Flu view—tests, predictions for the upcoming season

Create PDF

Predicting flu intensity

Dr. Jeffrey Shaman’s multi-model ensemble for flu forecasting was informed by data assimilation methods similar to those used in weather prediction. His models relied on recent flu activity as reported by three sources: real-time, municipal-scale data from Google Flu Trends; data from the CDC’s Influenza-like Illness Surveillance Program; and regional- and national-scale influenza positivity rates from the CDC and World Health Organization. The team analyzed multiple ensemble forecasts each week to evaluate and adjust the model’s reliability in real time.

Dr. Shaman

While the CDC’s “Predict the Influenza Season Challenge” focused on predicting flu intensity at a national level, the Shaman laboratory has been modeling flu intensity at the municipal level for a few years. Flu predictions for 100 cities across the nation are published on the Shaman laboratory’s website (http://cpid.iri.columbia.edu/), along with forecasts for other infectious diseases such as Ebola. In 2012, their most accurate year to date, the lab’s models accurately predicted a peak in flu activity in 63 percent of the cities at least two weeks before the actual peaks occurred.

“We can also assign certainties to our forecasts. So we don’t just say that flu is predicted to peak in five weeks in New York City; we can say there’s a 70 percent chance that flu will peak in five weeks. And that’s very different than if we said there’s a 10 percent chance that flu will peak in five weeks,” says Dr. Shaman, noting that the former prediction carries a much higher probability of accuracy because the ensemble of forecasts are in greater agreement.

Dr. Shaman cautions that infectious disease forecasting is just one tool in the public health arsenal. “It’s not intended to do more than it is laid out to do, which is to provide a view of the influenza incidence that may be coming down the pike for various communities. It is used in complement with surveillance, and it relies on that surveillance. People have to start becoming comfortable with what it means, and how you might respond to a 70 percent chance of flu peaking in five weeks, versus a 10 percent chance of flu peaking in five weeks, versus a 70 percent chance of flu peaking in three weeks. Those are very different pieces of information.” We have to learn how to use them, Dr. Shaman says, and how to incorporate them into decisions about public health responses and preparedness activities.

Forecasting efforts still have quite a way to go before the models are ready for mainstream use, says Dr. Daniel B. Jernigan of the CDC. “But as one of the participants in the forecast challenge said, 60 years ago weather forecasting was not very good either. And two things improved that. One was the technology and the analytics and all of the modeling and so forth. The second…was increasing the number of places that were actually collecting the data.”

In the meantime, flu forecasting models can enhance situational awareness.

“We currently put out information through FluView and other mechanisms,” Dr. Jernigan says. “There are apps you can download to tell you when things are trending upwards, and whether the circulating viruses are H3 or B, for example. We’re certainly continuing with this kind of information, but we’re seeing that we can actually improve our situational awareness through alternative data sources like Google and Wikipedia that can provide granular findings at the community level. That’s an area where we would like to continue working with our partners.”

As for Dr. Shaman’s predictions about the coming flu season, his influenza modeling efforts aren’t expected to kick into high gear until November. Until then, his team continues to explore multi-ensemble forecasting, including different model types and data assimilation methods, and different combinations of data that could enhance forecasts.

“We’re also looking at some other areas of the world, and forecasting, where some of the dynamics of flu may be a little bit different,” he says. –Ann Griswold

Ann Griswold is a writer in San Francisco.