Identification of predisposing factors for hematotoxicity in occupational benzene-exposed workers using machine learning and SHAP interpretation.

Benzene is an important industrial material and environmental pollutant, and even low-level exposure has been associated with hematotoxicity. Therefore, the identification of predisposing factors for hematotoxicity induced by occupational benzene exposure is crucial for managing occupational benzene poisoning. Herein, we recruited 1513 petrochemical workers in southern China, including 847 control workers and 666 workers with hematotoxicity. After randomly dividing the dataset into a training set (75%) and a test set (25%), we employed six machine learning algorithms construct a prediction model. Through recursive feature elimination, 22 feature variables, including age, work years, department, benzene, toluene, xylene, noise, AST/ALT, ALP, GGT, TBIL, DBIL, TC, TG, LDL-C, HDL-C, GLU, SCr, Urea, UA, FEV1/FVC%, and DBP were initially identified. Styrene was included due to its occupational relevance, resulting in 23 input variables for model development. We found the XGBoost algorithm was the optimal model for identifying factors affecting hematotoxicity. SHAP analysis based on the XGBoost model further characterized the contributions of variables at both the overall and individual levels, identifying benzene, noise, work years, age, department, and liver function indicators as important factors while also showing the direction and magnitude of their associations with the predicted risk of hematotoxicity. In vitro experiments also shown that impaired liver function exacerbated 4-benzoquinone (1,4-BQ)-induced cytotoxicity in bone marrow cells and modulated the underlying molecular pathways. Together, liver function-related indicators emerged as important predictors of hematotoxicity risk among benzene-exposed workers. These findings may help inform occupational health surveillance and risk stratification in this population.