Personal and environmental determinants of oxidative stress among the residents of Ulsan, South Korea: a combined approach of statistical and machine learning models

Chiwon Sung; Dong Yeop Lee; Miso Lee; Chang Sun Sim; Inbo Oh; Jin-Hee Bang; Dong Yoon Kang; Jiho Lee

doi:10.35371/aoem.2026.38.e41

Articles

Page Path: HOME > Ann Occup Environ Med > Accepted articles > Article

Original Article Personal and environmental determinants of oxidative stress among the residents of Ulsan, South Korea: a combined approach of statistical and machine learning models: Chiwon Sung¹, Dong Yeop Lee^2,3, Miso Lee¹, Chang Sun Sim¹, Inbo Oh⁴, Jin-Hee Bang⁴, Dong Yoon Kang^2,3^,*, Jiho Lee¹^,*; Annals of Occupational and Environmental Medicine [Epub ahead of print]
DOI: https://doi.org/10.35371/aoem.2026.38.e41
Published online: September 1, 2026

¹Department of Occupational and Environmental Medicine, Ulsan University Hospital, University of Ulsan College of Medicine, Ulsan, Korea

²Department of Preventive Medicine, University of Ulsan College of Medicine, Ulsan, Korea

³Department of Preventive Medicine, Ulsan University Hospital, Ulsan, Korea

⁴Environmental Health Center, University of Ulsan College of Medicine, Ulsan, Korea

*Corresponding author: Jiho Lee Department of Occupational and Environmental Medicine, Ulsan University Hospital, University of Ulsan College of Medicine, 877 Bangeojinsunhwando-ro, Dong-gu, Ulsan 44033, Korea E-mail: oemdoc@naver.com

Dong Yoon Kang Department of Preventive Medicine, Ulsan University Hospital, University of Ulsan College of Medicine, 25 Daehakbyeongwon-ro, Dong-gu, Ulsan 44033, Korea E-mail: kdy@uuh.ulsan.kr

Jiho Lee and Dong Yoon Kang contributed equally to this work.

• Received: May 26, 2026 • Revised: August 1, 2026 • Accepted: August 18, 2026

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

722 Views
43 Download

Full Article

Download PDF

Abstract
BACKGROUND
METHODS
RESULTS
DISCUSSION
CONCLUSIONS
Abbreviations
NOTES
SUPPLEMENTARY MATERIAL
REFERENCES

Abstract

Background
This study aimed to identify personal and environmental factors influencing malondialdehyde (MDA) levels, measured using thiobarbituric acid reactive substances (TBARS) assay, among residents of Ulsan, South Korea.
Methods
A total of 909 adult residents of Ulsan, recruited between 2022 and 2024, participated in a survey and urine analysis. The dependent variable was the natural log-transformed TBARS value (Ln_TBARS). The study examined 47 independent variables, including personal characteristics, socioeconomic factors, work-related exposures, residential area, and living environment. Residential areas were classified into three groups (A, B, and C) according to potential exposure to industrial hazardous air pollutants, determined jointly by proximity to large-scale industrial complexes and modeled concentrations of benzene, toluene, and xylene. Stepwise multiple regression, least absolute shrinkage and selection operator (LASSO) regression, and random forest (ntree = 500) were applied and compared to identify factors influencing oxidative stress levels among Ulsan residents.
Results
Residential area had the greatest impact on oxidative stress across all three models. Group C showed significantly lower Ln_TBARS levels than those in group A (stepwise: β = −0.395, p < 0.001; LASSO coefficient = −0.529; random forest: %IncMSE = 17.07%). Age and female sex were also consistently associated with oxidative stress across all three models (stepwise: β = 0.319, p < 0.001; LASSO coefficient = 0.014; random forest: %IncMSE = 12.67%, and stepwise: β = 0.152, p < 0.001; LASSO coefficient = 0.154; random forest: %IncMSE = 6.93%).
Conclusions
Residential area was the strongest predictor of oxidative stress across all three analytical models, with higher urinary MDA among residents of areas closer to industrial complexes; older age and female sex were also consistent predictors. These findings, robust across conventional statistical and machine-learning approaches, support systematic environmental health monitoring and targeted policy measures for residents near industrial zones.
Keywords: Least-squares analysis; Oxidative stress; Random forest; Regression analysis; Residential characteristics

BACKGROUND

Oxidative stress, both a cause and consequence of numerous diseases including cardiovascular disease, diabetes, dyslipidemia, neurodegenerative disorders, and cancer, has drawn growing research interest given its high social and economic burden.¹^-⁵ It is a pathological state marked by excessive reactive oxygen species (ROS) production and insufficient antioxidant defense. Although ROS are essential for metabolism, immunity, and inflammatory responses, their excess can cause cell death, organ damage, and DNA damage.⁶

