Detecting Influenza Epidemics on Twitter

TL;DR

This paper develops a Twitter-based predictive model for influenza outbreaks by analyzing keyword frequencies, correlating them with official health data, and integrating the predictions into an epidemic phase model, enabling early warnings.

Contribution

It introduces a novel approach combining Twitter data analysis with epidemic modeling to improve early detection of influenza outbreaks.

Findings

Twitter data correlates strongly with official ILI data.

The model provides earlier epidemic warnings than traditional sentinel systems.

Twitter-based predictions are effective for real-time epidemic monitoring.

Abstract

This paper presents a predictive model for Influenza-Like-Illness, based on Twitter traffic. We gather data from Twitter based on a set of keywords used in the Influenza wikipedia page, and perform feature selection over all words used in 3 years worth of tweets, using real ILI data from the Greek CDC. We select a small set of words with high correlation to the ILI score, and train a regression model to predict the ILI score cases from the word features. We deploy this model on a streaming application and feed the resulting time-series to FluHMM, an existing prediction model for the phases of the epidemic. We find that Twitter traffic offers a good source of information and can generate early warnings compared to the existing sentinel protocol using a set of associated physicians all over Greece.