Monitoring virus-laden mosquitoes is nothing new, but forecasting their next move through predictive modeling is.
This story was originally published by Data-Smart City Solutions.
Yes, it is Bill Gates, known to many as the richest man in the world, who compared the threat of mosquito-borne illnesses to Shark Week on the Discovery Channel. Gates has a point: mosquitoes are the transmitters of a long list of illnesses from malaria to Zika that pose a threat to people worldwide, yet many view them only as pests, not as a public health threat.
Chief among these mosquito-borne threats in the United States has been West Nile Virus (WNV). In 1999, WNV’s first reported cases appeared in Queens, NY; by 2001, it was detected in the Chicago area after an increase in bird deaths was linked to the virus. In 2002, Chicago suffered its first, and most severe, WNV epidemic: 225 human cases were reported, including 22 fatalities.
Despite the reach of the disease, not all mosquitoes carry West Nile Virus, mitigating the risk of transmission. In Chicago, for example, the most abundant species are only “nuisance” mosquitoes, which are not vectors (carriers) of the disease. For those who do contract WNV, the risk that it will advance to more serious symptoms is very low. Most people infected with the virus will not get sick; some (around 20 percent) may experience flulike symptoms. According to the US Centers for Disease Control (CDC), less than 1 percent of people bitten by an infected mosquito will become severely ill. Still, the number of WNV cases in an area could rise in any given year. The state of Texas, for example, had an unexpected outbreak in 2012: after only 27 cases statewide in 2011, the state was hit with 1,868 the following year. Across the United States, nearly 2,000 US residents have died from WNV complications since 1999. That number would likely be higher if not for surveillance measures to monitor human and non-human cases of infection. This surveillance allows local governments to take action (mainly in the form of mosquito spraying) to prevent the spread of WNV.
As a changing climate allows mosquito population to increase, these control efforts will become more difficult. Towards the end of this century, Chicago is projected to have plant-growing conditions equal to present-day Alabama. As the warmer climate shifts north, the potential for mosquito-friendly conditions comes with it.
This risk is why the Chicago Department for Public Health (CDPH) established a surveillance and control program in 2004, which includes annual spraying. It’s also why the city’s Department of Innovation and Technology (DoIT) has teamed up with CDPH to build a predictive model that can determine one week in advance whether or not a particular area will have WNV-carrying mosquitoes. With such a model, the city can target mosquito-spraying efforts towards areas that pose the greatest risk, mitigating the potential of another outbreak.
And this year, following a successful pilot program, the City of Chicago is now operationalizing WNV prediction. The hallmark launch provides lessons for current and future mosquito control efforts.
The concept for the WNV analytics project came via Kaggle, an online platform for analytics and predictive modeling competitions. Kaggle lets organizations post data online so that data scientists can compete to produce the best predictive model. In early 2015, CDPH, in partnership with DoIT, launched a Kaggle competition of its own by posting some WNV visualizations and starter code in R and Python on Kaggle Scripts.
The competition, which offered $40,000 in prizes for the three best models, was sponsored by the Robert Wood Johnson Foundation, which has become an integral partner for health-related data and analytics efforts in Chicago. In connection with the competition, the city published relevant data on its open data portal, including mosquito spraying program locations, schedules, and historical records of WNV detection at trap sites. The city made additional public data, such as local weather information from NOAA, easily available for the challenge as well. With data now ready, a large number of competitors set out on their quest: to predict where and when different species of mosquitoes will test positive for West Nile virus.
Armed with such predictions, CDPH could have the ability to more efficiently and effectively allocate resources towards preventing residents’ exposure and transmission of the virus. Yet getting those predictions right is not easy: any given modelling task allows for numerous strategies—and from the outset, it’s difficult to know which strategy or method is most effective. Kaggle’s crowdsourced approach was therefore a great starting point for CDPH’s venture into WNV analytics.
On June 17, 2015, the competition closed, bringing Chicago submissions from more than 1300 competitors. Thanks to these entries, Chicago’s team better understood their options for moving forward with WNV predictive model development. Perhaps most importantly, the competition made WNV analytics possible in the first place, as it provided an impetus for the city to get all its mosquito-related data ready for analysis.
Because of the data published for the competition, Chicago also connected with Nicolas Defelice, a researcher at Columbia University’s Department of Environmental Health Sciences, who had been conducting extensive research on forecasting WNV cases in urban areas around the country. Defelice had been working on a comprehensive research paper of his work, and was soon in regular contact with the Chicago team.
The paper, Ensemble forecast of human West Nile virus cases and mosquito infection rates, was published in Nature Publications in February 2017 and proved to be an important reference point for Chicago’s WNV model development process.
Gene Leynes, a senior Data Scientist at Chicago’s Department of Innovation Technology (DoIT), has worked on a wide range of analytics projects throughout his career. Yet becoming an expert on mosquito control is not something he ever predicted he would do.
