Models correctly predicted 200K deaths. Here's what they're warning for the coming months.

Denise Chow and Joe Murphy
·6 min read

When the coronavirus pandemic hit the United States this year, scientific models forecasting hundreds of thousands of deaths were met by some people with derision.

The models, unfortunately, have been vindicated. And they're providing fresh warnings that a recent uptick in numbers of cases could mean the U.S. death toll could almost double in the next four months.

"If we go back to March, at that time, we were saying if this thing is not handled very carefully, we could end up with 200,000 or 300,000 deaths," said a coronavirus modeler, Alessandro Vespignani, director of Northeastern University's Network Science Institute. "At that time, everyone was saying that's impossible. I think we should use that perspective now, especially when we think about the future."

After beating back an initial wave of coronavirus infections, some countries in Europe find themselves in familiar territory: facing a spike in the number of new cases and weighing which restrictions could help drive the numbers down. In the U.S., after a brief dip earlier this month, the number of new cases daily is creeping up again. Since Sept. 18, the seven-day average of new Covid-19 cases in the country hasn't fallen below 40,000 a day, according to an NBC News tally.

For coronavirus modelers, the writing has been on the wall. Many have watched with a mixture of horror and frustration as their projections of the pandemic's evolution, and its potential death toll, have come to fruition.

Now, a widely cited model developed by the Institute for Health Metrics and Evaluation at the University of Washington suggests that the U.S. could total more than 378,000 coronavirus deaths by January.

But infectious disease modeling can be a tricky science easy to criticize for its uncertainties. Experts say coronavirus models have come a long way since the early days of the pandemic, to the point where some researchers are moving away from long-term projections and focusing instead on forecasts that can more accurately predict Covid-19 trends up to six weeks.

Not all the models were accurate. Their biggest mistake came in terms of Africa, where some predicted severe outbreaks. Instead, much of the continent has avoided the worst of the pandemic.

The director of the institute, Dr. Christopher Murray, a professor of health metrics sciences at the University of Washington, said his team's U.S. model has undergone numerous refinements throughout the pandemic. Behavioral changes — such as diligent mask-wearing — could drive their projections for January down, but he also worries about fatigue settling in.

Murray said the new trajectory can already be seen in some European countries, including Spain, France and the United Kingdom.

That's why modelers are hoping people heed their warnings about the coming weeks, when, they say, growing complacency and changing behaviors tied to the fall and winter seasons could result in a new wave of infections.

"I think some people think the worst is over," he said. "That progressive decline in vigilance will fuel part of the fall and winter return."

The University of Washington institute's model, which is one of several the Centers for Disease Control and Prevention uses to track the pandemic, has been criticized for often including high degrees of uncertainty, which can lead to imprecise predictions. Early on, the model underestimated the number of Covid-19 deaths nationwide, projecting that the U.S. could hit 60,415 by the end of August.

Still, the model is updated frequently, and refinements are made as data on case numbers, hospitalizations and a host of other factors become available. By June, the institute's model was estimating that the U.S. death toll could hit 200,000 by Oct. 1, a projection that ended up being accurate to within two weeks.

But infectious disease models are never static, and several unknowns could significantly alter the existing projections.

One such factor is how the virus's spread may be affected by the changing seasons. No firm evidence suggests that the coronavirus will be more or less transmissible in the fall and winter. Rather, it's the effect that falling temperatures have on human behavior that concerns researchers, particularly because cold weather is likely to draw people indoors and make it difficult to practice social distancing.

"In the winter, people tend to stay inside, which could make it easier to transmit the disease," said Sen Pei, an associate research scientist at Columbia University, who has done extensive Covid-19 modeling work. "But we still don't know how the virus will perform in the winter."

Pei said that there were enormous challenges with modeling a novel coronavirus but that with nine months of data from the pandemic, his team's projections have become significantly more sophisticated. Yet one of the most difficult things to predict in a model is also one of the most important factors that could change the outcome of an outbreak: how humans respond.

"It's a fluid situation, because people's behavior changes over time, which is essentially unpredictable," Pei said.

The uncertainty is partly why Pei and other modelers avoid long-term projections like the institute's model and focus instead on producing short-term outlooks for the next four to six weeks.

"Nobody really knows what's going to happen past the next few weeks," said Youyang Gu, a data scientist who runs a coronavirus model known as Covid-19 Projections. Gu, who doesn't have a background in epidemiology or infectious disease modeling, designed a model that uses machine learning to "study" certain parameters that evolve with the pandemic, such as the virus's reproduction number, or R-naught, which represents how contagious a disease is.

"We don't rely on any implicit assumptions," Gu said. "We look at the data and say: This is what we learned from what is happening."

Gu said his model, which runs forecasts only until November, was able to predict that the surge in new cases in June and July wouldn't subsequently lead to an equivalent spike in deaths on par with what the country experienced in March and April.

"We compared what happened in the U.S. to other places around the world, and the data didn't support deaths' going up as quickly as cases," Gu said. "We ended up peaking at about 1,000 deaths per day, which is obviously still very significant but less than what a lot of people in the scientific community had been expecting."

The shift from long-term projections is a desire shared by other modelers, who say long-term projections are often less accurate because they need to include a wide range of estimates to account for uncertainties. By January, for instance, travel bans, lockdowns or other restrictions could be introduced, dramatically altering long-term predictions.

The switch "allows us to get away from these scenario projections that we were initially doing and move closer to forecasting, which is the goal," said Shaun Truelove, an assistant scientist and modeling expert at the Johns Hopkins Bloomberg School of Public Health. "The forecasts are more understanding what is really going to happen given the situation, rather than this is what could happen."

Vespignani likened it to weather forecasts, which are harder to nail down the further out they target. He said he hopes people will pay close attention to coronavirus forecasts, especially as the country braces for what could be an uptick in new cases in the coming weeks and months.

"We still have quite a run ahead of us," he said. "We have to fight this battle, because it's not over."