Predictive association between control measures and chikungunya fever incidence based on random forest and SHAP analysis

Abstract

Background Using the 2025 Chikungunya fever (CHIK) outbreak in Chancheng District, Foshan City, this study applied a random forest (RF) regression model combined with SHapley Additive exPlanations (SHAP) to explore predictive associations, nonlinear relationships and potential thresholds between environmental–social factors and village/community-level cumulative incidence, to inform stratified control of mosquito-borne diseases. Methodology/principal findings In this cross-sectional ecological study of 143 villages/communities, the outcome was cumulative incidence, and nine candidate covariates were assessed, including the hospitalization isolation rate, construction-site density, and population density. Multicollinearity was checked using the variance inflation factor. Model fit was evaluated by the out-of-bag (OOB) R², RMSE, and MAE, and variable importance by %IncMSE. Robustness was tested with 100 repeated runs, bootstrap thresholds from SHAP dependence plots, and a sensitivity analysis excluding the endogenous isolation rate. On the log(1 + incidence) scale, the OOB R² was 0.206, with underestimation of high-incidence areas. The hospitalization isolation rate had the highest importance (%IncMSE = 18.43) and was negatively correlated with predictions (ρ = −0.749), but this likely reflects reverse causation and is predictive only. Construction-site density was strongly positive (ρ = 0.855), with a stable threshold near 12.8 sites/km²; it remained the most robust predictor after removing the isolation rate (%IncMSE = 8.54). Conclusions/significance Construction-site density was the most robust environmental predictor, whereas population density contributed little. These exploratory, predictive associations—not causal effects—should guide risk stratification and require prospective validation with time-matched longitudinal data.