomatic, and the infected and hospitalized, according to a primer on modeling provided by Ohio State University.
“We are trying to predict the peak number of cases, how many that will be at the peak, to inform whether hospitals will be overwhelmed, when we will be on the downswing, when we can come out of the isolation, and how many are going to die,” said William Miller, a professor of epidemiology at Ohio State University.
From COVID-19’s earliest days, public officials from the White House to local government bunkers have trotted out these graphs to explain the wide disparity of potential impacts, to motivate populations to comply with social distancing orders and to demand and marshal equipment among hospitals.
More importantly, by offering the ability to factor in variables such as the potential effect of social distancing, they can show how specific changes in human behavior can alter the outcomes, chiefly by limiting infections and death. That’s helped motivate compliance.
It's also frustrated some who have used the model's projections to paint dire consequences, and set the stage for massive movements of equipment and staff. "They have turned out to be wrong," Cuomo has said more than once of the models.
Right or wrong isn't the point, modelers say.
"They're actually surprising accurate given that we’re trying to predict the future, but I think they are sometimes taken a little too concretely. There’s a lot of variance," said Sean Clouston, associate professor for Stony Brook University's Program in Public Health at the Renaissance School of Medicine.
For weeks, he's been working on models to project the virus' potential impact on the university hospital and Suffolk County, work crucial to determining how to plan for hospital capacity.
So far, it's continuing to project that Suffolk will see a lower and slightly later peak infection rate than New York City, and a lower percentage of the population infected because "we have lower contact with one another," all reflected in the models, Clouston said.
The models are fed constantly. "Every day it changes basically," he said. "As things go forward you try to use as much as possible to inform" refined projections.
One of the earliest U.S. models in the pandemic was developed by the University of Washington’s Institute for Health Metrics and Evaluation. It was considered instrumental in Washington state’s early recognition of the pandemic’s potential spread across the state after a nursing home became overrun with infections, according to a report in Wired, and it was used to provide a graphic illustration of how instituting precautions such as business closures and social distancing could impact the spread of the virus. It’s a factor being credited for Washington’s relatively low infection rate now, despite being the country’s first epicenter.
Experts say it's no surprise that lines between projections and actual results begin to intersect the longer the pandemic continues, and the more data that's piled in.
“They will get more accurate the longer this goes on and the more they can refine the parameters that underlie these models,” said Skiena, the Stony Brook director.
With the increasing accuracy also comes refinement of the models.
Last week, the Institute for Health Metrics and Evaluation revised forecasts in a model it offers for each U.S. state to determine how many hospital beds, intensive care units and ventilators are needed, as well as project how many will die.
“As we obtain more data and more precise data, the forecasts we at IHME created have become more accurate,” Christopher Murray, director of the institute, said in a publicly issued statement. Equipment and hospital beds shortages have been revised downward, and high-end of the range for estimated deaths has been lowered, most recently to just over 60,000 from a one-time "best-case" range of 100,000 to 200,000.
For Cuomo, the revisions have been mostly good news, even to the point of providing a rallying cry.
“The statisticians when they did their curve did not know how New Yorkers would respond and didn’t know whether or not New Yorkers would comply,” he said Friday, calling New Yorkers unified responsible and caring. “That’s what they couldn’t count in those models … That’s what they couldn’t figure out on their computers.”