Methods and Applications: Construction of Prediction Model of Foodborne Disease Outbreaks and Its Trend Prediction — Guizhou Province, China, 2023–2025

This study utilized data from the “Guizhou Foodborne Disease Outbreak Surveillance System” at the Guizhou Center for Disease Control and Prevention. Meteorological data from the World Weather Information Service (http://worldweather.wmo.int/zh/home.html) and China Weather Network (http://www.weather.com.cn) were collected for each city and state in Guizhou Province from 2012 to 2022. Five monthly average weather indicators were analyzed: average monthly rainfall, average monthly rainfall days, average monthly relative humidity, monthly relative humidity, and average hours of sunshine.

The data was organized using Excel 2019 (Microsoft, Redmond, WA, US), and SPSS 24.0 (IBM, Armonk, NY, US) was utilized to perform the χ² test for the reported prevalence of foodborne disease outbreaks. The statistical analysis considered differences to be significant at P<0.05. The rate of foodborne disease outbreaks was the dependent variable, while climate factors were the independent variables. Each climate factor was categorized into five groups based on its magnitude. The χ² test was employed to assess the statistical significance of differences in disease rates among the various factor groups.

Due to the seasonal patterns of foodborne disease outbreaks and their correlation with climatic conditions, this research utilized the seasonal autoregressive integrated moving average (SARIMA) model. The study also applied the autoregressive integrated moving average with exogenous regressors (ARIMAX) model and a combination of SARIMA-ARIMAX models (7). The Prophet model served as a reference for comparison. Given the complexity of the analysis process with fewer random variables (8), a substantial sample size spanning from 2012 to 2022 was necessary for reliable results. The time series prediction model was developed using R 4.2.2 (R Core Team, Vienna, Austria), with parameters selected for model fitting assessment (9).

For the investigation into foodborne disease outbreaks under varied climatic conditions, the study classified average monthly rainfall into five categories: from 5 to 65.2 mm, 65.2 to 125.4 mm, 125.4 to 185.6 mm, 185.6 to 245.8 mm, and 245.8 to 306 mm. The number of rainy days per month was similarly grouped: 4 to 7.2 days, 7.2 to 10.4 days, 10.4 to 13.8 days, 13.8 to 16.8 days, and 16.8 to 20 days. Average monthly temperature was divided into the ranges of 2.7 to 8.6 °C, 8.6 to 14.5 °C, 14.5 to 20.4 °C, 20.4 to 26.3 °C, and 26.3 to 32.2 °C. For monthly relative humidity, the categories were set from 62.0% to 66.8%, 66.8% to 71.6%, 71.6% to 76.4%, 76.4% to 81.2%, and 81.2% to 86.0%. Lastly, average sunshine hours per month were categorized into ranges of 1.8 to 3.06 hours, 3.06 to 4.32 hours, 4.32 to 5.58 hours, 5.58 to 6.84 hours, and 6.84 to 8.1 hours. The χ² test results revealed statistically significant differences in the incidence rates of foodborne disease outbreaks across the various climatic categories in Guizhou Province from 2012 to 2022. The χ² values are respectively: 2,122.142, 1,066.166, 2,753.543, 1,656.289, and 1,739.290, all P≤0.001 (Supplementary Table S1).

Based on the Prophet model: According to Supplementary Figure S1, the prediction plot generated by the Prophet model (Supplementary Figure S1A) displayed that all predicted values fell within the 95% confidence interval (CI). The assessment metrics indicated a good fit of the Prophet model with RMSE=67.645 and MAPE=42.357%. This demonstrates the model's capability in capturing the general incidence trend and seasonal patterns of foodborne disease outbreaks. The upper segment of Supplementary Figure S1B suggests a potential increasing trend in foodborne illnesses in Guizhou Province in the future. The lower segment of Figure S1B illustrates the seasonal pattern of foodborne disease outbreaks, highlighting a peak season from June to September.

Based on the SARIMA model: In this study, 16 SARIMA models were finally listed, and three better models were selected for the series based on the AIC and BIC criteria. Each of the three models is expressed as SARIMA (1,1,1) (0,1,1)₁₂, SARIMA (0,1,2) (1,1,1)₁₂, SARIMA (0,1,2) (1,1,1)₁₂. The BIC values for the three models were 1,317.555, 1,322.103, and 1,321.861; the AIC values were 1,306.864, 1,308.739, and 1,308.497, respectively. For the three alternative models initially selected, RMSE and MAPE were used as the main prediction accuracy evaluation indexes, and the RMSE values of the three models were 53.953, 60.489, and 62.301, respectively; and the MAPE values were 37.740%, 36.021%, and 37.209%, respectively. A comprehensive comparison of the AIC and BIC values of the alternative models revealed that the SARIMA (1,1,1) (0,1,1)₁₂ model is the best, as can be seen in its prediction graph, fits well with the actual reported values (Supplementary Figure S2).

Based on the ARIMAX model: In this study, 16 ARIMAX models were evaluated, and three models were chosen based on AIC and BIC criteria. The selected models are ARIMAX (1,1,1) (0,1,1), ARIMAX (0,1,1) (0,1,1), and ARIMAX (1,1,1) (1,1,1). The BIC values for these models were 1,321.170, 1,325.520, and 1,324.470, while the AIC values were 1,308.400, 1,310.320, and 1,309.270, respectively. RMSE and MAPE were used to assess prediction accuracy, with the RMSE values being 26.544, 28.614, and 67.999, and the MAPE values being 15.289%, 20.441%, and 44.102%, respectively. Comparing the AIC and BIC values, the model ARIMAX (1,1,1) (0,1,1) was found to be the best, showing good agreement with actual data (Supplementary Figure S3).

