The ILI% is the proportion of patients with ILI divided by the total number of physician visits. The ILI% based on weekly reports in Langfang, Hebei Province, China, was extracted from the Chinese National Notifiable Infectious Disease Reporting System (NIDRIS) for influenza from October 2013 to March 2024. Patients with ILI were defined as outpatients of any age with acute respiratory infection syndrome with fever ≥38 °C and cough or sore throat.

Baidu index is a statistical indicator that represents the search volume of demanding keywords or phrases based on Baidu’s search query logs, which is the largest search engine in China [23]. This study summarized influenza-related keywords that might correlate with the trend of ILI% and selected the keywords and the time range from October 2013 to March 2024. The search data were based on personal computer and personal mobile phone data.

A descriptive analysis was used to show the characteristics of the current ILI% and the current search indices of different keywords on the Baidu index. This study divided the study period based on the characteristics of influenza epidemics. According to the latest influenza surveillance plan [24-26], the epidemic season is from week 40 of each year to week 13 of the following year, while the nonepidemic season is from week 14 to week 39. This division helps align with the seasonal patterns of influenza outbreaks [27]. The cross-correlation coefficient was calculated to explore the association between the influenza-related search terms data and the ILI% of the epidemic season and the nonepidemic season, respectively. The study also analyzed the correlation between the previous week’s search index (from week 1 to 4) and the ILI% of the epidemic season and the nonepidemic season. A correlation coefficient closer to 1 or –1 indicates a stronger correlation, and a correlation coefficient closer to 0 indicates a weaker correlation. The cross-correlation coefficient was calculated between each variable to observe the correlations among the variables. After performing the correlation analysis, the variables with correlation coefficients above 0.5 were selected. These variables were used in different lags to develop predicting models. This approach aimed to improve prediction accuracy based on previous studies [11,28].

To effectively leverage the useful information in surveillance data and Baidu index data, this study proposed a deep learning framework for multifeature time series to address the challenge of predicting real-world influenza trends (Figure 1). By fully mining the inherent characteristics of data and establishing the mapping relationship between features and results, the framework could complete the task of time series prediction in a scientific and robust manner. The mathematical formulation of the framework is shown in Multimedia Appendix 1.

The models used in this study were based on an encoder-decoder framework. The input of the encoder was the data, including historical ILI% and Baidu keyword searches. The advanced ILI% forecast results were obtained from the output of the decoder. For observed ILI%, the variational mode decomposition algorithm [29] was applied to decompose original signal into multiple intrinsic mode functions. This helped remove redundant information from the observed ILI%, thereby improving prediction accuracy [30]. A time window was set and slid during each decomposition to generate new subsignals while preventing information leakage. For the Baidu index, transformer was used for the feature extraction in this study. The method used a self-attentive mechanism to capture temporal dependencies and semantic relationships in the sequences. This allowed it to extract key features that are useful for predicting ILI%. In this way, feature extraction could more accurately reflect the link between search data and influenza activity.

This study used the classical encoder-decoder framework for algorithm construction. The encoder part was divided into backbone and fusion parts. Data from different modalities were input into the encoder through 2 separate channels. The backbone part, which is a convolutional long short-term memory (CLSTM) network, was used to learn the corresponding features of each modality {M₁, M₂}. Gated attention fusion model was used in the fusion part. Gated attention fusion integrated features {M₁, M₂} from the backbone. It generated a query, key, and value using the attention mechanism and computes attention weights. Then, these weights were processed by softmax, used to weight the value. Subsequently, the weighted values were used to enable the model to selectively focus on key features through the gating mechanism, resulting in the joint feature, M_y. The joint feature M_y was used as input data and sent into the decoder. In the decoder, the CLSTM module was used to decode the state vector, while the multiple-layer perceptron transformed the features generated by the CLSTM into the final ILI% prediction. In the model validation, the study introduced long short-term memory (LSTM) and transformer as baseline comparators, ensuring fair evaluation through identical training sets and assessment metrics. To further verify generalizability, the model was validated using independent datasets from Tianjin, China.

