Analysing Twitter and web queries for flu trend prediction

Area of Science:

• Infodemiology and Public Health Surveillance
• Computational Epidemiology
• Natural Language Processing for Health Data

Background:

• Social media platforms facilitate sharing of personal health information.
• User-generated content can be leveraged for public health monitoring.
• Estimating disease incidence often relies on traditional surveillance methods.

Purpose of the Study:

• To evaluate the use of Twitter messages and search engine query logs for estimating and predicting influenza-like illness (ILI) incidence in Portugal.
• To adapt infodemiology approaches for non-English languages.
• To assess the predictive power of models for subsequent flu seasons.

Main Methods:

• Manual classification of 2704 Portuguese tweets to identify flu-related mentions.
• Training a Naïve Bayes classifier using 650 textual features.
• Developing a multiple linear regression model using tweet classification results and search query logs to estimate Influenzanet data.
• Applying classification and regression models to predict flu incidence.

Main Results:

• A Naïve Bayes classifier achieved a precision of 0.78 and an F-measure of 0.83 for identifying flu-related tweets.
• The regression model achieved a high correlation (0.89) with Influenzanet data.
• Prediction of flu incidence for the following season yielded a correlation of 0.72.

Conclusions:

• Infodemiology methods are adaptable to non-English languages with appropriate data preprocessing and feature selection.
• The proposed models demonstrate effectiveness in estimating and predicting ILI incidence.
• While predictive models show promise for future seasons, adaptation may be needed for optimal robustness.