Columbia University Researchers Develop State-Level Influenza Nowcasting Model
Researchers from Columbia University have developed their own “nowcasting” model, which leverages online search trends to gain a picture of current, local influenza outbreaks.
Flu activity is on the rise in the United States, and, to this end, over 130 million doses of flu vaccine have been administered thus far this season. Monitoring and forecasting influenza and flu-like illnesses can be a critical tool for public health officials. However, while forecasting models continue to become more widely available, they generally don’t have the ability to predict accurate results at the local level.
To remedy that, scientists have developed ways to use data from local health departments, along with surveys of flu symptom mentions in search queries and social media posts, to gain a picture of what’s happening in a given area.
Now, a new study, researchers from Columbia University have leveraged online search trends to gain a picture of current, local influenza outbreaks. Although the study shows signs of promise, researchers remain hamstrung by a lack of access to complete live data.
The most prominent effort at influenza “nowcasting” thus far was Google Flu Trends (GFT), a program first developed by Google in 2008 to surveil internet postings for signs that flu outbreaks were occurring. However, the company shelved that effort in 2013, after it was shown to have badly overestimated instances of the flu.
Picking up the baton, researchers from Columbia University developed their own “nowcasting” model, which used flu-like illness data at the state level, along with web-based search activity from Google Extended Trends. Findings from this new study, published in the Journal of Medical Internet Research, were mixed.
“These results suggest that the proposed methods may be an alternative to the discontinued GFT and that further improvements in the quality of subregional nowcasts may require increased access to more finely resolved surveillance data,” the authors wrote.
In short, the team’s method beat an autoregressive model but underperformed compared to GFT. When the method was altered to include subregional surveillance data, it fared better versus GFT. Due to the proprietary nature of Google’s product, it’s not clear whether GFT performed better due to access to a fuller set of data, or it simply had a better methodology.
The study also found that states with larger populations were easier to accurately assess, likely due to the increase in available data points.
Study co-author Sasikiran Kandula, MS, of Columbia University, says acquiring accurate local data is a major hurdle.
“Starting this season, CDC is releasing data at state level, and this may improve the nowcasts, although we have not yet had a chance to assess,” Dr. Kandula told Healthcare Analytics News.
However, public health agencies are likely the best route to accurate and robust information, since private companies have other interests and may not wish to regularly publish such data, according to Dr. Kandula.
“And if they do make them public, translating this data into reliable estimates of disease incidence is far from straightforward,” Dr. Kandula said. “I doubt they are an alternative to public health surveillance systems.”
As the team from Columbia works to improve its nowcasting modeling, the researchers envision a future where subregional data can be leveraged to help health care organizations better prepare for influenza outbreaks.
“The forecasts may be useful to hospital administrators to allocate beds, manage staffing, [and] supplies; and for local departments of health to distribute PSAs, push vaccination drives or decide on school closings,” Dr. Kandula concluded.