Pitfalls of bigdata to predict “epidemics”: Lessons from Google Flu Trends

Most universities are wholly transfixed on the “promises” of “global health”. To what effect? Most studies focus on communities, but a pilot rarely scales up as a health policy.

Either there’s a lack of political will or the original authors “complete their studies” and leave it in the lurch. As such, it represents a graveyard of good ideas, but that never scaled up.

The following is a fascinating read on the failure of Google’s Flu Trends. Very few remember it (it was laid to rest in 2015) but had promised the big breakthrough in “artificial intelligence”. It aggregated the “flu symptom search” from “worried users”, but it could be anyone.

The author misses one point completely. Most projects in Google start as “betas” only as technology demonstrators. They almost always are “laid to rest” once the “project” is over or internal standards have been met. The period of 2009-2012/3 was one of the feverish pitch to refine “real-time search”. It coincided with the rise of Twitter. Google News was in its infancy. Most companies were trying to define methodologies to figure out what “people are looking for”.

Google Flu Trends was probably part of the experiment. The company is a black hole as far as the status of projects is concerned. For example, they have a division called Verily that makes moonshots on exciting healthcare projects. How are they funded and addressed is a subject of intense speculation.

This article examines some of the conceptual and practical challenges raised by the online algorithmic tracking of disease by focusing on the case of Google Flu Trends (GFT). Launched in 2008, GFT was Google’s flagship syndromic surveillance system, specializing in ‘real-time’ tracking of outbreaks of influenza. GFT mined massive amounts of data about online search behavior to extract patterns and anticipate the future of viral activity. But it did a poor job, and Google shut the system down in 2015. This paper focuses on GFT’s shortcomings, which were particularly severe during flu epidemics, when GFT struggled to make sense of the unexpected surges in the number of search queries. I suggest two reasons for GFT’s difficulties. First, it failed to keep track of the dynamics of contagion, at once biological and digital, as it affected what I call here the ‘googling crowds’. Search behavior during epidemics in part stems from a sort of viral anxiety not easily amenable to algorithmic anticipation, to the extent that the algorithm’s predictive capacity remains dependent on past data and patterns. Second, I suggest that GFT’s troubles were the result of how it collected data and performed what I call ‘epidemic reality’. GFT’s data became severed from the processes Google aimed to track, and the data took on a life of their own: a trackable life, in which there was little flu left.

Here’s another insight:

The story of GFT challenges common narratives in which big data modelling and mining appear as quasi-naturalized operations. It undermines the dichotomy between all-embracing technical ordering and the apparently indomitable complexity of life itself, a dichotomy that often remains implicit in discussions of algorithmic futures. There is no clear-cut distinction, in GFT, between algorithmic formalization and viral activity. 

Needless to say, it only validates my assertion- the purpose of Google Flu Trends was to demonstrate viability of “real-time search”. Twitter held the mantle (as it was a platform designed to have people speaking to each other before bots took over). Hence, they made significant advancements in scale (and database architectures). Twitter is primarily “dead” now (it has been taken over by another set of algorithms and ad networks), but more importantly, the “fad” of real-time searches has finally worn off. Google’s core search algorithms have become better (especially with target slotting and targeted advertising networks with supercookies, etc.).

via MAT