Key points are not available for this paper at this time.
BACKGROUND: Global climate change has significantly altered the reproduction conditions and transmission patterns of foodborne pathogens, leading to dynamic shifts in the seasonal distribution of foodborne diseases (FBDs). Existing studies primarily rely on traditional statistical models or single machine learning (ML) algorithms, which have limitations in capturing the nonlinear associations between meteorological factors and FBDs. METHODS: This study employed spatiotemporal scanning to investigate the spatiotemporal clustering characteristics of FBDs in Wuxi City, China, from 2019 to 2023. Four ML models, including decision tree (DT), backpropagation neural network (BPNN), extreme gradient boosting (XGBoost) and long short-term memory network (LSTM), were constructed by integrating FBD surveillance data and concurrent climate data to predict FBD risks. Shapley additive explanations (SHAP) were used to quantify the contributions of climatic factors to model predictions. RESULTS: Spatiotemporal scanning results revealed a significant seasonal clustering of FBDs in summer and autumn, with a 19.0% increase in the incidence rate in the primary clustering area during 2022–2023 compared to 2019–2021, and actual cases consistently exceeding predicted values across all periods. Model performance comparison showed that LSTM outperformed other models, achieving root mean square errors (RMSE) of 9.1021 and 8.1854, mean absolute errors (MAE) of 7.0461 and 5.7671, and symmetric mean absolute percentage errors (SMAPE) of 49.8365% and 43.2618% on the validation and test sets, respectively. SHAP value analysis identified temperature as the key climatic factor with a strong positive correlation with FBD risks, whereas the contributions of non-temperature factors varied significantly across different models. CONCLUSION: This study provides a scientific basis for accurate risk prediction and prevention strategies of FBDs through a multi-model comparison framework and interpretable analysis.
Qin et al. (Mon,) studied this question.