An ARIMAX model integrating Baidu index for improved hand, foot, and mouth disease incidence prediction Authors: Zeshan Chen, Yijue Qin Published: 2026 ABSTRACT AbstractHand, foot and mouth disease (HFMD) constitutes a global public health concern. Internet search data offers advantages including vast data volumes, provision of real-time information, and the potential for earlier infectious disease surveillance. The objective of this study is to utilise Baidu Search Index (BSI) to construct a predictive model for HFMD epidemiological surveillance, thereby enhancing HFMD incidence forecasting and early warning capabilities. HFMD cases reported by the National Health Commission of the People’s Republic of China from January 2011 to March 2025 were collected. Keywords highly correlated with HFMD were identified using Spearman’s rank correlation and cross-correlation analysis, and a comprehensive search index (CSI) for HFMD was constructed. Subsequently, based on the monthly number of HFMD cases and the CSI, autoregressive integrated moving average (ARIMA) and autoregressive integrated moving average with exogenous inputs (ARIMAX) models were established. Finally, the predictive accuracy of the models was evaluated using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and standardised mean absolute percentage Error (SMAPE). From January 2011 to March 2025 (total of 171 months), China reported a total of 24,856,589 HFMD cases, with an average of 145,360.2 cases per month. The highest incidence of HFMD occurred between May and July each year. After correlation analysis, five keywords highly associated with HFMD were ultimately included, with a potential time lag of 0 months. The CSI of HFMD exhibits a high Spearman rank correlation (rs = 0.937) with monthly reported HFMD cases. The developed ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) performs better in terms of fitting and prediction compared to the ARIMA(2,0,1)(1,1,1)(12). The MAE values for ARIMAX and ARIMA are 42,085.93 and 80,260.93, respectively, the RMSE values are 52,235.39 and 98,444.62, respectively, the MAPE values are 0.96% and 2.10%, respectively, and the SMAPE are 0.86% and 0.82%, respectively. The ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) model constructed from extensive internet search data in this study has effectively enhanced the prediction of HFMD incidence. Furthermore, it can serve as an early warning system for HFMD. This research provides valuable support for HFMD surveillance.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-51022-y.Keywords: Hand, foot and mouth disease; ARIMAX; ARIMA; Baidu Search Index; InternetSubject terms: Computational biology and bioinformatics, Diseases, Health care, Medical research, Risk factors Introduction Hand, foot and mouth disease (HFMD) is a common childhood infectious disease caused by infection with various enteroviruses, most frequently Coxsackievirus A16 and Enterovirus 711. HFMD primarily manifests as fever accompanied by rashes or vesicles on the hands, feet, and oral mucosa. Some patients may develop severe complications including encephalitis, pulmonary oedema, viral myocarditis, or even fatal outcomes2. Between 2013 and 2022, HFMD ranked as the most prevalent enteric infectious disease in China (incidence rate: 130.40 per 100,000 population). This indicates a substantial HFMD burden in China3. Given the country’s large population and high infection numbers, there is an urgent need to reduce HFMD incidence. Currently, autoregressive integrated moving average (ARIMA) and long short-term memory (LSTM) models are widely employed traditional approaches for forecasting HFMD in China4,5. However, ARIMA exhibits limitations in capturing non-linear relationships within data and in its long-term forecasting capabilities6. LSTM overcomes these constraints by effectively handling extended time series data and demonstrating superior performance in long-term predictions. However, LSTM requires substantial data volumes for training, resulting in high computational costs and considerable complexity7. Consequently, developing new HFMD surveillance models is essential. ARIMA model is one of the most effective linear models for time series forecasting and is frequently employed in predicting infectious disease incidence rates. Recent studies have confirmed the efficacy and applicability of ARIMA models in forecasting hand, foot and mouth disease incidence8. Owing to its straightforward parameter configuration, high predictive accuracy and simple modelling approach, it