Our study used a total of 3 time series data about Yunnan province, including the incidence of mumps, the Baidu index [27], and environmental factors.

We obtained the daily number of mumps cases with their age, sex, and careers, and the annual population data from the Yunnan Center for Disease Control and Prevention, which collected accurate case characteristics from the Chinese information system for infectious diseases control and prevention, for the period between January 1, 2014, and December 31, 2023. From the system, we collected a total of 20 career types in our study, reporting the number of patients of the top 6 most common ones while grouping the remaining careers under the category “other careers.” In addition, since more than 80% of patients were younger than 20 years, we organized the age distribution into 5-year groups for those younger than 20 years, while others were presented as a separate category. The Baidu index [27] is the weighted sum of the search frequencies of various search terms by scientifically analyzing and calculating based on users’ search volumes on Baidu, which is the most popular search engine in China. By extensive search, we initially collected daily Baidu index for 30 search terms in Yunnan from January 1, 2016, and December 31, 2023. After group discussion, we excluded 3 search terms due to irrelevance and another 3 because they were introduced after 2016, resulting in incomplete data (Table S1 in Multimedia Appendix 1). Ultimately, we included 24 search terms in our study.

We collected information about environmental factors during January 1, 2016, and December 31, 2023, from China Meteorological Administration [28], including maximum temperature (°C), minimum temperature (°C), air quality index, particulate matter (PM) with an aerodynamic diameter less than or equal to 2.5 μm (PM_2.5; measured as μg/m³), PM with an aerodynamic diameter less than or equal to 10 μm (PM₁₀; measured as μg/m³), sulfur dioxide (SO₂; measured as μg/m³), nitrogen dioxide (NO₂; measured as μg/m³), carbon monoxide (CO; measured as mg/m³), and ozone (O₃; measured as μg/m³).

Our research was approved by the Medical Ethics Committee, Xiangya School of Public Health, Central South University, Changsha, China (XYGW-2024-65). Informed consent was waived by our institutional review board because all data were fully anonymized before being accessed by the researchers, ensuring that no identifiable information about individuals was available. Therefore, compensation is not applicable in our study.

In Yunnan Province, the incidence of mumps increased from 10.9 per 100,000 population in 2014, peaked at 37.5 per 100,000 population in 2019, and then steadily decreased to 12.8 per 100,000 population by 2023 (Figure 2). The outbreaks of mumps occurred from 2017 to 2020 in different cities in Yunnan. From 2014 to 2023, the proportion of female patients remained fairly stable, ranging from 41.3% in 2015 to 45.7% in 2020. (Table 1). The median age of patients showed some fluctuations over the years, starting at 7 years (IQR 4-12) in 2014, peaking at 11 years (IQR 7-14) in 2019, and then decreasing to 6 years (IQR 4-10) by 2023. The proportion of cases in children between 5 and 10 years old remained substantial, with the highest proportion recorded in 2023 (43.7%). Students consistently represented the largest affected group, comprising 44.3% to 67.5% of cases in 2014 and 2019, respectively.

The temporal distribution from 2016 to 2023 revealed distinct trends in both the Baidu search index and environmental factors (Figure 3). The distribution of the 3 Baidu index mirrored the pattern observed in mumps incidence, with a peak in 2019 and followed by a decline. Most environmental factors except SO₂ displayed obvious seasonal cycles. Maximum and minimum temperatures reached peaks in summer, while air pollutants, including CO, PM_2.5, and PM₁₀, had higher concentrations in winter. In addition, SO₂ and CO showed a long-term decreasing trend over the 10-year period.

The Pearson correlation analysis reveals that 3 Baidu search terms, disease name (r=0.67), symptom (r=0.66), and treatment (r=0.53), show strong positive correlations with mumps incidence in Yunnan Province from 2016 to 2023 (Table 2). Among environmental factors, both minimum (r=0.17) and maximum temperature (r=0.16) have weak positive correlations, while pollutants including SO₂ (r=–0.16), CO (r=–0.12), PM_2.5 (r=–0.20) and PM₁₀ (r=–0.10) show weak negative correlations with mumps incidence.

