After October, the seasonal march of the cold, wet months and their viral baggage can down much of an office, or school, for weeks. But now, researchers are radically cutting out the guesswork as to when that might happen in the stretch between winter and late spring: An algorithmic model developed by scientists at Columbia University’s Mailman School of Public Health is the first to reliably predict flu outbreaks up to nine weeks in advance in 108 American cities.
“This [flu forecast] is for predicting when the time of the peak would be,” explains Jeffrey Shaman, lead author of the paper showing results from the last flu season published in Nature Communications this week. “So about six weeks in advance of the peak is when we started seeing a lot of movement and change.”
Shaman’s model, something called the susceptible-infected-recovered-susceptible model, or SIRS, uniquely factored in weekly reports from both Google Flu Trends and the CDC. Last year, Google Flu Trends, which derives local flu estimates from measuring panicked flu-ish search terms around the globe, came up with a set of measurements that far overestimated who was getting sick–and was promptly slapped on the wrist by a report in Nature showing where the search engine went wrong. Google altered its Flu Trends algorithm in response, but SIRS also combines Flu Trends reports with real-time weekly CDC data of influenza diagnoses.
Shaman acknowledges that Flu Trends veered wildly off-course last year, estimating nearly twice the number of cases reported by the CDC, a problem which he kept track of and documented as well. “But you shouldn’t throw the baby out with the bathwater,” he said. “[Flu Trends] had a very bad year. It overshot the mark and produced unrealistically large outbreaks.”
But considering the disadvantage, the Columbia model still produced unprecedentedly accurate predictions in 63% of the cities it analyzed. Much of that had to do with machine learning: The Columbia model trained itself with incoming data by week, tweaking its prediction trajectory, and growing more accurate over time.
Some cities were easier to predict than others. Because the model assumes that everyone gets exposed to everyone else in a certain area, Shaman found that smaller areas, smaller populations, or cities with high population densities yielded the most accurate predictions. Diffuse, spread-out cities like Los Angeles challenged that assumption, but the researchers’ prediction that Portland’s flu outbreak would peak four weeks into the season, for example, was spot on. Outbreaks in Birmingham, New York City, and Kansas City also struck with nearly 100% accuracy to the model.
Next year, Shaman hopes to fine-tune these predictions, using cities within Los Angeles County, or predicting borough-by-borough in New York City. But as Shaman and his team learn more about predicting the flu, he hopes the research can give public health officials a tool to better prepare for outbreaks in the future, just as a city might prepare for a tornado or hurricane. Better yet, as the model becomes increasingly accurate, why not just throw a flu forecast up on TV?
“If [flu forecasts] become accurate or reliable enough, you might end up seeing them on the local weather report,” Shaman said. “It’s entirely possible given that we already hear about pollution and pollen levels, why not hear about flu and what the forecast is so that people have a little window into what’s coming down the pike?”
Yes, it’s all becoming clearer now. Maybe one day we’ll even be able to accurately time those hand sanitizer deliveries by Amazon drone.