July 21, 2015
Reich attended a previous workshop in the White House series focused on forecasting dengue fever. His lab has been working with the Thai government on creating real-time forecasts of dengue fever outbreaks there—one of only a handful of such efforts.
“We have merged over 2.5 million records of dengue cases dating back 50 years in one central database. But there is a lot more work to be done to understand this complex system enough to be able to make actionable predictions,” notes Reich.
Approximately 50 attendees at the DARPA workshop discussed how current forecasting methodologies fail, ways to evaluate the usefulness of forecasting models, what degree of accuracy is needed before a forecast can be actionable, and key scenarios for disease forecasting. According to Reich, participants came up with several key conclusions. First, different types of forecasts may be needed for different audiences. For example, vaccine producers need to know a prediction of what strains of influenza may be circulating months in advance, but the general public might be more interested in very short-term local forecasts of when peak flu season will take place. Second, forecasting influenza is a huge challenge. Although accurate forecasting itself is difficult, Reich notes biostatisticians also face problems communicating uncertainty about a given prediction.
“A 50% chance of an outbreak is essentially where our uncertainty is highest. You might as well be saying ‘flip a coin and predict from that’. But if we say there is a 10% chance or a 90% chance of some event happening, those show equal amounts of certainty about the event happening or not,” he points out.
Despite these challenges, Reich is optimistic about the future of disease forecasting.
“This is a really exciting time to be a part of these efforts. We are thinking about ways to leverage multiple real-time streams of data - weather data, disease surveillance data, travel data - to improve public health.”