Considering the potential lag impact of these predictors on mumps incidence, we leveraged DLNM to explore their short-term impacts (Figure 4). The impact of various predictors on the incidence of mumps varies across different lag weeks. For 3 Baidu search terms, the RR consistently increases with the rising Baidu index. The analysis also revealed nonlinear associations between temperatures and incidence, with minimum temperature showing a U-shaped curve and maximum temperature exhibiting a complex pattern. In addition, elevated levels of SO₂, CO, PM_2.5, and PM₁₀ generally corresponded with decreases in RR at a 1-week lag.

Table 3 shows the comparisons of the performance metrics of 4 models with different combinations of predictors. Model IBE, which includes all 3 predictors, consistently outperformed the other models in both the train and validation set, as well as the test set. Specifically, model IBE achieved the highest R² values (0.89 in the train and validation set, 0.72 in the test set) and the lowest error metrics, with an MAE of 0.51, MAPE of 15.2%, and RMSE of 0.74 in the train and validation set, and an MAE of 0.33, MAPE of 15.9%, and RMSE of 0.43 in the test set. In comparison, the IE model (R²=0.54, MAE=0.45, MAPE=28.2%, and RMSE=0.55) and the IB model (R²=0.49, MAE=0.45, MAPE=29.2%, and RMSE=0.58) showed similar levels of prediction accuracy in the test set, both outperforming model BE, which had the worst performance (R²=0.33, MAE=0.54, MAPE=32.0%, and RMSE=0.66). The actual values and model predictions of the weekly incidence of mumps are presented in Figure 5.

Our study presented the incidence of mumps from 2014 to 2023 in Yunnan, a representative province of irregular mumps activity, and found continuous mumps outbreaks in different cities from 2017 to 2020. Using multisource data, we constructed models for predicting the incidence of mumps one week in advance based on LSTM. Model IBE (ie, the model including the incidence of mumps, Baidu index, and environmental factors) showed the best performance among 4 models.

The mean incidence of mumps in Yunnan from 2014 to 2023 was 16.9 per 100,000 persons, which is lower than that of the period between 2004 to 2012 (26.2/100,000) in Yunnan and comparable to the incidence in mainland China from 2013 to 2018 (16.9/100,000) [19]. The 10-year incidence pattern of mumps in Yunnan shows an increase followed by a decrease starting in 2020. The decline may be related to the COVID-19 pandemic, which extensively reduced opportunities for interpersonal contact, thereby blocking the spread of infectious diseases. However, we do not have a definite explanation for the peak in 2019. Given the rapidly increasing number of cases among students in 2019, factors such as vaccine coverage, waning immunity, and the close gathering of adolescents on campuses may have contributed to the outbreak [3,33].

Previous studies have found nonlinear associations between meteorological factors and the incidence of several infectious diseases including mumps [34,35], influenza [23,24,36], and chickenpox [37]. These findings suggest meteorological factors are common predictors for the field of infectious disease prediction. Our study found 6 environmental factors with a PCC of >0.10 and presented their lag correlations with the incidence of mumps. The most relevant factors were PM_2.5 and SO₂, both showing weak negative correlations (PCC=–0.2) with mumps incidence in Yunnan. Notably, the PCC is obviously lower than that of other infectious disease. For example, a study in Singapore found a stronger correlation (PCC=0.32) between minimum temperature and the COVID-19 pandemic incidence [38]. This variability suggests that while environmental factors can be valuable for predicting some infectious diseases, they may not be sufficient for accurate prediction across all diseases, highlighting the need for additional predictors.