Based on the SARIMA-ARIMAX combination model: The RMSE and MAPE values were used to compare the predictive performance of two models simultaneously. The optimal sub-models selected were SARIMA (1,1,1) (0,1,1)₁₂ with a MAPE value of 37.740% and ARIMAX (1,1,1) (0,1,1) with a MAPE value of 15.289%. Weight coefficients of 0.246 and 0.654 were assigned to the SARIMA and ARIMAX models, respectively, based on calculations. The expression for the combined SARIMA-ARIMAX model is:

The respective predictive outputs of two submodels are weighted according to their associated coefficients and then aggregated to determine the forecast of the combined SARIMA-ARIMAX model. This integrated approach was employed to model the occurrence of foodborne disease outbreaks in Guizhou Province from 2012 to 2022. The resulting fitted data aligned well with the original trend (Supplementary Figure S4). Data from foodborne disease outbreaks between January and December 2022 constituted the test set. Evaluation of this test set indicated that the combined SARIMA-ARIMAX model achieved RMSE of 26.196 and MAPE

Upon a thorough examination of the prediction curves of various models forecasting the occurrence of foodborne disease outbreaks in Guizhou Province from January to December 2022, it is evident that all models' projected values align closely with the actual data. Analyzing the RMSE and MAPE metrics, the SARIMA, ARIMAX, and SARIMA-ARIMAX models developed in this study outperform the benchmark Prophet model. Additionally, the ARIMAX model surpasses the SARIMA model individually, while the combined SARIMA-ARIMAX model excels over the three standalone models. Forecasts using the optimal SARIMA-ARIMAX model for 2023 to 2025 indicate a stable trend in foodborne diseases in Guizhou Province, with approximately one to two peak periods each year (Supplementary Figure S4).

Foodborne illness represents a significant public health challenge globally, and in China, it stands as the paramount concern for food safety (10–11). Factors influencing the incidence of foodborne disease outbreaks are manifold, including human, natural, and geographic variables, with distinct characteristics observed across various regions (12). Enhancing research on foodborne disease outbreaks within different localities aids in devising prevention and control strategies that are more effectively customized and targeted, thereby diminishing the impact and burden of these outbreaks (13). In this study, a time series analysis was performed using surveillance data of foodborne disease outbreaks in Guizhou Province spanning from 2012 to 2022 to forecast future patterns. Findings indicated that outbreaks in Guizhou Province exhibited marked seasonal trends, with statistically significant correlations between incident rates and meteorological factors, and predicted a relatively stable trend moving forward.

Analysis of foodborne disease outbreaks in Guizhou Province indicates that variations in outbreak rates across climatic subgroups are statistically significant, aligning with findings from the study by Xiaojuan Qi et al. (14). Predictive models also revealed seasonal spikes in outbreaks, with higher incidences occurring during the warmer and wetter months of summer and autumn. These patterns suggest a probable link between climate change and the prevalence of foodborne illnesses (15). While the seasonal proliferation of such diseases can be attributed to a range of factors, including environmental conditions, climate, insect vectors, and human behavior (16), research has established that shifts in climate can influence the frequency of foodborne infections. For instance, a rise in average temperatures may enhance the growth of pathogens like Salmonella and Campylobacter, thereby escalating the risk of foodborne illnesses (17). Nonetheless, these influences are multifaceted rather than straightforward and warrant comprehensive study and analysis.

Forecasting plays a crucial role in decision-making and planning, especially in predicting foodborne disease outbreaks. The combination of various prediction models suggests a gradual increase in foodborne disease outbreaks in Guizhou Province over the next few years. This indicates the importance of maintaining rigorous monitoring, warning, prevention, and control measures. Analysis of prediction graphs highlights June–September as peak incidence months, with possible yearly peaks. Timely intervention strategies, effective communication, and proactive measures are essential for reducing the occurrence of foodborne disease outbreaks during these critical periods.

Each model possesses unique strengths and weaknesses. Although the Prophet model yields clearer results, it lacks the predictive capabilities of the SARIMA, ARIMAX, and the combined SARIMA-ARIMAX models. The SARIMA model demonstrated superior predictive performance compared to the Prophet model in forecasting episode numbers within a single model but fell short of the multivariate analysis model, ARIMAX. Overall, the combined SARIMA-ARIMAX model exhibited the highest predictive accuracy among the four models.

The SARIMA-ARIMAX combination model, weighted by MAPE, demonstrated superior predictive performance compared to other models. The forecast suggests that the frequency of foodborne disease outbreaks in Guizhou Province may exhibit a relatively stable trend during the period of 2023 to 2025, with one or two peak occurrences annually.

The findings highlight the pivotal role of accuracy, completeness, and chain consistency in foodborne disease outbreak reports for the stability of prediction models. Factors affecting foodborne outbreaks extend beyond meteorological conditions to include local economic and dietary cultural aspects. Future prediction models should prioritize authentic data acquisition, incorporate various influencing factors, and integrate multidisciplinary approaches to enhance accuracy and reliability.

No conflicts of interest.

All participating institutions for providing information and support throughout the study. The authors also acknowledge the research team members involved in data collection.