Machine Learning in Air Pollution Research: Assessing Chemicals, CO, and NOx Exposure Effects on Liver and Endocrine Systems

Abstract

Abstract Purpose Air pollution is a recognized threat to metabolic health, but the complex relationships between specific pollutant sources, exposure duration, and liver and endocrine outcomes remain poorly understood. Traditional statistical methods often fail to capture the non-linear, high-dimensional nature of these interactions. Methods This study leverages a machine learning (ML) framework—including XGBoost, Random Forests, and neural networks—to model these complex associations using a large, integrated dataset of environmental and simulated clinical data. We generated a synthetic cohort to analyze the impact of multi-source exposure to chemicals including CO, NOx, and volatile organic compounds. Results Our models identified high-risk pollutant-source combinations, such as formaldehyde from wildfire smoke and benzene from oil and gas operations, as major predictors of simulated hepato-endocrine impact. Tree-based models demonstrated superior performance, and feature importance analysis revealed that interactions between pollutants and socioeconomic factors significantly improved predictive power. While longer exposure durations were associated with higher impact scores, the specific source was a critical determinant of risk. Discussion This study implemented a sophisticated machine learning framework to investigate the complex relationships between multi-source air pollution exposure and simulated hepato-endocrine health outcomes, demonstrating the potent utility of ML models in deciphering the high-dimensional, non-linear interactions that characterize environmental health data. By moving beyond traditional linear models, this study was able to identify specific pollutant-source combinations and exposure dynamics that may contribute to liver and endocrine dysfunction, building upon established epidemiological links between air pollution and metabolic diseases. The identification of pairs like Formaldehyde-Wildfire Smoke and Benzene-Oil and Gas Production as high-impact combinations suggests that the health impact of a chemical is not merely a function of its concentration but is profoundly shaped by its source. In this regard, wildfire smoke is a complex mixture where formaldehyde may act synergistically with other pyrolytic compounds like polycyclic aromatic hydrocarbons, potentially leading to enhanced hepatotoxicity through oxidative stress pathways [20, 21]. Similarly, benzene from oil and gas operations might co-occur with other volatile organic compounds, creating a mixture with greater endocrine-disrupting potential than any single pollutant alone [22, 23], which underscores a significant drawback of regulatory frameworks that often target individual pollutants and highlights the need for a source-oriented risk assessment strategy [24]. A consistent, yet modest, signal linked longer exposure durations to higher impact scores, a trend that aligns with the biological plausibility of cumulative damage [2]. However, the small effect sizes could be attributed to factors inherent in this simulation. The discretization of exposure into broad "buckets" may have obscured more subtle, continuous dose-response relationships, a known issue in environmental health studies [25], and our synthetic data may not fully capture complex biological mechanisms like adaptive responses or threshold effects that modulate the relationship between duration and harm [26]. From a methodological standpoint, the strong performance of tree-based models is consistent with their known strengths in handling tabular data with complex interactions [27, 28], and their compatibility with SHAP analysis was crucial. It revealed that the interaction between a pollutant and an individual's age or BMI could be a more significant predictor than the pollutant's concentration alone, echoing findings on susceptibility factors in air pollution health effects [29]. While the neural network's ability to capture complex, non-linear patterns is promising, its "black box" nature presents a challenge [30], and its lower estimated impact might reflect its modeling of more subtle, interactive effects that are harder to distill into a single feature importance score. The feature importance analyses further underscore the value of our integrated data approach. The prominence of features like "cumulative exposure" (pollutant concentration * log(duration)) and "pollutant-wind speed interaction" suggests that accounting for toxicokinetics and atmospheric dispersion is critical for accurate risk characterization. This aligns with a growing body of research emphasizing that the biologically effective dose is a function of both concentration and time [31], and that meteorological conditions can significantly amplify personal exposure by influencing pollutant buildup in specific areas [32, 33]. Furthermore, the model's identification of certain chemical-source interactions over others provides a data-driven ranking for prioritizing toxicological investigation into these specific mixtures, which has been a major challenge in environmental health [34]. The pioneering use of a synthetic cohort in this context is both a strength and a primary limitation, allowing for a controlled, proof-of-concept validation of our ML pipeline and the generation of hypotheses about high-risk exposures [35, 36]. However, this approach introduces critical constraints. The health outcomes, generated based on predefined functions, inevitably simplify the immense biological complexity of liver and endocrine pathophysiology, meaning the "impact scores" are proxies, not clinically validated endpoints, and their relationship to real-world disease states remains unknown [37]. Furthermore, the present model cannot account for all potential confounders that exist in real populations, such as genetic predisposition or diet [38, 39], and the spatial linkage of exposure based on postcode is a common but crude proxy for personal exposure, which varies based on individual mobility and indoor air quality [40, 41]. The discretization of exposure windows may also miss critical acute exposure events or sensitive developmental windows pivotal for endocrine disruption [42, 43]. Notwithstanding these limitations, this study provides a data-driven framework for prioritizing public health research and interventions. The identification of high-impact chemical-source pairs offers actionable intelligence for policymakers, who could focus on implementing stricter source-specific controls on industrial flaring and investing in wildfire management, as suggested by studies on the efficacy of source control [44, 45]. Other actions include enhancing air quality monitoring networks near identified high-impact sources and developing targeted public health advisories for vulnerable populations [46, 47], as well as using model insights to inform urban planning, such as increasing buffer zones to reduce the population-level exposure burden [48]. This work thereby lays the groundwork for several critical next steps. The foremost priority is to apply this validated ML pipeline to a large-scale, longitudinal cohort with precisely measured clinical endpoints and personal exposure data [37, 49]. Future work should also incorporate wearable sensor data to refine exposure assessment [40] and integrate causal inference frameworks to better isolate the causal effect of pollutants from confounding factors [38]. Finally, expanding this framework across diverse geographic and socio-economic contexts will test its generalizability and help ensure that public health strategies are equitable and context-aware [50]. In conclusion, this study demonstrates that machine learning provides a powerful and necessary toolkit for untangling the complex web of air pollution's effects on metabolic and endocrine health. By identifying high-risk pollutant-source combinations and highlighting the nuanced role of exposure duration, we have generated specific, testable hypotheses for future research. While the synthetic nature of our cohort calls for cautious interpretation, it successfully validates an analytical framework ready for deployment on real-world clinical and environmental data, with the ultimate goal of translating these data-driven insights into evidence-based public health policies and targeted interventions. Conclusions This study demonstrates ML's utility in generating actionable hypotheses for targeted public health interventions and prioritizing toxicological research, providing a validated framework for future studies with real-world clinical data. Keywords: Machine learning; Air pollution; CO exposure; NOx exposure; Liver function; Environmental health