Last year, the city of Lahore, Pakistan, was hit with the worst outbreak of dengue fever in its history. The mosquito-transmitted disease infected some 16,000 people and took 352 lives. This year was a completely different story. There were only 234 confirmed cases and no deaths. The magnitude of the disease varies year to year, but some of the turnaround could be credited to a new system of tracking and predicting outbreaks in the region.
Researchers working for the Pakistani government developed an early epidemic detection system for their region that looked for telltale signs of a serious outbreak in data gathered by government employees searching for dengue larvae and confirmed cases reported from hospitals. If the system’s algorithms spotted an impending outbreak, government employees would then go to the region to clear mosquito breeding grounds and kill larvae. “Getting early epidemic predictions this year helped us to identify outbreaks early,” says Umar Saif, a computer scientist at the Lahore University of Management Sciences, and a recipient of MIT Technology Review’s Innovators Under 35 award in 2011.
“This year, because of the tracking system and the efforts of government employees on the ground, we could look at a map and tell if certain areas were going to develop into an epidemic,” says Saif, who has been working with the government during a sabbatical. “The key is to be able to localize and quarantine a disease like this and prevent it from developing into an epidemic,” he says.
The groundwork for the early detection system was another project headed by Saif: Flubreaks. This system processes data from Google Flu Trends, which estimates the spread of flu based on search terms related to the disease. “That whole idea of being able to scrape digital data has helped us find outbreaks faster,” says Mark Smolinski, director of Global Health Threats at Skoll Global Threats Fund, a nonprofit that recently helped launch a site called Flu Near You, which tracks flu based on a weekly electronic survey that asks people about their health and any flu symptoms.
Smolinski was part of the team at Google to develop Google Flu Trends, which he says can speed up outbreak identification. “You can gain a couple of weeks just by aggregating data of search terms on the Internet,” he says.
While Google Flu Trends identifies outbreaks as they occur, Flubreaks can see them before they start by teasing out global flu trends and making early epidemic predictions.
The results from Flubreaks closely matched actual outbreaks reported by the Centers for Disease Control, says Saif. “We found that idea very exciting,” says Saif. Countries like Pakistan typically do not have a well established disease surveillance network, he says. “We want one for dengue in Pakistan, but it’s a very expensive and difficult thing to manage.”
The researchers have adapted the algorithms designed for early-detection of flu outbreaks for dengue fever outbreaks. “If you are trying to track a disease like dengue, which happens only occasionally, you want a different algorithm compared to a disease that happens every year, such as flu,” says Saif.
The dengue monitoring system relies on real-world field testing of mosquito larvae and reports from hospitals to predict where dengue outbreaks are starting. If a certain neighborhood is suspected to be at the beginning of an outbreak, then government officials could search out mosquito-larvae reservoirs such as pools of water that are likely causing the problem.
The system was put to use this summer. Using 1,500 Android phones, government workers in the region tracked the location and timing of confirmed dengue cases and the mosquito larvae that carry the disease. Each case was tagged by time and location. “Because of the Android phones, we could localize the outbreak to a couple of hundred houses. Inevitably, we would find some water in or near these houses.”
Saif and colleagues plan to verify their dengue epidemic prediction tools using Google Dengue Trends. Like Google Flu Trends, the dengue version scans search terms for disease-related words, but the data set is sparse at this point, says Saif. He and his colleagues are also looking into scanning newspaper articles and social media updates, in collaboration with researchers at New York University, to add to their disease surveillance data.
Digital data such as Internet search terms have potential to help monitor diseases all over the globe, says Smolinksi, just by tracking 15 different symptoms. “If you could detect syndromes of different symptoms in time and space,” he says, it then might be possible “to see where they are coming from and when they are being reported. We might actually pick up outbreaks of respiratory disease, diarrheal disease, and other things.”