Mediaboss Marketing

Epidemiologist want to forecast disease like meteorologists forecast rain. And the way people browse Wikipedia could be the key, they say.

This time last year, the Centers for Disease Control and Prevention in Atlanta launched a competition to find the best way to forecast the characteristics of the 2013-2014 influenza season using data gathered from the internet. Today, Kyle Hickmann from Los Alamos National Laboratories in New Mexico and a few pals reveal the results of their model which used real-time data from Wikipedia to forecast the ground truth data gathered by the CDC that surfaces about two weeks later.

They say their model has the potential to transform flu forecasting from a black art to a modern science as well-founded as weather forecasting.

Flu takes between 3,000 and 49,000 lives each year in the U.S. so an accurate forecast can have a significant impact on the way society prepares for the epidemic. The current method of monitoring flu outbreaks is somewhat antiquated. It relies on a voluntary system in which public health officials report the percentage of patients they see each week with influenza-like illnesses. This is defined as the percentage of people with a temperature higher than 100 degrees, a cough and no other explanation other than flu.

These numbers give a sense of the incidence of flu at any instant but the accuracy is clearly limited. They do not, for example, account for people with flu who do not seek treatment or people with flu-like symptoms who seek treatment but do not have flu.

There is another significant problem. The network that reports this data is relatively slow. It takes about two weeks for the numbers to filter through the system so the data is always weeks old.

Thats why the CDC is interested in finding new ways to monitor the spread of flu in real time. Google, in particular, has used the number of searches for flu and flu-like symptoms to forecast flu in various parts of the world. That approach has had considerable success but also some puzzling failures. One problem, however, is that Google does not make its data freely available and this lack of transparency is a potential source of trouble for this kind of research.

So Hickmann and co have turned to Wikipedia. Their idea is that the variation in numbers of people accessing articles about flu is an indicator of the spread of the disease. And since Wikipedia makes this data freely available to any interested party, it is an entirely transparent source that is likely to be available for the foreseeable future.

Hickman and co use the flu-article data from earlier years to train a machine learning algorithm to spot the link with the influenza-like illness figures collected by the CDC. They then used the algorithm to predict flu levels in real time during last years flu season.

Read this article:
How Wikipedia Data Is Revolutionising Flu Forecasting