Diagnosing the Outbreak and Spread of the Chikungunya Virus in Central & South America - The DARPA CHIKV Challenge
Abstract:
In 2014, the U.S. Defense Advanced Research Projects Agency (DARPA) issued a “challenge” to forecast the future outbreaks and spreads of the virus Chikungunya Virus (CHIKV) in Central and South America. We took on the challenge of predicting the cumulative number of Pacific Atlantic Health Organization (PAHO) reported Chikungunya or CHIKV cases by South and Central America PAHO country, and by week. Our predicted variables include cumulative numbers of suspected, confirmed and imported cases. To generate these predictions, we developed a methodology that combines statistical and data-driven empirical epidemiological modeling approaches. The results of our model system, the sources of information and a predictive algorithm capability are documented and provided. At the end of the CHIKV Challenge, DARPA announced ten Top 10 Teams, out of 487 that entered the challenge. Our Team was one of the Top Ten Awardees. The challenge period of the CHIKV Challenge only lasted from late August of 2014 until March 2015, just seven months, and thus we believe that the results of many of the DARPA Challenge predictive model systems based on Southern Hemisphere spring to fall conditions were skewed accordingly. Basically, many of the 487 models were simple power law curve fits, including some who finished in the Top 10. Had the context continued for five more months, these curve fitting exercises would have failed as the Southern Hemisphere moved from its fall to winter, and the power law curves would have failed to decrease due to planetary seasonal progressions. Our data based mathematical model approach included related weather factors and seasonal intrinsic modes of variability, which would have continued for more seasons and even years. Those other approaches were mathematically correct but epidemiologically implausible. Unfortunately, the appraiser that DARPA employed to evaluate the mathematical prognostic methodologies was not up to the task. Our approach is both mathematically and epidemiologically generic, and will continue to produce accurate forecasts well into the future in the South and Central Americas. As such, we were honored by DARPA to have been in the Top 10 of the contestants. The cash award was divided up amongst several graduate students. This manuscript documents the successful mathematical methodology that we employed.
