Using multiple data streams and artificial intelligence to ‘nowcast’ local flu outbreaks


Because influenza is so contagious, it’s been challenging to track and forecast flu activity in real time as people move about and travel. While the CDC continuously monitors patient visits for flu-like illness in the U.S., its information can lag by up to two weeks. A new study led by the Computational Health Informatics Program (CHIP) at Boston Children’s Hospital combined multiple approaches, providing what appear to be the most accurate local flu predictions to date.

The ARGONet system, described last week in Nature Communications, combines machine learning with two flu detection models. The first model is the team’s earlier high-performing forecasting approach, ARGO. It leverages information from electronic health records, flu-related Google searches and historical flu activity in a given location.

The second model draws on spatial-temporal patterns of flu spread in neighboring areas. “It exploits the fact that the presence of flu in nearby locations may increase the risk of experiencing a disease outbreak at a given location,” explains Mauricio Santillana, PhD, a CHIP faculty member and the paper’ senior author.

flu prediction
The heatmap at left (click to enlarge) shows state-to-state correlations of influenza-like illness from September 30, 2012 to May 14, 2017. The darker the circle, the stronger the correlation (quantified by the key at right). The black boxes indicate clusters where flu overspread multiple states. At right, five clusters of flu-like illness are shown geographically. (CREDIT: Mohammad Hattab, Wyss Institute for Biologically Inspired Engineering, and Fred Lu, Boston Children’s Hospital)

Putting flu prediction to the test

ARGONet’s final element is the machine learning system. The team “trained” the system by feeding it flu predictions from both models as well as actual flu data. This helped to reduce errors in the predictions.

“The system continuously evaluates the predictive power of each independent method, and recalibrates how this information should be used to produce improved flu estimates,” says Santillana, who is also an assistant professor at Harvard Medical School.

We think our models will become more accurate as more online search volumes are collected and as more healthcare providers incorporate cloud-based electronic health records.

Santillana and his colleagues applied both ARGO and ARGONet to flu seasons from September 2014 to May 2017. ARGO alone outperformed Google Flu Trends, the previous forecasting system that operated from 2008 to 2015. But ARGONet made even more accurate predictions in more than 75 percent of the states studied. It was able to predict flu activity across the U.S., at the state level, a week ahead of traditional healthcare-based reports.

“We think our models will become more accurate over time as more online search volumes are collected and as more healthcare providers incorporate cloud-based electronic health records,” says Fred Lu, a CHIP investigator and first author on the paper.

Precision public health

The investigators believe their “nowcasting” approach will help local health officials mitigate epidemic outbreaks. The timely information may help the public be more aware of potential risks, says Santillana.

The work was funded by the Centers for Disease Control and Prevention and the National Institute of General Medical Sciences of the NIH. Mohammed Hattab of the Wyss Institute for Biologically Inspired Engineering in Boston, Cesar Leonardo Clemente of Tecnológico de Monterrey (Monterrey, Mexico) and Matthew Biggerstaff of the CDC were coauthors on the paper.