A spatiotemporal risk prediction framework for PRRSV based on multisource data integration

Objective To propose and validate a spatiotemporal risk prediction framework for porcine reproductive and respiratory syndrome virus (PRRSV) by integrating epidemiological, environmental factors, and phylodynamic data, thereby addressing the limitations of existing transmission risk models in dynamically capturing transmission processes and quantifying risk gradients.

Method A multidimensional feature system was constructed, and a continuous risk score was used to consistently quantify transmission risk at the “provincial level administrative region-month” scale. Comparative experiments were performed using a classical baseline model (Gradient boosting), traditional machine learning models (SVR, SVR-L, and XGBoost), and deep learning models (LSTM and Transformer). Model performance was further evaluated by provincial level administrative region-wise cross-validation and rolling time-window validation.

Result All six models yielded lower mean absolute error (MAE) values when continuous risk score labels were used than when conventional binary labels were used. Phylodynamic/transmission features and historical features were the principal sources of information contributing to model performance; When used independently, they recovered 93.6% and 90.3% of the predictive performance of the full-feature baseline model, respectively. When engineered features were used and the training data proportion increased from 20% to 100%, the MAE of the Transformer model decreased from approximately 0.39 to 0.09, while that of the LSTM model decreased from approximately 0.14 to 0.07. In validation across provincial-level administrative regions, XGBoost and Gradient boosting exhibited strong overall robustness, whereas the Transformer model showed regional adaptability in Henan, Shandong and other regions.

Conclusion The proposed spatiotemporal risk prediction framework improves the accuracy of PRRSV transmission risk prediction and enhances the utilization efficiency of small-sample data, providing a methodological reference for the targeted prevention and control of PRRSV and other animal infectious diseases.