Predicting Dengue Fever Outbreaks with Machine Learning

CAMBRIDGE, MA—Dengue infection is the most prevalent mosquito-borne viral disease throughout the world—and there is no available treatment. Each year an estimated 500,000 people are hospitalized with severe dengue, also known as dengue hemorrhagic fever, and about 2.5 percent, or 12,000, die. And the problem is only growing worse. Over the past 50 years, dengue cases have increased thirtyfold, and currently half of the world’s population, some 3.9 billion people, are at risk.

Rainfall patterns drive dengue transmission as the Aedes mosquito—a main vector of the disease—needs standing water to reproduce. To better understand Aedes breeding patterns, Draper scientists, working in collaboration with Boston University and the Massachusetts Institute of Technology, developed the Predictive fLUshing Mosquito (PLUM) model. The PLUM model may help public health officials in areas at risk to predict and decrease outbreaks.

Previous models assumed that increased rainfall in warm climates leads to more breeding opportunities for mosquitoes, and subsequently a higher incidence of dengue fever. However, the PLUM model has found that too much rainfall can disrupt the Aedes reproductive cycle by flushing out aquatic larval stages from these breeding sites.

“Rainfall conditions have never been studied for their potential effect on dengue transmission,” said Natasha Markuzon, formerly principal scientist at Draper. “Our research suggests that predictive and early warning systems for dengue outbreaks need to consider flushing, among other contributing factors, to accurately predict near-term risk.”

PLUM uses machine learning and regression models to identify rainfall conditions associated with flushing, based on earlier entomological observations collected in Singapore, where dengue cases were observed to actually decrease after a very wet monsoon. The team successfully extended the model using weather data and dengue cases in Peru and Puerto Rico.

Flushing events, characterized by the interaction of cumulative and intense daily rainfall, were associated with a 15 to 85 percent decrease in the risk of dengue incidence. The researchers observed an association between flushing and reduced dengue outbreak occurrence for two to five weeks after a flushing event.

Mosquitoes are one of the deadliest animals in the world, according to the World Health Organization. With more than half of the world’s population living in areas where mosquitoes are present, health officials, aid organizations and biopharmaceutical companies have long searched for a better way to accurately forecast the spread of mosquito-borne diseases.

“With the PLUM model, we can predict how big an outbreak will be in the coming weeks,” Markuzon said. “PLUM can give us near-term, near real-time prediction that can be extended to different parts of the world. PLUM can help to optimize vector control strategies and better understand dengue, particularly as the expansion of the disease is expected to continue due to climate change, globalization, travel, trade, socioeconomics, resettlement and viral evolution.”

The current research into dengue fever, just published in PLOS Neglected Tropical Diseases and funded by the Defense Threat Reduction Agency (DTRA), is among the most recent innovations in Draper’s image and data analytics portfolio. Other machine learning and data mining tools developed at Draper include applications for finding online extremists in social networks, a deep learning technique for predicting cognitive decay in Alzheimer’s disease and a model that uses information derived from Internet-based news sources to detect and analyze the spread of disease epidemics.

Troy Lau, group leader in Machine Intelligence at Draper, said, “Machine learning and data analytics continues to show its applicability across a wide of uses. At Draper we are applying these tools to cyber defense, human-machine teaming, ocean sensing, biomedical technologies, finance and a number of other applications.”