“The main goal here for our analytics effort is to put systems in place that reduce people’s exposure to mosquitoes that carry WNV—or any vector-borne illness, really.” Leynes noted. WNV mosquito sprayings have been going on in the Chicago area for well over a decade. In early spring of every year, CDPH staff meets with numerous city departments to assess the season’s efforts and review levels of WNV risk. The city then moves forward with plans to spray insecticide in potentially affected areas. To inform the process, CDPH conducts mosquito and bird surveillance operations, which includes regularly trapping and collecting mosquitoes for WNV analysis.
Trapping and spraying isn’t Chicago’s only defense against WNV, either: other efforts include larviciding catch basins, which involves dropping tablets in storm drains along the public way that slowly dissolve over a five-month period to prevent mosquito larvae from hatching, and eliminating standing water by ensuring that swimming pools and construction sites are regularly maintained.
In Chicago’s spraying program—its main effort against WNV—the city sprays areas when CDPH mosquito traps test positive for WNV two weeks in a row. This is where Leynes and the efforts of Chicago’s advanced analytics team come in.
“Having a host of WNV-related data ready for analysis presents us with an opportunity,” noted Leynes. “Given this week’s results, can we predict if there will be WNV at a given trap location next week? Building an algorithm that can detect at-risk sites earlier means that we can spray sooner, reducing public exposure to the virus.”
Using data from 2010 through 2016, Chicago built a model that relies on the presence of WNV cases in the past, as well as weather data and trap locations, to predict whether a trap will have WNV present on a given trap collection week. By the summer of 2017, the model was ready to be tested.
The key output from the WNV algorithm is a score indicating the risk that a specific site could test positive for WNV in an upcoming week. To use this score, Chicago needed to be sure the model provided reliable, accurate results. During the summer of 2017, DoIT’s data science team tested the model to measure how frequently alerts for potential presence of WNV would actually occur. The team compared predictions for WNV to whether it actually occurred in the following week, finding that its model was able to correctly predict consecutive results for WNV virus about 80 percent of the time.
Armed with a working algorithm, Chicago’s next step was to integrate it into current operations. The city’s goal was to get the algorithm up and running without changing or disrupting current processes (as was the case with its food inspection forecasting and predictive rodent baiting programs).
Fortunately, the city has had a long-running tool to help make the program easy to implement: WindyGrid. WindyGrid is a computer application designed for city personnel that makes Chicago’s big data easily and strategically accessible in one place. As a geographic information system, the application presents a unified view of city operations across a map of Chicago, giving key personnel access to all of the city’s spatial data. This includes both historic and real-time data, with featured information including 911 and 311 service calls, transit and mobile asset locations, building information, and even public geospatially-enabled tweets.
DoIT has since added WNV prediction data into WindyGrid. With all city employees able to access WindyGrid, CDPH’s field staff has an already-running tool to use in its spraying and monitoring processes, making the transition to algorithm-optimized operations seamless.
“The number of West Nile-related illnesses in Chicago remains low for the last couple of years, but we're maintaining our vigilance. There are nearly 100 traps across the city where we continue to monitor for mosquitoes infected by West Nile virus, which can be transmitted to humans,” said Dr. Julie Morita, Commissioner of CDPH, in a recent press release. “Sometimes, those infected by West Nile virus, especially children and adults with weakened immune systems, can experience unpleasant flu-like symptoms or even neurological damage. This new tool helps us prepare and proactively direct our resources to the locations that need it most.”
Chicago’s efforts to combat WNV aren’t related to only one ailment, of course; its process can be applied to any mosquito-borne illness, including the more threatening Zika virus. As Jonathan Jay has noted for Data-Smart Cities, Aedes aegpyti, the mosquito responsible for most cases of transmission of the Zika virus, can be found in warm, wet climates throughout the American South, as far west as California, and as far up the Eastern Seaboard as New York City.
For Chicago, this means that the seasonal buzzes of Aedes aegptyi are not far off. In preparation for this, the City of Chicago’s Department of Public Health has launched a public awareness campaign to better understand Zika’s risks. Should Zika continue its spread northward, the city’s WNV modeling program will be able to provide immediate value.
#StopZika, CDPH’s public awareness resource for the Zika Virus, includes general information, tips for travelers, and a response line for any Zika-related issues or questions. Source: Chicago Department of Public Health.
Furthermore, Chicago’s WNV algorithm can assist cities affected by mosquito-borne illness in the present—whether it be WNV, Zika, or something else. Chicago’s model, as well as instructions on how to run it, is free for any city or locality to replicate via the city’s GitHub page.
For Leynes, the work he’s done on WNV prediction—along with the efforts of those across the board, from Kaggle entrants to CDPH—is well worth it.
“Climate change is already affecting tick and mosquito populations, and it’s only going to get worse. If my work can help save even one life, then it’s all worth it for me.”