On The Horizon: Climate Scientists Forecast the Flu
For most of us, the flu is a minor inconvenience. A box of Kleenex, a day or two of missed work, and that’s it. But not everyone is so lucky. A bad case of the flu can send a person to the hospital, or even to the grave.
The number of American fatalities associated with influenza—a highly contagious respiratory virus—varies wildly from year to year: in the winter of 2011–2012, 12,000 people died; the next year, nearly 50,000. During the 1918–1919 Spanish flu pandemic, more than a quarter of the country caught the bug, and 675,000 died. The virus claimed more American lives than World War I and, by some estimates, 50 million people around the world were felled by that one pandemic.
Predicting the magnitude and timing of outbreaks has long been a guessing game, because the information we’ve had about influenza has been old and incomplete. Like the visible tip of an iceberg, the data on documented cases fails to account for the underlying dynamics of how the virus spreads. But in recent years, scientists have developed methods drawn from computer modeling and weather forecasting to provide a more complete picture of these invisible forces.
Every year since 2013, Jeffrey Shaman, PhD, associate professor of Environmental Health Sciences and director of Columbia’s Climate and Health Program, has published weekly forecasts of the flu season. Specific to 108 cities across the country, the online projections predict whether cases are expected to rise or fall and by how much. A new version of the system drills down to the neighborhood level, using data on commuter patterns. In New York City, where it was tested, the system might, for example, tell residents in the South Bronx that they are more at risk than someone living in Soho.
In temperate climates like North America’s, the flu arrives every year sometime between October and May. What drives this seasonal pattern has been the subject of scientific debate. Some have argued that reduced sunlight exposure during the cooler months suppresses our immune systems; others, that more time spent indoors increases our exposure to infected people. In the fall of 2007, Shaman, a Columbia University–trained climate scientist with a special interest in hydrology, had a hunch about humidity. Reanalyzing data from an experiment involving guinea pigs, he found that drier air increased the likelihood that the lab animals would catch the flu. Using that insight, he created a computer model that was able to reproduce the average rise and fall of three decades of human cases.
While Shaman’s humidity model could predict the average timing of many flu seasons, it wasn’t powerful enough to draw conclusions on any single season. In December 2008, Shaman and his former PhD classmate Alicia Karspeck discussed the problem over lunch. Karspeck, now a scientist at the National Center for Atmospheric Research, suggested that they adapt the mathematical methods used in meteorological forecasting, which simulate the weather based on thousands of observations—from weather satellites, ocean buoys, and the like—and extrapolate into the future.
Using a similar approach, Shaman and Karspeck built a computer model to simulate how influenza spreads through a population, drawing on hospital admissions reports and data provided by Google on cold- and flu-related search queries. Their model generates hundreds of estimates, creating visual “spaghetti,” with some trajectories indicating a sharp rise in sickness and others a gradual uptick or downward trend. Then the algorithm applies a technique from weather forecasting called data assimilation to weave those strands of spaghetti into a single prediction.
To test their model, the pair made what’s known as a retrospective forecast, using historical data and comparing their model’s prediction to actual disease patterns. Their test run, on five flu seasons in New York City—published in 2012 in Proceedings of the National Academy of Sciences—showed that their computer model could predict the peak of the outbreak more than seven weeks in advance. “The model mimics the behavior of the outbreak as it has thus far transpired,” says Shaman. “Doing so gives you a better chance of predicting where it will go in the future.”
Soon Shaman was generating his predictions in real time. When he first showed his work to officials at the U.S. Centers for Disease Control and Prevention back in 2012, he was met with skepticism. Undeterred, he sent them unsolicited weekly forecasts. Then in 2013, CDC officials invited Shaman’s group and five other teams to forecast the current flu season. The Mailman School team routed the competition. Their triumph owed to the use of the same kind of confidence probabilities people have come to expect from weather forecasts—if there’s an 80 percent chance of rain, we’ll take an umbrella. Similarly, influenza forecasts have the potential to provide actionable information to both health officials and the public.
One administrator at a large medical center told Shaman that knowing the chances of a surge in symptoms would help hospitals avoid being caught off guard, short on supplies like gloves and masks or with too few clinical staff on hand. Greater awareness could help the public, too, says Shaman, by motivating people to take steps to avoid getting sick in the first place. He foresees a day when his forecasts become a regular part of local weather reports, alongside pollution levels and pollen counts. “Flu forecasts,” he says, “can help people stay informed and consider whether they’re up to date on their shots and whether they might need to keep their kids home from school if they’re beginning to get sick.”
