Just as weather forecasting has improved over recent decades, the accuracy of forecasting influenza and other infectious diseases is expected to improve, says Jeffrey Shaman, PhD, associate professor of environmental health sciences at the Mailman School of Public Health, who led a team that placed first in the Centers for Disease Control and Prevention’s “Predict the Influenza Season Challenge.”
Dr. Shaman’s team developed a scientifically validated system for predicting seasonal peaks of influenza in cities across the United States. “A number of government agencies are beginning to recognize the value of disease forecast and understand this is a research area in which we should invest.”
Contest entrants were asked to forecast the timing, peak, and intensity of the 2013–14 flu season using digital data. Eleven teams completed the challenge, using a variety of data sources. First-place recognition and a $75,000 prize were awarded to Dr. Shaman’s team, which included Wan Yang, PhD, a postdoctoral researcher in environmental health sciences at Mailman. Other team members were from the National Center for Atmospheric Research and Harvard School of Public Health.
Funding for the flu forecasting research was provided by a joint program of the National Institute of General Medical Sciences and the National Science Foundation to support research at the interface of the biological and mathematical sciences; the Biomedical Advanced Research and Development Authority of the Department of Health and Human Services; and the Models of Infectious Disease Agent Study of the National Institute of General Medical Sciences.
As part of National Flu Vaccination Week, the CUMC Newsroom spoke with Dr. Shaman about his model.
How does your flu prediction model work?
We mimic the prediction framework used to produce weather forecasts. For this we rely on three basic ingredients: 1) a mathematical model describing the propagation of influenza through a local population; 2) real-time observations of the system we are forecasting (e.g., estimates of influenza incidence for the modeled population); and 3) a class of statistical methods called data assimilation or sequential Monte Carlo methods. The issue is that if we run our mathematical model alone, it will typically make a poor forecast. We improve this forecast by iteratively informing the model of the state of observed conditions. This process, which optimizes the model to better replicate local flu activity as it has thus far transpired is performed using the second and third ingredients.
Specifically, we initiate an ensemble of simulations 10 to 20 weeks in the past. With each weekly observation of influenza incidence, we stop these integrations and use the data assimilation methods to adjust the simulations to better align with observations. We then integrate forward to the next observation and repeat all the way up to the present. Through this recursive adjustment process, the model is optimized to represent the outbreak of flu as it has thus far unfolded. The idea is that if the ensemble of simulations replicate what has already transpired, it will be better positioned to predict what will happen. The forecast is then generated by integrating the optimized model ensemble from the present into the future.
How is this year’s flu season shaping up as of early December?
It is still early. We are seeing some activity, particularly in Texas and Louisiana.
When do you expect your weekly forecasts to begin?
We started posting forecasts last week at our website, Columbia Prediction of Infectious Diseases.
How did winning the CDC’s “Predict the Influenza Season” contest affect your work?
The recognition is wonderful. But in truth, we need to keep developing and testing these systems for a range of diseases and applications. I think it has motivated my group to explore these larger possibilities.
How will Ebola cases affect the public’s attention to flu season?
It may make people generally more aware of infectious diseases and the need to take personal precautions. However, it will be important to recognize the distinctions between Ebola and influenza, particularly the risk of infection from each (virtually nil for Ebola, but substantial for influenza). People are accustomed to influenza and its associated risks; as a result, we are often complacent about getting vaccinated, staying home when sick, or covering our face when sneezing. Perhaps Ebola will make us all a little more proactive in working to stop infectious disease in general and flu in particular.
What are your plans for improving on your model?
We are always working to improve our forecasts. We test and develop new strategies, new models, new data types, and new data-assimilation methods. This kind of investment is needed. The forecast error of numerical weather predictions over the past 30 years has steadily dropped, i.e., forecast accuracy has improved, as a result of continued research and development. If we invest similarly in infectious disease prediction, we will begin to realize its full potential and benefit.
Infectious Disease Forecasts:cpid.iri.columbia.edu
This story was originally published by Columbia University Medical Center on Dec. 8, 2014.