enjoys broad application and is readily mastered by government officials and researchers. However, it lacks the capacity to incorporate external variables. Autoregressive integrated moving average with exogenous Inputs (ARIMAX), also termed the extended ARIMA model, builds upon ARIMA by integrating the influence of external variables. By incorporating China’s internet search index, ARIMAX better aligns with actual national incidence patterns, further enhancing the predictability of HFMD incidence in China. Zheng et al. developed ARIMA and ARIMAX models, finding that ARIMAX incorporating environmental indicator exogenous factors demonstrated superior performance and reduced error in influenza prediction models9. Additionally, other methodological models can be used for epidemic monitoring, such as the holt-winters model and LSTM10. However, like ARIMA, holt-winters is a traditional forecasting model whose linear modelling approach has numerous limitations, including an inability to handle external variables and weak long-term forecasting capabilities11. In addition, Research has shown that even when incorporating key predictors, the performance of LSTM models remains lower than that of ARIMAX models11. ARIMAX models demonstrate robust performance in forecasting infectious disease incidence12. With the advancement of big data, increasing research employs internet search data for epidemiological studies and infectious disease early warning systems13,14. Compared to traditional infectious disease surveillance systems, internet search data offers superior timeliness. Traditional systems involve a lag of several days or longer from symptom onset, through medical consultation and diagnosis, to final reporting to disease control centres15. However, when individuals experience discomfort, their initial response may involve searching online for symptoms, treatments, hospitals, and relevant departments. Such internet activity often precedes actual medical consultations by days or even weeks, enabling near-real-time dynamic surveillance for epidemic early warning16. Moreover, many patients with mild or atypical symptoms may not seek medical care, instead self-diagnosing through simple internet searches. Traditional monitoring systems may overlook this group, failing to accurately reflect the true scale of infection. Moreover, internet search data can be pinpointed to the province, city, or district of the searcher, enabling early warnings of outbreak hotspots17. Internet data requires minimal human oversight, whereas traditional monitoring systems necessitate multi-tiered reporting and substantial manpower for ‘active surveillance.’ Internet data facilitates ‘passive surveillance’ while reducing monitoring costs18. Baidu is China’s most widely used search engine, with approximately 86.7% of internet users prioritising baidu search19. Its baidu search index (BSI) function reflects the actual level of public concern and demand regarding diseases. Huang et al.20 employed BSI to analyse the correlation between actual HMFD cases and composite indices in urban areas of Guangdong Province, China, yielding a correlation coefficient of 0.8. They utilised the biased sentinel hospital-based area disease estimation (B-SHADE) method combined with historical HMFD actual cases to enhance correlation and reduce error. However, the applicability of this model for HMFD surveillance in other Chinese provinces remains to be verified. Similarly, Chen et al. study only analysed data from two cities in Guangdong Province using BSI, without conducting larger-scale research or validation21. Du et al. employed BSI and climate variables to forecast HMFD incidence in Guangdong Province from 2011 to 2014, yielding an akaike information criterion (AIC) of − 345.332 and a mean absolute percentage error (MAPE) of 101.745% (lower values being preferable)22. However, owing to the limited time span (2011–2014 data only), the model’s performance requires improvement and validation in other provinces22. Several prior studies have demonstrated that the ARIMAX model, comprising ARIMA plus BSI, exhibits significant predictive value in forecasting the incidence and early warning of infectious diseases such as scarlet fever, tuberculosis, and hepatitis C. The inclusion of BSI markedly enhances the model’s performance23–25. This study utilised Baidu Search Index data and monthly HFMD case counts from January 1, 2011 to March 31, 2025 to establish an ARIMAX based predictive model for HFMD in China. This approach aims to enhance the forecasting