New Flu 'Nowcasting' Model Outperforms Previous Methods

Investigators say they’ve developed a more accurate model for tracking and predicting the spread of influenza.

Mauricio Santillana, PhD

Scientists have developed a new framework for influenza “nowcasting,” which they say outperforms existing flu tracking models because it combines 2 distinct but valuable data streams.

Investigators from Boston Children’s Hospital, Harvard University, Tecnológico de Monterrey, Mexico, and the US Centers for Disease Control and Prevention, say their “ensemble” model can give highly accurate indications of current flu conditions, and also predict the spread of the virus with fewer errors than previously proposed methods of flu tracking.

The ensemble contains 2 components. The first is the Auto-Regression with General Online information (ARGO) model, which combines location-specific historic flu data with flu-related Google search data and electronic health record data to track the flu.

However, the data scientists behind the new study sought to improve upon ARGO. In order to do so, they combined ARGO with a second model that is based on spatial-temporal patterns of the flu. The research team describes the second model, abbreviated as “Net” as a check that stops ARGO from major errors. The result is highly accurate “nowcasting” at the state level.

“Using the values of past CDC influenza reports in an autoregression adds robustness by preventing our models from creating outsize errors in prediction,” wrote senior author Mauricio Santillana, PhD, and colleagues. “Similarly, incorporating spatial synchronicities adds stability by maintaining state-level inter-correlations evident in historical influenza activity.”

ARGO alone was already an improvement over its most high-profile predecessor, Google Flu Trends. When Santillana and colleagues tested ARGO and Google Flu Trends against historical flu data from 2012 to 2017, ARGO came out on top. (Google Flu Trends was shuttered in 2015.)

Next, the investigators compared ARGO with the combined ARGONet model, which adds the layer of spatial-temporal patterns. When ARGO and ARGONet were tested against retrospective flu estimates from 2014 to 2017, ARGONet was the winner in more than three quarters of the states studied. Importantly, these estimates can be made one week ahead of when traditional state level flu data reports would come out.

That early warning can make a big difference.

“Timely and reliable methodologies for tracking influenza activity across locations can help public health officials mitigate epidemic outbreaks and may improve communication with the public to raise awareness of potential risks," Santillana said, in a press release.

With better tracking and prediction, local public health authorities will be able to provide the right interventions at the right time, the co-authors wrote.

Moreover, because the models are based on a stockpile of data that will only get larger over time, they should become even more accurate in the future.

“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," said Boston Children’s Hospital’s Fred Lu, BS, the study’s first author, in the press release. Both Santillana and Lu are on the faculty of the hospital’s Computational Health Informatics Program (CHIP).

The study, “Improved state-level influenza nowcasting in the United States leveraging Internet-based data and network approaches,” was published in Nature Communications.