Work

Machine learning-driven analysis of occupational heat exposure and metabolic associated fatty liver disease in petrochemical workers

Year: 2025

Type: article

Abstract: OBJECTIVES: This study aimed to investigate the pathogenesis of metabolic-associated fatty liver disease (MAFLD) in petrochemical workers chronically exposed to heat and industrial chemicals, using interpretable machine learning to identify synergistic risk factors. METHODS: A nested case-control study was conducted among 1277 petrochemical workers (mean follow-up: 8.20 ± 1.17 years). A total of 347 MAFLD cases were matched with 930 controls (1:4 ratio) by age (± 5 years) and gender. Demographic, occupational, and clinical data were analyzed. After Relief-F feature selection, nine machine learning algorithms, including eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), Gaussian Naive Bayes (GNB), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Decision Tree (DT), were optimized for MAFLD prediction. Model performance was rigorously evaluated using area under the receiver operating characteristic curve (AUC-ROC) analysis with DeLong's test for statistical comparison. SHAP values were used for model interpretability. RESULTS: Multivariate analysis revealed that prolonged heat exposure (> 10 years) significantly increased the risk of MAFLD when combined with diesel (OR = 2.31, χ² = 9.407, P = 0.009) or polyethylene dust (OR = 2.78, χ² = 10.645, P = 0.005). Certain socio-demographic factors were also played a significant role: widowed/divorced workers had a 1.89-fold higher risk of MAFLD compared with their married counterparts (χ² = 15.756, P < 0.001), and high-income earners (> $40,000/year) showed a 4.29-fold higher prevalence (χ² = 129.155, P < 0.001). The XGBoost model demonstrated greater stability, achieving a training AUC of 0.919 and maintaining a validation AUC of 0.761. The relatively smaller performance decline observed in XGBoost (ΔAUC = 0.158) reflects better robustness and resistance to overfitting than RF. Shapley value analysis revealed visceral adiposity (0.412), marital status (0.387), and heat-diesel co-exposure (0.218) as primary drivers of MAFLD risk. CONCLUSIONS: Chronic occupational heat exposure is a significant risk factor for MAFLD among petrochemical workers, particularly when combined with chemical exposures.

Source: Ecotoxicology and Environmental Safety

DOI: 10.1016/j.ecoenv.2025.119512