of HFMD incidence rates and improve early warning capabilities. Materials and methods Data sources The data used in this study were officially released by the National Health Commission of the People’s Republic of China26 (available at http://www.nhc.gov.cn/). These data document in detail the monthly number of newly reported cases of HFMD over a period of 171 months from January 1, 2011 to March 31, 2025. Restricted to aggregate statistics of HFMD cases, the dataset contains no personally identifiable information or sensitive private data of individual patients. In addition, the entire dataset is complete with no missing values whatsoever. BSI data were obtained from the Baidu Index official website27, with data collection spanning from January 1, 2011 to March 31, 2025, and were converted into monthly counts. Keywords selection and filtering Keyword database construction The acquisition of keywords was divided into the following three steps: (1) extracting any keywords related to HFMD from the ChinaZ website28; (2) obtaining HFMD keywords based on the Baidu Search Index demand map29; (3) extracting keywords using Baidu Encyclopedia and Baidu Health Medical Dictionary. We then excluded keywords with low relevance to HFMD based on search volume. Following previous studies23,24, we categorised keywords into five types: general, etiology, examination, symptoms, and prevention and treatment. Keyword screening This study used Microsoft Excel 2019 to develop a statistical analysis database. First, the daily search volume from Baidu Index was converted to a monthly average. We performed a Spearman rank correlation analysis using SPSS 26.0 on the monthly average search volume of keywords and the number of HFMD cases. Keywords meeting the criteria of a rank correlation coefficient (rs) >0.8 and P < 0.05 were retained30. Subsequently, we conducted a cross-correlation analysis on the screened keywords. Those with a maximum cross-correlation coefficient (CCF) > 0.6 were retained, while the remaining keywords were eliminated. Comprehensive search index calculation The comprehensive search index for HFMD is grouped according to the lag of keywords. Then, the Spearman correlation coefficient is used to calculate the weight of each keyword and the comprehensive search index (CSI) for HFMD24,25. The formula is as follows: The formula’s ‘k’ represents the potential time lag (in months), and ‘n’ represents the number of keywords included in each time lag. ρki refers to the Spearman correlation coefficient for a specific keyword (i) at a specific time lag (k). Weightki and keywordki represent the weight and BSI within a specific lag period, respectively. Model construction and evaluation ARIMA and ARIMAX model components ARIMA and ARIMAX models were developed using R4.3.0 software and the forecast package. ARIMA is commonly used for predicting the incidence of infectious diseases31. The expression for the ARIMA model is ARIMA(p, d,q)(P, D,Q)(S), where d and D represent the non-seasonal and seasonal difference orders, respectively. p and P denote the non-seasonal and seasonal autoregressive orders, respectively. q and Q represent the non-seasonal and seasonal moving average orders, respectively; s is the seasonal period. ARIMAX is based on ARIMA, incorporating BSI as an external condition and modelling it jointly with the monthly number of HFMD cases. Time series stationarity test Insertion smoothness test: We plotted a time series chart of monthly reported HFMD cases from January 2011 to March 2025 to determine the existence and characteristics of long-term trends and seasonal fluctuations. Additionally, the augmented dickey-fuller test (ADF test) and Ljung-Box Q test were used to confirm the stationarity of the series. When the series is non-stationary, it can be transformed into a stationary series using stationarisation methods such as differencing and time series decomposition. Dataset splitting We used the 120 months of data from January 1, 2011 to December 31, 2020 as the training set (70.2%) and the 51 months of data from January 1, 2021 to March 31, 2025 as the test set (29.8%). Model construction and evaluation This study determines the optimal parameters of ARIMA and ARIMAX models based on the minimum principle of the akaike information criterion (AIC)9. First, the auto.arima function is used to automatically fit the model parameters. Subsequently, parameters are adjusted by plotting the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots, followed by manual model fitting. The ARIMAX model is