Abstract
BACKGROUND: Severe fever with thrombocytopenia syndrome(SFTS) is an emerging tick-borne disease with an expanding range and increasing public health burden. Meteorology-driven frameworks that integrate qualitative prediction with quantitative risk estimation while accommodating lag, regional heterogeneity, autoregressive case count effects, and zero-inflated counts remain scarce. METHODS: Monthly SFTS case counts and meteorological data from thirteen prefecture-level cities in Liaoning Province, China, from 2010 to 2024 were analyzed. Fushun was excluded because all counts were zero. Predictors were screened by correlation and variance inflation factor (VIF), and Boruta plus conditional permutation importance selected nine variables. Cities were grouped by k-means clustering. Four algorithms, including random forest (RF), extreme gradient boosting (XGBoost), gradient boosting decision tree (GBDT), and light gradient boosting machine (LightGBM), classified case presence using 2010-2022 training with ten-fold cross-validation and 2023-2024 testing. Shapley additive explanations (SHAP) interpreted variable importance and lagged associations in Dalian and Dandong. A mixed generalized additive model (MGAM) with distributed lag nonlinear modeling (DLNM) estimated exposure-lag effects of each meteorological main exposure. RESULTS: Nine meteorological variables were retained: wind speed (WS), relative humidity (RH), precipitation (PRCP), air pressure(AP), sunshine duration (SD), diurnal temperature range (DTR), surface air temperature difference (STD), standardized precipitation evapotranspiration at one month (SPEI1), and six months (SPEI6). K-means clustering grouped the thirteen Liaoning cities into three climatic groups. Across four classifiers, RF performed best in high-incidence areas, XGBoost was most stable; SHAP revealed opposite lag effects for some variables, indicating nonlinear delayed influences. Quantitative risk estimation selected the optimal covariates for each main exposure, characterized exposure response shapes: inverted U for WS, AP, PRCP, DTR, and SPEI6; monotonic increase for RH and SD; monotonic decrease for STD; bimodal for SPEI1. CONCLUSIONS: This study identifies meteorological heterogeneity in high-incidence regions while quantifying province-wide risk windows for each meteorological exposure, thereby informing regional and provincial prevention and early warning strategies.