Influenza Surveillance using Search Engine, SNS, On-line Shopping, Q&A Service and Past Flu Patients

TL;DR

This study evaluates multiple online data sources and regression models to improve influenza outbreak prediction accuracy in Japan, highlighting the effectiveness of Huber Regression with combined data resources.

Contribution

It introduces a comprehensive analysis of various UGC types and models for influenza prediction, identifying the best model and resource combination for accurate forecasting.

Findings

Huber Regression with multiple data sources yields the highest accuracy.

Search query logs and social media are strong predictors over three years.

Huber Regression effectively handles outliers in flu prediction data.

Abstract

Influenza, an infectious disease, causes many deaths worldwide. Predicting influenza victims during epidemics is an important task for clinical, hospital, and community outbreak preparation. On-line user-generated contents (UGC), primarily in the form of social media posts or search query logs, are generally used for prediction for reaction to sudden and unusual outbreaks. However, most studies rely only on the UGC as their resource and do not use various UGCs. Our study aims to solve these questions about Influenza prediction: Which model is the best? What combination of multiple UGCs works well? What is the nature of each UGC? We adapt some models, LASSO Regression, Huber Regression, Support Vector Machine regression with Linear kernel (SVR) and Random Forest, to test the influenza volume prediction in Japan during 2015 - 2018. For that, we use on-line five data resources: (1) past flu patients, (2) SNS (Twitter), (3) search engines (Yahoo! Japan), (4) shopping services (Yahoo! Shopping), and (5) Q&A services (Yahoo! Chiebukuro) as resources of each model. We then validate respective resources contributions using the best model, Huber Regression, with all resources except one resource. Finally, we use Bayesian change point method for ascertaining whether the trend of time series on any resources is reflected in the trend of flu patient count or not. Our experiments show Huber Regression model based on various data resources produces the most accurate results. Then, from the change point analysis, we get the result that search query logs and social media posts for three years represents these resources as a good predictor. Conclusions: We show that Huber Regression based on various data resources is strong for outliers and is suitable for the flu prediction. Additionally, we indicate the characteristics of each resource for the flu prediction.