The ILI% surveillance data were divided into 2 parts, the training set and the test set. The training set was from the 14th week of 2013 to the 10th week of 2022, and the test set was from the 11th week of 2022 to the 13th week of 2024 for the whole study period. The training and test sets for the epidemic and nonepidemic seasons were divided according to their respective study times in a ratio of 8:2 and assigned in time order. This study used 3 types of indicators to evaluate the performance of the prediction model, including model stability metric, accuracy metrics, and explanatory power metric (Multimedia Appendix 2). Stability metric mainly uses the mean absolute percentage error (MAPE). Accuracy metrics include mean square error (MSE), root-mean-square error (RMSE), and mean absolute error (MAE). Explanatory power metric mainly uses the coefficient of determination (R²). When R² exceeds 0.7, while MSE, RMSE, and MAE values approach 0 and MAPE falls below 20%, it indicates strong predictive performance of the model [31]. In addition to model performance, the models’ computational efficiency and computational complexity analysis were used to the models’ comparison. To evaluate computational efficiency, floating-point operations per second (FLOPs) can be used because it directly reflects the amount of computational resources and processing speed required by the model when performing computational tasks. Similarly, parameters indicate the model’s complexity and memory footprint, which are crucial for practical deployment. To evaluate computational complexity analysis, the Big-O complexity estimates were used to assess scalability. Moreover, hyperparameter sensitivity analysis was conducted to investigate how variations in model parameters (the number of attention heads, the kernel size, and the number of layers) affect predictive performance. Shapley Additive Explanations (SHAP) analysis was used to interpret feature contributions in the CLSTM model. Finally, the convergence analysis was used to demonstrate training stability through comparative learning curves across multiple influenza seasons.

A box plot was introduced to illustrate the distribution of prediction errors. All data analyses were performed using SPSS Statistics software (version 23; IBM Corporation) and Python (version 3.4.0; Python Software Foundation). All data were checked for completeness and accuracy before analysis.

The study used deidentified surveillance data from the Chinese NIDRIS in China. The study received formal ethics approval from the research ethics board of Tianjin University (approval number TJUE-2025-222), and the requirement for informed consent was waived. The study did not involve direct human participants research, and no compensation was provided to participants.

The ILI% presented a regular seasonal high incidence in China. The average weekly ILI% was 4.36%, respectively. The highest ILI% was in the 50th week in 2022 (21.03%) and the lowest was in the 40th week in 2016 (0.86%). The influenza season is clearly reflected in the period from October to March each year, and there is only 1 influenza peak in the whole influenza cycle (Figure 2). ILI% has a trend of increasing annually (Figure 3). The observed ILI% showed a clear periodic pattern, with the ILI% being higher than the yearly average for the period from October to March (Figure 3).

Based on the literature searched on the web, the Baidu search terms in this study that were screened out were all related to influenza. Overall, the study observed a high correlation between the Baidu search terms and the ILI% in all databases. The terms were classified into 4 categories: influenza essential facts, influenza symptoms, influenza treatment, and influenza prevention (Table 1 and Multimedia Appendix 3). The correlation coefficients differed between the influenza epidemic and nonepidemic seasons.

When analyzing the use of keywords in different time periods, it was found that the focus on keyword categories differed between the influenza epidemic and nonepidemic seasons (Table 1 and Multimedia Appendix 3). Specifically, in the influenza epidemic season, the significant keywords with a cross-correlation coefficient greater than 0.5 mainly concentrated on the category of influenza essential facts. These keywords usually focused on basic information about influenza-related diseases, transmission routes, and other basic information. In contrast, during the nonepidemic season, the most important keywords in the framework mainly focused on the category of influenza treatment. It covered the use of medications, treatment protocols, and other content. This suggests that there are remarkable differences in the public’s concerns during different time periods. In addition, this study observed that the categories of keywords throughout the year mostly overlapped with those used during the epidemic season.

