Robust two-stage influenza prediction model considering regular and irregular trends

Influenza causes numerous deaths worldwide every year. Predicting the number of influenza patients is an important task for medical institutions. Two types of data regarding influenza-like illnesses (ILIs) are often used for flu prediction: (1) historical data and (2) user generated content (UGC) data on the web such as search queries and tweets. Historical data have an advantage against the normal state but show disadvantages against irregular phenomena. In contrast, UGC data are advantageous for irregular phenomena. So far, no effective model providing the benefits of both types of data has been devised. This study proposes a novel model, designated the two-stage model, which combines both historical and UGC data. The basic idea is, first, basic regular trends are estimated using the historical data-based model, and then, irregular trends are predicted by the UGC data-based model. Our approach is practically useful because we can train models separately. Thus, if a UGC provider changes the service, our model could produce better performance because the first part of the model is still stable. Experiments on the US and Japan datasets demonstrated the basic feasibility of the proposed approach. In the dropout (pseudo-noise) test that assumes a UGC service would change, the proposed method also showed robustness against outliers. The proposed model is suitable for prediction of seasonal flu.