Occupational thermal exposure and perceived thermal load in naturally ventilated polyhouses: a field-based exposure-modeling study under dry-humid conditions.

PURPOSE: This study evaluated occupational thermal exposure and perceived thermal load among agricultural workers in naturally ventilated polyhouses under dry-humid climatic conditions using a worker-centered analytical framework integrating field-based environmental measurements, worker-reported thermal sensation vote (TSV), Bayesian modelling, and interpretable machine-learning analyses. METHODS: A cross-sectional field study included 154 independent worker observations collected during routine agricultural activities. Workplace air temperature, globe temperature, relative humidity, and air velocity were measured, and workers reported TSV. Agreement between predicted mean vote (PMV) and TSV was assessed using correlation, Bland-Altman, intraclass correlation coefficient (ICC), Wilcoxon signed-rank, and receiver operating characteristic (ROC) analyses. Bayesian linear regression, accumulated local effects, feature interaction analysis, and SHapley Additive exPlanations (SHAP) were used to identify determinants of perceived thermal load. RESULTS: Workers were routinely exposed to thermally demanding conditions, with 48.4% reporting thermal sensations outside the conventional thermal acceptability range. PMV showed modest correlation with TSV (Pearson's r = 0.246; Spearman's ρ = 0.338) and moderate agreement (ICC = 0.671), but poor discrimination of thermal acceptability (area under ROC curve = 0.39). Bayesian regression identified metabolic workload (MET) as the strongest positive predictor of perceived thermal load, whereas SHAP ranked age, body mass index, predicted percentage dissatisfied, and MET as the most influential predictors. CONCLUSION: Worker-reported TSV provides valuable information to conventional steady-state thermal comfort indices for occupational heat assessment in naturally ventilated polyhouses. Integrating Bayesian modelling with interpretable machine-learning techniques offers a transparent worker-centered framework for occupational heat assessment under real agricultural working conditions.