Big data diagnostics.

Abstract

A little over a year ago, a damning report published in the journal Nature (1) sent a ripple through the public health and social science communities: in a season of unusually high flu outbreak, one of the world's newest biosurveillance systems was inaccurately predicting flu trends. Google Flu Trends (GFT),3 launched in 2008, is an attempt by Google to predict, in real time, the prevalence of influenza-like infection (2). In combination with computer modeling, the company mines data records for flu-related keyword searches, producing a map of on-the-ground flu activity. GFT predicts trends at the local and state levels, and nationally in at least 29 countries around the world (3). “The GFT effort is a prominent one in the field,” says Dr. David Lazer, a professor of political science and computer and information science at Northeastern University, in an e-mail to Clinical Chemistry . As his professional title would indicate, Lazer is no stranger to big data and has been a proponent for big data analysis in public health. The Nature report highlights that in the peak of the 2012–2013 US flu outbreak, GFT's national “peak of flu” was almost double that of the CDC. (The report also points out that in other countries, GFT was hitting projected targets.) Lazer, an early supporter of GFT, admits he had not closely examined the product. “It then was widely reported that it had missed by a huge margin last winter, which begged the question: why?” he says. This past March, Lazer and a team of colleagues from Northeastern and Harvard universities published their answer to that question in the journal Science (4). In looking at GFT trends, Lazer and colleagues write that since well before 2013, “GFT had been persistently overestimating flu prevalence.” For example, the team found …