In recent years, using search engine data, such as Google Trends [39], and the Baidu index [24,40], to predict the incidence of infectious disease has become popular and demonstrated excellent predictive performance. The advantages of search engine data included timeliness, low cost, and sometimes stronger correlation with disease incidence than other data sources. Given the different disease outcomes, fairly comparing our study with other research using search engine data is challenging, as this is, to our knowledge, the first study to construct a mumps prediction model using search engine data. However, we noted that Yang et al [24] and Choi and Ahn [39], who predicted influenza incidence in mainland China and South Korea respectively, achieved similar or even better performance using Baidu index and Google Trends alone, compared with our study’s combination of the Baidu index and environmental factors. It suggests that the contribution of search engine data to incidence prediction may vary depending on types of disease, regions, and data sources. In our study, the PCCs of the Baidu index of 3 search term groups were higher than those of environmental factors. This may be because the Baidu index, derived from individuals’ active search activities, better reflected the incidence peak between 2018 to 2020. In contrast, the relatively stable environmental factors might contribute more to the seasonal trends in incidence but failed to capture the sharp fluctuations observed during outbreaks.

Our study deployed multisource data to construct and compare 4 models, with the model IBE demonstrating superior performance. Notably, when Baidu index data was incorporated, model IBE exhibited even greater predictive accuracy compared with the IE model. This demonstrates the capability of the Baidu index to predict mumps incidence and the potential of integrating multiple data sources to complement the surveillance systems in predicting mumps more accurately, as these data sets can complement one another and provide a more comprehensive understanding of the influencing factors and early signs of disease spread. Further research could build on the proposed framework in our study or explore the integration of additional data sources, such as social media activity and movement behaviors, to further enhance the predictive models.

To our knowledge, our study is the first to develop a prediction model for the incidence of mumps that combines the Baidu index and environmental factors as predictors, and it is also the first to create such a model specifically for Yunnan province, China. In addition, by leveraging data such as the Baidu index, our model offers a novel and practical approach to monitor mumps incidence 1 week in advance.

However, there are several limitations to our study. First, at the city level rather than the provincial level, we observed different outbreak times among cities from 2017 to 2020, which may indicate the spread trajectory of mumps and be important for its control and prevention. Unfortunately, the time-series forecast model in our study cannot leverage this geographical information. Future studies could explore integrating temporal and spatial data, which may be crucial for preventing widespread transmission. Second, some Baidu search terms were introduced after 2016 (eg, the search term “Symptoms and treatment of mump” was introduced on April 2, 2018), thereby decreasing the number of available predictors of the Baidu index, which may compromise the performance of Baidu index. Nonetheless, after extensive research, our study identified and included a sufficient number of predictors. Third, we did not have access to some meteorological factors such as sunshine duration and wind speed, which have been identified as effective predictors in other studies, potentially hindering better performance of our models. Future research should incorporate a broader range of meteorological variables to improve prediction accuracy further. Fourth, applying the model to practical public health interventions still faced several challenges. Future work should focus on developing a user-friendly tool that can automatically collect predictors and output results, making the model more accessible and actionable for public health authorities. Finally, due to economic and geographic variations in China [19], various epidemic patterns of infectious diseases, and the different nature of data sets from other search platforms, this model may not generalize well across different regions, diseases, or data sources without careful adaptation and validation. While the model in this study was applied to mumps, similar approaches could potentially be applied to other infectious diseases with seasonal or outbreak-driven incidence patterns. Diseases such as influenza, chickenpox, and hand, foot, and mouth disease, where public awareness, mobility, and environmental factors play a role in transmission, might benefit from similar predictive frameworks. Future studies should focus on tailoring models to specific diseases and data sets to evaluate their broader applicability.

In conclusion, our study showed the incidence of mumps in Yunnan province, China, from 2014 to 2023, highlighting a sharp peak around 2019. We used data from 2016 to 2023, including mumps incidence, the Baidu index, and environmental factors, to construct 4 prediction models based on LSTM networks. Among these models, the model IBE demonstrates the best performance. This indicates the feasibility of using search engine data and environmental factors to forecast the incidence of mumps, offering a valuable tool for public health surveillance and rapid outbreak assessment.