A Deep Learning Framework for Using Search Engine Data to Predict Influenza-Like Illness and Distinguish Epidemic and Nonepidemic Seasons: Multifeature Time Series Analysis.

Background: The seasonal influenza epidemic poses a persistent and severe threat to global public health. Web-based search data are recognized as a valuable source for forecasting influenza or other respiratory tract infection epidemics. Current influenza prediction studies typically focus on seasonal trends in traditional monitoring data, neglecting the sensitivity of different web-based search terms to seasonal changes, thereby increasing prediction challenges. Objective: The aim of this study was to propose a deep learning framework for different influenza epidemic states based on Baidu index and percentage of influenza-like illness (ILI%). Methods: Official weekly ILI% data from 2013 to 2024 were extracted from the Chinese National Notifiable Infectious Disease Reporting System (NIDRIS). Based on the Baidu index, influenza-related search indexes were acquired for the corresponding time periods. To explore the association between influenza-related search queries and ILI%, the study conducted a cross-correlation analysis. The study period was divided into influenza epidemic and nonepidemic period. The study finally used the convolutional long short-term memory (CLSTM) network framework to predict influenza epidemics with 1-3 weeks ahead for the all-time period and epidemic + nonepidemic period. The evaluation metrics included model stability metric, accuracy metrics, and explanatory power metric. Results: The ILI% presented a regular seasonal high incidence in China. Meanwhile, the prediction of ILI% after dividing the epidemic and nonepidemic seasons (mean absolute percentage error [MAPE]=10.730%, mean square error [MSE]=0.884, mean absolute error [MAE]=0.649, root-mean-square error [RMSE]=0.940, and R2=0.877) was better than that of the all-time period (MAPE=12.784%, MSE=1.513, MAE=0.744, RMSE=1.230, and R2=0.786). In addition, we found that the ILI% + Baidu search index predicts better than only the ILI% regardless of the time period and lag time of the study. Comparative analysis with long short-term memory (LSTM) and transformer models demonstrated that CLSTM achieved superior performance in 1 week-ahead ILI% predictions using ILI% + Baidu index data in epidemic + nonepidemic period (MAPE=11.824%, MSE=1.243, MAE=0.723, RMSE=1.115, and R2=0.827). Furthermore, CLSTM comprehensively surpasses LSTM in computational efficiency, complexity, extrapolation capability, and stability while partially outperforming transformer models. Conclusions: This study shows strong potential for influenza prediction by combining Baidu index data with traditional surveillance and specific keywords for epidemic and nonepidemic seasons. It provides a new perspective for public health preparedness. This research is expected to support early warning systems for influenza and other diseases. Future work will further optimize these models for more timely and accurate predictions, enhancing public health responses.