Shaman and his collaborators are still refining their approach. To boost the model’s accuracy, they’ve started combining data on hospital admissions for flu-like symptoms with cases of influenza verified by a lab, helping them to distinguish respiratory viral strains from the common cold. More recently, Shaman’s team—including Wan Yang, PhD, associate research scientist in Environmental Health Sciences, and epidemiologists at New York City’s Department of Health and Mental Hygiene—reported success in generating forecasts for specific neighborhoods. The results, published in PLOS Computational Biology in 2016, showed that their algorithms could provide nearly the same local specificity as weather forecasts—something important, notes first author Yang, because “most public health decisions are made on the local level.”
Meanwhile, the researchers tested their forecasts in Hong Kong, where semitropical conditions allow influenza rates to ebb and flow throughout the year. In partnership with researchers at Hong Kong University, Shaman and Yang tweaked their computer model to generate three-week forecasts with up to 93 percent accuracy. The team is now setting up a system for real-time predictions to guide public health decisions for the city—and potentially beyond. “Having a foothold there and being able to make predictions in that area of the world is going to be very valuable on the global stage,” says Shaman, who notes that new flu strains, particularly pandemic varieties, often emerge from Southeast Asia.
In recent years, Shaman and Yang have extended their approach to other pandemic threats. In the summer of 2014, Shaman got an urgent call from the CDC. The official, familiar with his flu forecasts, wanted to know if the professor could help him get a handle on the rapidly worsening Ebola outbreak in West Africa. Despite a lack of research into the disease dynamics, the strength of the outbreak “signal” allowed Shaman and Yang to generate forecasts. These and other predictions, along with traditional surveillance on the number of patients being treated, helped officials understand the urgency of the situation and commit resources to fight the epidemic. “In a situation like the Ebola outbreak where there is little data available,” says Shaman, “forecasts provide information to guide decisions so officials aren’t completely flying blind.”
Near the end of the outbreak, Shaman’s team obtained district-level data in Sierra Leone. With the additional information, they were able to recreate the internal dynamics of the outbreak and identify critical junction points in the districts of Kenema and Port Loko. In a retrospective analysis published in 2015 in the Journal of the Royal Society Interface, Yang and Shaman showed that had the forecast been done in real time, officials might have interrupted the epidemic’s spread by focusing interventions on those two districts.
No matter the finer points, generating accurate forecasts requires a thorough understanding about how contagions spread—the underwater part of the infectious-disease iceberg. In ongoing work, Shaman’s team is developing increasingly complex models that account for hundreds of subpopulations, and stitching together multiple models into “superensembles,” evocative of the approach meteorologists use to anticipate the path of a hurricane. Other models incorporate the life-cycle dynamics of the insects and animals that spread disease, such as the 70 mosquito and 300 bird species that carry West Nile virus. Ongoing work with colleagues at China’s Academy of Military Medical Sciences focuses on H7N9, an emerging and deadly strain of avian influenza that kills approximately one-third of those infected. Currently, the illness spreads exclusively from birds to humans and not among infected people. But health officials worry that if the virus mutates so that it can spread person-to-person, the consequences could be catastrophic—in Shaman’s estimation, “Ebola on steroids.”
Ultimately, computer simulations still rely on observational data. Over the years, investments in more-sophisticated weather satellites have bolstered public confidence in weather forecasts, says Shaman. Likewise, he suggests, health officials should do more to seek out information on illnesses, particularly as the cost of identifying pathogens drops. In the meantime, the scientist is undertaking an $11 million study funded by the Defense Advanced Research Projects Agency to understand the transmission dynamics of respiratory illnesses in New York City. Shaman’s research team is collecting nasal swabs at local high schools, day care facilities, and a pediatric emergency department to understand, among other things, how many people are infected with mild illnesses and what role these semihealthy people play in the spread of disease. The U.S. Department of Defense is funding the study, in part, says Shaman, because “in the history of the U.S. military, more people have died from infectious disease than in combat.”
This kind of proactive approach to data collection is central to efforts to anticipate infectious outbreaks, with implications from aids to Zika, as well as more obviously weather-dependent conditions like diarrheal disease and famine. In theory, even social contagions like obesity and opioid addiction could also be forecast. “There is a lot of room to grow,” says Shaman. “We’re only just getting started.”
Tim Paul edits the Mailman School's Transmission newsletter. His report on immigration and refugee health appeared in the 2016 edition of this magazine.