To compare the training stability of models across epidemic and nonepidemic periods, this study presents the loss functions of 3 models (CLSTM, transformer, and LSTM) (Figure 8A and B). Comparative analysis revealed consistent performance rankings: transformer achieved optimal convergence, LSTM exhibited the weakest performance, and CLSTM maintained robust intermediate performance. CLSTM demonstrated remarkable stability despite seasonal variations in data distribution, highlighting superior robustness and cross-distribution adaptability. All-time period convergence analysis (Multimedia Appendix 7) yielded consistent results with seasonal evaluations.

To further validate the generalizability of the CLSTM model, the study conducted additional experiments using ILI% from Tianjin and corresponding Baidu index data. The model demonstrated robust performance for 1-week-ahead ILI% predictions using ILI% + Baidu index data in epidemic + nonepidemic period (Table 6), achieving an R² of 0.755 >0.7, MAPE of 17.832% <20%, and low error metrics (RMSE=1.716, MSE=2.943, and MAE=1.216). These results indicate the model’s strong predictive capability on external datasets. Comparative analysis revealed the CLSTM model’s superior performance over LSTM and transformer. It confirmed the model’s enhanced ability to capture spatiotemporal patterns in ILI transmission.

Although current models of transmission epidemiology largely rely on surveillance data and mathematical models, there have been many studies exploring the potential of alternative data sources and machine learning techniques. This study integrates Baidu index data with traditional surveillance data to propose a deep learning framework that uses multifeatured time series for different influenza epidemic states. By considering the impacts of both epidemic and nonepidemic seasons, the framework enhances the accuracy of influenza predictions. Analysis revealed that the weekly search volumes for 4 types of Baidu search terms related to influenza—essential facts, symptoms, treatment, and prevention—were strongly correlated with ILI%. By distinguishing between epidemic and nonepidemic seasons, this study found that these search terms had different focuses during each period. Based on these findings, we developed a CLSTM deep learning framework using multifeatured time series from multiple data sources. The study was conducted over 4 research periods: all time periods, the combined epidemic and nonepidemic periods, the epidemic period, and the nonepidemic period. The framework can predict influenza trends up to 2 weeks in advance. This research investigates the application of combining public opinion data with traditional surveillance systems to predict influenza epidemics. It provides a valuable reference not only for northern China but also for advancements in modern surveillance methods.

Dividing the study period into epidemic and nonepidemic seasons can more accurately capture the seasonal characteristics of the public’s search behavior for influenza information (Baidu index). The ILI% was used to predict trends in influenza virus incidence [32]. An increase in the ILI% usually signals the peak of an influenza pandemic, and health care providers and public health departments can take timely action based on the change in the ILI% to intensify their influenza prevention and treatment efforts. Nevertheless, influenza risks vary during different seasons. Therefore, dividing the study period into epidemic and nonepidemic seasons better captures seasonal variations in public concern and search behavior, which are crucial for predicting influenza trends. During the influenza epidemic season, members of the public, with heightened crisis awareness and perceived increased risk of contracting influenza, proactively obtain relevant information for self-assessment and prevention. Specifically, they pay attention to the symptoms of influenza for self-diagnosis, learn about transmission routes to prevent infection, and seek preventive measures to protect themselves. These behaviors led to a significant increase in Baidu search terms related to influenza essential facts and prevention, which in turn led to an increase in the Baidu index. In addition, extensive media coverage of influenza cases and outbreak developments also increased public attention. Public health departments usually promote preventive measures such as vaccination and hand hygiene during the epidemic season, which further directs the public to search for related information. In contrast, during the nonepidemic season, the public is relatively less aware of the crisis and less worried about the risk of contracting influenza. In this case, they mainly focus on information related to influenza treatment. Therefore, Baidu search terms focus more on recognizing influenza symptoms, choosing treatments, and effective medications. Media reports and other health information received relatively less attention, while public interest in vaccination and preventive measures declined. Overall, public attention during the nonepidemic season shifted more toward the influenza treatment.