manually fitted by incorporating time-delayed CSI into the ARIMA model framework. The ‘least squares method’ is employed for model significance testing. Besides AIC, the criteria for final model selection also encompass residual testing and white noise testing. QQ plots, residual plots, ACF plots, and PACF plots are generated to assess whether model residuals conform to a normal distribution. The Ljung-Box Q test is employed to determine whether model residuals constitute a white noise sequence (P > 0.05 indicates white noise)32. Model performance evaluation We used ARIMA and ARIMAX models to predict the incidence levels of HFMD from January 31, 2021 to March 31, 2025 and compared the predicted values with the actual incidence rates. We further evaluated the predictive accuracy of the models using the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). MAE reflects the average level of absolute error between predicted and actual values. RMSE is particularly sensitive to outliers in the data and effectively captures the magnitude of deviation between predicted and actual values. MAPE is employed to measure the average percentage error between model predictions and actual values. The closer the values of MAE, RMSE, and MAPE are to zero, the better the model’s fit. However, these values increase indefinitely as the model’s error grows. MAPE is unsuitable for predictive models with large expected errors. however, symmetric mean absolute percentage error (SMAPE), accounts for the scale of the data; smaller values indicate predictions closer to actual values. SMAPE constrains the error range between 0% and 200%, thereby correcting this limitation of MAPE33. The workflow of this study is shown in Fig. 1. Flow chart. Results Analysis results of HFMD case numbers and baidu index search data From January 2011 to March 2025 (171 months), China reported a total of 24,856,589 HFMD cases, with an average of 145,360.2 cases per month. The highest incidence of HFMD occurred between May and July each year. We plotted time series graphs of HFMD incidence rates and related keyword search volumes. As shown in Fig. 2A, incidence rates fluctuated, with a decline observed between 2020 and 2023, likely due to the COVID-19 pandemic. The time series graphs of related keyword search volumes and HFMD case numbers exhibited similar trends (Fig. 2B–F). Time series plot of HFMD cases and baidu search index of selected keywords. (A) Time series chart of HFMD cases. (B) Time series chart of HFMD search index. (C), (D), (E), and (F) are time series charts of the keywords ‘medication for HFMD ,’ ‘HFMD in children,’ ‘Hand foot and mouth (HFM),’ and ‘HFM complications,’ respectively. A total of 56 keywords were collected, including 15 general keywords, 24 symptom keywords, 13 prevention and treatment keywords, and 4 etiology keywords. Detailed information is shown in Supplementary document 1. Keyword selection and filtering Correlation analysis between the baidu search index and HFMD The Spearman rank correlation method was used to analyse the correlation between the monthly average search volume of 56 keywords and the number of HFMD cases. We selected keywords with rs > 0.8 and P < 0.05, ultimately including 5 keywords, as shown in Table 1 and Supplementary document 1. Among these, the keywords ‘HFMD in children’ and ‘Medication for HFMD’ exhibited the strongest correlation between their monthly average sequences and the monthly incidence rate sequence, reaching 0.925 and 0.845, respectively. Figures 2C and F illustrate their trends in relation to the HFMD incidence rate series. Table 1Monthly baidu search index for keywords related to HFMD.KeywordAverage ± Standard deviationP value r s a HFMD in children165.07 ± 74.69< 0.010.925Medication for HFMD79.84 ± 42.63< 0.010.845Hand foot and mouth1610.36 ± 984.1< 0.010.842HFM complications24.16 ± 26.66< 0.010.834What medicine to take for HFMD547.43 ± 420.98< 0.010.818Ars, Spearman rank correlation. Monthly baidu search index for keywords related to HFMD. Ars, Spearman rank correlation. Cross-correlation analysis and HFMD comprehensive search index synthesis We then performed a cross-correlation analysis on these five keywords. After excluding keywords with a cross-correlation coefficient (CCF) < 0.6, the results showed that the maximum cross-correlation coefficients of these five keywords were all greater than 0.6 (Table 2). Table 2Correlation between five keywords and the number of new HFMD cases.KeywordsMaximum cross-correlation coefficientLagP valueStandard errorHFMD in