Prediction of brucellosis incidence in China's five highest-incidence provinces: Comparing time-series models with multi-source environmental predictors

Abstract

Background Brucellosis is a severe zoonotic disease with pronounced seasonality and regional heterogeneity in high-incidence areas of China. Reliable forecasting tools are needed to inform prevention strategies, but the optimal modeling approach across different regions remains unclear. Principal Findings We collected monthly brucellosis incidence and 17 environmental variables from 2014 to 2024 across five high-incidence provinces: Inner Mongolia, Xinjiang, Shanxi, Heilongjiang, and Hebei. A three-step procedure--cross-correlation analysis, multicollinearity diagnostics, and stepwise regression--was used to select exogenous predictors. We then compared four time-series models: seasonal autoregressive integrated moving average (SARIMA), SARIMA with exogenous variables (SARIMAX), long short-term memory (LSTM), and LSTM with exogenous variables (LSTMX). All five provinces showed a unimodal seasonal pattern with peaks between April and July, though environmental drivers and optimal lag periods varied substantially by region, ranging from 1 to 6 months. In forecasting performance, LSTM achieved the highest accuracy in Shanxi (R2=0.925), Hebei (R2=0.876), and Xinjiang (R2=0.829), outperforming SARIMA and SARIMAX. LSTMX performed best in Inner Mongolia (R2=0.759) and Heilongjiang (R2=0.772) but showed weaker performance than LSTM in Shanxi and Hebei. Overall, adding exogenous variables did not consistently improve predictions across provinces. Conclusions Our findings demonstrate that LSTM-based models offer clear advantages for brucellosis forecasting in most high-incidence provinces, but the value of incorporating environmental predictors is region-dependent. These results support the development of tailored early warning systems and precision prevention strategies for brucellosis in high-risk areas of China.