The inclusion of Baidu index is effective for ILI% prediction. Using separate Baidu search terms for epidemic and nonepidemic seasons will improve prediction performance. The search engine data, as a leading signal of influenza trends, can sense the spreading trend of influenza earlier than the ILI% [27]. Baidu index is a globally applicable big data tool, widely used in Baidu-dominated regions (eg, parts of Asia and South America). By the second quarter of 2025, Baidu will have a wide range of active users globally. Baidu operates similarly to Google Trends and other international search indices, making the model adaptable to local platforms (eg, Naver in Korea) without algorithmic modifications. Currently, Baidu index has been used in deep learning–based prediction of influenza on several occasions [31,33,34]. Search terms filtered from Baidu index are often divided into 4 categories [31,34,35]: basic understanding of influenza, symptoms, treatment and medicine, and prevention. It is generally consistent with this study’s categorization of influenza-related terms.

At the same time, it is necessary to study influenza prediction using deep learning algorithms. Yang et al [34] used data from northern and southern China and integrated a model based on gated recurrent units and multiple attention mechanisms to achieve approximately 2 weeks’ prediction, with an R² value exceeding 0.77. Jung et al [36] used a self-attention mechanism-based deep learning model to predict influenza trends, demonstrating significant effectiveness. Additionally, studies [31,33] combined ILI case, virological surveillance, climate, demographic, and search engine data. They applied LSTM models to predict ILI%, achieving R² values from 0.67 to 0.9. Athanasiou et al [35] used weekly ILI monitoring data and multisource data, including Twitter and weather conditions, to predict ILI% in Greece with an LSTM model. In this study, using only Baidu index and traditional monitoring data, the CLSTM framework predicted ILI% in 1-2 weeks. The framework’s R² value exceeded 0.82, outperforming most studies using multisource data. Comparative analysis demonstrates that CLSTM achieves better performance to classical machine learning models in terms of computational efficiency, complexity, extrapolation capability, and stability.

It was found that the current papers rarely achieve more than 3 weeks of accurate prediction, which is deemed ineffective [31,35]. The Baidu index reflects public search behavior for influenza-related information. It typically increases 1-2 weeks before an outbreak and weakens afterward, reducing predictive effectiveness. Traditional ILI% surveillance data are more reliable for longer-term predictions and provide comprehensive information.

Most influenza prediction studies do not divide epidemic and nonepidemic seasons. This study found that using specific Baidu search terms for each period significantly improves prediction performance and reliability during single epidemic or nonepidemic seasons. This approach allows the framework to capture season-specific patterns more accurately, providing reliable and detailed prediction results.

However, this study had a few limitations. First, this study did not separately consider the effect of the COVID-19 pandemic on ILI% but allowed the framework to learn from existing trends. Second, computational challenges (eg, real-time latency) must be addressed through optimization and cloud solutions.

To enhance robustness, future work should explore the following: first, hybrid artificial intelligence strategies (eg, reinforcement learning) for dynamic parameter tuning; second, multisource data integration (social media, electronic health records, and climate variables) to improve sensitivity and reduce confounders; and third, federated learning for privacy-preserving cross-region collaboration, despite computational inequities.

In summary, this study demonstrates significant potential for influenza prediction by integrating Baidu index with traditional surveillance data and specific keywords (for both epidemic and nonepidemic seasons). The CLSTM framework performed exceptionally well, effectively improving prediction accuracy through the use of seasonal Baidu search terms, thereby offering a novel perspective for public health early warning systems. Although Baidu’s search engine is widely used internationally, its data are subject to demographic, geographical, and media-driven biases, necessitating population-weighting adjustments and integration with traditional surveillance data to mitigate prediction distortions. Future efforts will focus on further optimizing the model to achieve more timely and accurate predictions: on one hand, leveraging transfer learning to extend this framework to other disease predictions, and on the other hand, using multiobjective optimization to balance accuracy and interpretability. It will ultimately enhance public health response capabilities.