children0.9270< 0.010.076Hand foot and mouth0.8420< 0.010.076HFM complications0.8340< 0.010.076What medicine to take for HFMD0.8180< 0.010.076Medication for HFMD0.8450< 0.010.076 Correlation between five keywords and the number of new HFMD cases. After applying the weighting formula, the CSI for HFMD was determined. The results showed that the Spearman rank correlation coefficient between the CSI for HFMD and the monthly reported cases of HFMD was 0.937. Construction of ARIMA and ARIMAX models Automatic model construction Based on the time series of reported HFMD cases and the AIC minimisation principle, the optimal ARIMA model automatically fitted by the auto.arima function is ARIMA(0,0,0)(2,1,0)(12). The least squares method (LSM) value was < 0.01, the residual white noise Ljung-Box test P-value was 0.516, and the AIC value was 2648.77. However, ARIMAX is modelled by combining the ARIMA model with the HFMD composite search index. The optimal ARIMAX model is also determined based on the AIC minimisation principles and the auto.arima function, with an AIC value of 2804.02 (Table 3). Table 3Comparison between the ARIMA and ARIMAX models.ModelLSM, p valueLjung-Box test P valueMAERMSEMAPE (%)SMAPE (%)AICBICAutomatic fitting ARIMA(0,0,0)(2,1,0)(12)< 0.0010.51678388.698444.622.100.772648.772662.18Automatic fitting ARIMAX(0,1,0)errors< 0.0010.60126386.852235.390.960.472804.022809.58Manual fitting ARIMA(2,0,1)(1,1,1)(12)< 0.0010.90480260.9398444.622.100.862647.462663.55Manual fitting ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0)< 0.0010.94642085.9352235.390.960.822539.472558.24LSM, least squares method. MAE, Mean absolute error. RMSE, Root mean square error. MAPE, Mean absolute percentage error. SMAPE, Standardised mean absolute percentage error. AIC, Akaike information criterion. BIC, Bayesian information criterion Comparison between the ARIMA and ARIMAX models. LSM, least squares method. MAE, Mean absolute error. RMSE, Root mean square error. MAPE, Mean absolute percentage error. SMAPE, Standardised mean absolute percentage error. AIC, Akaike information criterion. BIC, Bayesian information criterion Manual model construction The ADF stationarity test indicates that the monthly HFMD incidence series exhibits p = 0.01 and the CSI series p = 0.01, suggesting both time series are relatively stable. Meanwhile, the HFMD incidence trend demonstrates seasonal periodicity. Consequently, the manually fitted model parameters yield d = 0 and D = 1; the HFMD seasonal cycle spans 12 months, with s = 12. The PACF plot (Fig. 3E) indicates truncation at order 2, while the ACF plot (Fig. 3D) exhibits tailing, thus yielding p = 2. To derive the optimal ARIMA model, we varied the parameters q, p, and Q within the range 0–2 for the ARIMA model (p,0,q)(P,1,Q)(12) model. Values were incrementally increased and input into the ARIMA(2,0,1)(1,1,1)(12) (Fig. 3A). The cross-correlation function (CCF) serves to measure the similarity between two time series, particularly the correlation between one series and another at different lags. The CCF indicates (see Supplementary Fig. 1) that the strongest correlation between the CSI series and HFMD incidence occurs when the lag is 0 months. Therefore, we incorporated the CSI series with a lag of 0 months into the ARIMA(2,0,1)(1,1,1)(12) model. This resulted in the final manual ARIMAX model of ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) (Fig. 4A)。The residual analysis indicated that the QQ plots for both models suggested normally distributed residuals (Figs. 3B and 4B). Residual plots fluctuate near zero (Figs. 3C and 4C). Most points on the ACF and PACF plots fall within the 95% confidence interval (Figs. 3D and E and 4D and E). This indicates that residuals from both manually fitted ARIMA and ARIMAX models exhibit white noise characteristics. Both models are suitable for HFMD forecasting. ARIMA(2,0,1)(1,1,1)(12) model fitting and prediction results. (A) is the model fitting and prediction results, where the black line and yellow line represent the true values and predicted values of the training set, respectively, and the blue line and red line represent the predicted values and true values of the test set, respectively. (B) is the QQ plot. (C) is the residual plot. (D) and (E) represent the ACF plot and PACF plot respectively (the blue line represents the confidence interval). ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) model fitting and prediction results. (A) is the model fitting and prediction results, where the black line and yellow line represent the true values and predicted values of the training set, respectively, and the blue line and red line represent the predicted values and true values of the test set, respectively. (B) is the QQ plot. (C) is the residual plot. (D) and (E) represent the ACF plot and PACF plot respectively (the blue line represents the confidence interval). Model selection and performance evaluation Compared with automatic fitting, the manually fitted ARIMA(2,0,1)(1,1,1)(12) and ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) models have lower AIC values, and the AIC of ARIMAX is even lower than that of ARIMA. As shown in Figs. 3A and 4A, the predicted values of the manually fitted ARIMA and ARIMAX models are very close to the actual values, indicating that the models have good accuracy. Additionally, when comparing the predictive performance of the two manually fitted models, the results show that the ARIMAX model performs better than the ARIMA model (Table 3). Discussion This study summarised the epidemiological characteristics of HFMD in China from January 2011 to March 2025 and established an ARIMAX model suitable for HFMD surveillance in China to predict future epidemiological trends. In 2008, China incorporated HFMD into infectious disease management34. Transmission routes for HFMD include faecal-oral transmission, respiratory droplet transmission, and contact transmission35. No specific antiviral treatment exists for HFMD; management typically involves symptomatic and supportive care. Most affected children recover spontaneously, though a minority may develop severe illness. The outbreak of Coronavirus Disease 2019 (COVID-19) in early 2020 led to a marked decline in HFMD incidence during the subsequent three-year pandemic control period. However, following the cessation of China’s COVID-19 dynamic zero-COVID policy, the incidence of HFMD has markedly increased36. Furthermore, the inherent lag in traditional surveillance models and potential underreporting may lead to delayed infectious disease alerts and underestimation of the true epidemic scale. Consequently, novel real-time monitoring methods require exploration to enhance the accuracy of HFMD early warning and forecasting systems. The vast volume of internet data and its provision of real-time information enable earlier warning systems for infectious diseases. As early as 2006, research utilised internet data to monitor influenza outbreaks37. Research indicates that patients may seek online consultations prior to seeking medical care38. Furthermore, due to a lack of specialised knowledge, their online searches for relevant information increase39. In China, Baidu Search Index (BSI) reflects public concern regarding health issues. As early as 2013, studies applied Baidu search queries to epidemiological surveillance in China40. Other research has confirmed that BSI can provide early warnings and predict infectious diseases41,42. Our study ultimately included five keywords related to HFMD. Among these, keywords such as ‘HFMD medication’ and ‘HFMD in children’ exhibit search volume fluctuations that align with changes in HFMD incidence rates, demonstrating consistent amplitude variations. Additionally, Spearman’s rank correlation analysis revealed a Spearman’s rank correlation coefficient of 0.937 between BSI and monthly reported HFMD cases, indicating a strong positive correlation between the two. This also implies that BSI can effectively capture variations in HFMD epidemics. ARIMA serves as a crucial method for monitoring numerous infectious diseases43,44. However, ARIMA is only applicable to single variables and cannot assess the impact of multiple external factors, resulting in suboptimal predictive performance. ARIMAX, conversely, can compensate for this shortcoming of ARIMA. In this study, both the manually fitted ARIMA(2,0,1)(1,1,1)(12) model constructed using HFMD time series and the ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0) model incorporating BSI as an external factor demonstrated good predictive performance. However, the ARIMAX model exhibited superior predictive performance, consistent with the conclusions from studies using ARIMA models for hepatitis C23, pulmonary tuberculosis24, scarlet fever25 and brucellosis45. The MAE, RMSE, and MAPE of the ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0)model were 42,085.93, 52,235.39, and 0.96%, respectively, all of which were smaller than those of the ARIMA ARIMA(2,0,1)(1,1,1)(12) model. This indicates that BSI can serve as an indicator of public concern regarding HFMD