Epidemiological forecasting is critically needed for decision-making by public health officials, commercial and non-commercial institutions, and the general public. The DELPHI group focuses on developing the technological capability of epi-forecasting, and its role in decision making, both public and private.
Our long term vision is to make epidemiological forecasting as universally accepted and useful as weather forecasting is today. As was the case with weather forecasting, this will likely take several decades. In the shorter term, we select high value epidemiological forecasting targets (currently Influenza and Dengue); create baseline forecasting methods for them; establish metrics for measuring and tracking forecasting accuracy; estimate the limits of forecastability for each target; and identify new sources of data that could be helpful to the forecasting goal.
We produce weekly forecasts of influenza activity in the US in real-time. You can see our current and historical forecasts here!
We are part of a large University of Pittsburgh-based MIDAS National Center of Excellence, which includes epidemiologists, virologists, public health experts, medical doctors specialized in infectious disease, legal and economic experts, and computationalists.
Roni's Proposal for Standardized Evaluation of Epidemiological Models whitepaper
Roni's Predicting The Predictable presentation
Seasonal influenza epidemics cause consistent, considerable, widespread loss annually in terms of economic burden, morbidity, and mortality. Public health officials and medical personnel could prepare more effective countermeasures if they had access to accurate and reliable forecasts about an upcoming epidemic, for example, when designing vaccination programs. Individuals and organizations would also benefit from such predictions when planning for potential sickness.
One goal of the Delphi group is to produce high-quality forecasts of the current season's influenza epidemic, providing not only single-number predictions of key characteristics, but a probability distribution over what the future could be. We leverage data from the CDC's ILINet surveillance system as well as Google Flu Trends to estimate the current prevalence of influenza-like illness, and to build a model of what influenza epidemics can look like. This data is used to generate forecasts with an empirical Bayes framework and importance sampling techniques from statistics. We have participated in the CDC's 2013-2014 and 2014-2015 "Predict the Influenza Season" challenges, submitting weekly forecasts of the season's timing and intensity, both nationally and for 10 regions of the USA. Our 2015 paper in PLoS Computational Biology, Flexible Modeling of Epidemics with an Empirical Bayes Framework, describes our methodology and results. Presently we are preparing forecasts for the 2015-2016 influenza season.
We also seek to understand the interaction between viral evolution and human immunity. Our 2015 paper in PLoS ONE, Computational Characterization of Transient Strain-Transcending Immunity against Influenza A, explores this interaction in depth.
Another goal that we have is to establish a reasonable baseline forecast to which we can compare our various machine-generated forecasts. For influenza and chikungunya, we create this baseline through an ongoing collection of basic forecasts from the general public using a website we created specifically for this purpose. With these user-submitted forecasts we create aggregate weekly forecasts, both as a performance baseline for comparison with data-driven methods and also as a serious attempt to reasonably forecast the spread of these diseases. A manuscript describing the Epicast system is currently under review.
Join in! We are always looking for more participants, so please check out Epicast (http://epicast.org) - and feel free to invite others to do so as well. No special skills or background are required, and the top epicasters are featured on the leaderboards!
Dengue fever is a mosquito-borne viral infection prevalent in the tropical/subtropical regions worldwide, in urban and semi/urban areas. Dengue fever's global incidence has grown dramatically in recent decades, and it is now estimated that about half of the world's population is at risk. There is no specific treatment yet developed for dengue fever, but early detection and medical treatment to early symptoms may lead to a fatality rate of less than 1%. Because of the lack of a vaccine or treatment, vector control (eradication of disease-hosting mosquitoes) is the sole effective prevention measure. As in flu, it is important for public health officials to have access to accurate forecasts, in order to communicate to the public, allocate resources, and implement strategies to combat the spread of the disease.
The Dengue Fever prediction effort of the Delphi group was initially similar to the one for influenza: to produce high-quality forecasts of the entire trajectory or certain facets of the current season's Dengue epidemic, along with measures of uncertainty. Our most recent data was provided through an agreement with the Brazilian Government, spanning over a decade's worth of data of rich spatial and temporal granularity. Together with Willem van Panhuis's group and Brazil's MoH, we produced estimates for game and training cities in the 2014 FIFA World cup in Brazil, where the threat of Dengue has resurfaced in recent years, and where about 600,000 foreign visitors were estimated to visit. Our 2014 paper in PLoS Neglected Tropical Diseases, Risk of Dengue for Tourists and Teams during the World Cup 2014 in Brazil, describes this forecasting effort in greater detail.
We maintain a list of publications relating to epi-forecasting here.