Explainable Artificial Intelligence (AI) Modeling of Co-Exposure Synergies in Tobacco Farming Health Risk Assessment with Epidemiological Analysis

Occupational health research in tobacco farming has long been constrained by the absence of structured, multi-hazard datasets and the prevailing use of single-exposure frameworks that fail to capture the risks faced by smallholder farmers.To fill the gaps, this study utilizes the Tobacco Farmers' Multi-Hazard Occupational Health and Vulnerability dataset that aims to record chemical, environmental, ergonomic, and physiological exposure dimensions, as well as clinical and selfreported health outcomes, for tobacco farmers in the northern part of Bangladesh, in Lalmonirhat district.This is the information that is used to build an ensemble machine learning-epidemiological interaction analysis-explainable artificial intelligence-comprehensive co-exposure vulnerability mapping framework integrated into a single multi-hazard pipeline.The three identified vulnerability archetypes show very different exposure profiles and Green Tobacco Sickness (GTS) prevalence patterns, thereby enabling stratification of vulnerability beyond that achievable with a single hazard analysis.The ensemble classifiers achieved excellent predictive performance (receiver operating characteristic-area under the curve ) for GTS, hypertension, and respiratory diseases.Physiological stress is a significant predictor of hypertension (odds ratio (OR) = 4.037, p = 0.0055) and respiratory illness (OR = 20.889,p < 0.001) in logistic regression.At the same time, the chemical and physiological co-exposures also demonstrated the potential to generate synergistic amplification, as indicated by non-significant bootstrap confidence intervals.SHapley Additive exPlanations -based feature explanations showed that the contributions of each model remained consistent and stable across models, enabling transparency and interpretability for potential real-world deployment.A comparative analysis using several performance indicators shows that the proposed ensemble framework outperforms the traditional single-model and single-hazard approaches.The results offer a scalable, data-driven, and interpretable method of occupational health risk stratification among tobacco farmers in resource-constrained, low-income agricultural communities.