while also predicting the epidemiological situation of the disease. Even during the COVID-19 pandemic, the ARIMAX(2,0,1)(1,1,1)(12) + CSI(Lag = 0)model demonstrated good stability in its predictive performance for HFMD. Furthermore, compared with previous studies, the model performance in this study demonstrated superior predictive capability. For example, Guo et al.46 and Yang et al.24 used Baidu Index to develop predictive models with MAPE values of 15.15 ± 0.07% and 5.34%, respectively. Our study demonstrated superior performance with a MAPE value of 0.96%. However, this study also exhibits certain limitations. Firstly, as keywords are susceptible to individual search behaviours, some terms may be omitted, potentially leading to an underestimation of relevance. Therefore, to ensure a high degree of correlation between internet search patterns and actual cases, future research could develop a lexicon that adapts to the dynamic shifts in online search behaviour, thereby enhancing the model’s accuracy. Additionally, prior research has confirmed that environmental factors such as temperature and humidity correlate with HFMD incidence rates, and these conditions may also facilitate disease transmission under certain circumstances47. Furthermore, low internet penetration among elderly populations and in certain regions may reduce relevant search volumes, leading to an underrepresentation of actual search activity and consequently diminishing the surveillance capabilities of infectious disease models48. These factors should therefore be incorporated into future research as key areas for investigation. Finally, this study did not incorporate data from other internet platforms such as Sogou Search, potentially omitting certain keywords. As Sogou serves as the exclusive gateway to WeChat, it can retrieve the majority of articles from public accounts across the entire web. WeChat Official Accounts represent one of China’s most influential platforms for content published by self-media and authoritative sources49. Moreover, Sogou fully integrates content from Zhihu, a knowledge-sharing question and answer community where professionals and medical practitioners may initiate discussions during the early stages of an outbreak. Incorporating such data could substantially enhance the model’s predictive accuracy. Future predictions incorporating additional internet platforms may yield more compelling results. Overcoming these limitations would significantly contribute to further improving the model’s predictive precision. Conclusion In summary, this study utilised extensive internet search data to construct an ARIMAX model, which addresses the shortcomings of China’s traditional HFMD surveillance system in terms of delayed early warning. Furthermore, this model enhances the predictive performance for HFMD incidence rates. These findings will provide valuable assistance and reference for HFMD surveillance in China. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Author contributions ZC wrote the manuscript and prepared the figures and tables. YQ conceptualized the study and oversaw the process. All authors reviewed the manuscript. Data availability The data of HFMD cases analyzed in this study were obtained from National Health Commission of the People’s Republic of China (http://www.nhc.gov.cn/). The search engine data was obtained from Baidu Index (https://index.baidu.com/v2/index.html#/). Declarations Competing interests The authors declare no competing interests. Ethical approval This study is not a clinical trial. Data were obtained from the public National Health Commission of the People’s Republic of China database and the researchers did not have access to individual patient details. The medical research ethics committee of the Guangxi Zhuang Autonomous Region People’s Hospital determined that this research did not require ethics approval according to the following Chinese law: Notice by the National Health Commission, the Ministry of Education, the Ministry of Science and Technology, and the National Administration of Traditional Chinese Medicine of Issuing the Measures for Ethical Review of Life Science and Medical Research Involving Human Being (No. 4 of the National Health Commission, 2023). Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement The data of HFMD cases analyzed in this study were obtained from National Health Commission of the People’s Republic of China (http://www.nhc.gov.cn/). The search engine data was obtained from Baidu Index (https://index.baidu.com/v2/index.html#/).