Urinary malondialdehyde (MDA), a byproduct of lipid peroxidation in cell membranes, is widely used as a biomarker of oxidative stress and can be measured using the thiobarbituric acid reactive substances (TBARS) assay.⁷^,⁸ Although high-performance liquid chromatography methods are used for 8-hydroxy-2′-deoxyguanosine or F2-isoprostane, TBARS-based MDA measurement remains predominant in large-scale, community-based studies for its cost-effectiveness and practicality.⁹^,¹⁰ Urine sampling is noninvasive, cost-effective, facilitates voluntary participation, and better reflects long-term exposure than blood, making it suitable for large-scale environmental studies.¹¹^,¹²

Ulsan, South Korea, contains large-scale heavy industrial complexes—shipbuilding yards, petrochemical plants, and automobile factories—located close to residential areas.¹³ National government surveys have documented elevated pollutant exposure and adverse health indicators among residents near the Ulsan-Onsan industrial complex compared to those living farther away.¹⁴^,¹⁵ Such conditions expose residents to air pollutants, heavy metals, and organic solvents, established contributors to oxidative stress.¹⁶ Previous studies have focused mainly on chemical exposures in specific occupational groups or lifestyle factors such as smoking and alcohol consumption,¹⁷^,¹⁸ with little attention to residential location and living environment—research essential for evidence-based policy in industrialized cities such as Ulsan.

This study analyzed associations between oxidative stress (urinary MDA) and personal characteristics, socioeconomic factors, work-related exposures, residential area, and living environment among 909 Ulsan residents. Given the complex interrelationships among environmental factors, we combined two machine learning methods with traditional analysis for a multilayered assessment of each factor’s contribution.

First, multiple regression examined linear associations and derived standard determinants of oxidative stress based on statistical significance. Second, least absolute shrinkage and selection operator (LASSO) regression addressed multicollinearity and overfitting, extracting a concise set of highly explanatory variables.¹⁹^,²⁰ Third, random forest identified nonlinear relationships and ranked variables by importance.²¹^,²²

This study aimed to compare these three analytical methods and identify key factors influencing oxidative stress, thereby providing a scientific basis for environmental health policies.

METHODS

Study design

This study employed a cross-sectional design to examine the associations of personal characteristics, socioeconomic factors, work-related exposures, residential area, and living environment with urinary MDA levels among adults residing in Ulsan.

Participants

The study comprised 909 Ulsan residents. Participants were recruited between 2022 and 2024 by the Ulsan Metropolitan City Environmental Health Center through administrative collaboration with local government authorities.

Participants were eligible if they were aged ≥19 years and had resided in the same area for at least 5 years to reflect long-term and consistent residential exposure. To ensure consistency in residence-based exposure assessment, participants were excluded if they had relocated within the previous 5 years or spent >8 hours daily outside their residential area for occupational purposes.

Setting

Participants were enrolled from the environmental health monitoring survey conducted by the Ulsan Metropolitan City Environmental Health Center and were initially organized by administrative district (dong/eup/myeon), the unit at which the survey was administered. By using districts as the base unit, each exposure group was kept spatially contiguous, whereas allocation by individual proximity or concentration alone would have resulted in groups dispersed across the city.

Participants organized in this way were then classified into three groups according to their potential exposure to industrial hazardous air pollutants (HAPs), determined jointly by two criteria: proximity of the district to the large-scale national industrial complexes (petrochemical and non-ferrous metal facilities), and concentrations of benzene, toluene, and xylene (BTX), which served as reference indicators of industry-derived air pollution. Group A was comprised of districts bordering the industrial complexes, representing those with the highest potential exposure to industrial HAPs. Group B was comprised of districts within the expected influence range of the complexes, representing intermediate potential exposure. Group C was comprised of districts beyond that range (i.e., without large-scale industrial emission sources), thus representing the lowest expected exposure. Table 1 summarizes the classification criteria and the corresponding potential industrial HAP exposure of each group.

BTX concentrations at participants' residential locations were estimated using Community Multiscale Air Quality–based model predictions of emissions from the Ulsan industrial complexes.²³^,²⁴ As participants were classified by district, the distances shown in Table 1 describe the range spanned by the constituent districts rather than individually measured distances; the ranges of groups A and B partially overlap. Throughout the analyses, residential area was treated as an ordinal surrogate for the rank of potential industrial HAP exposure rather than as a measure of absolute exposure concentration.

Variables

Dependent variable

The dependent variable was urinary MDA concentration (μmol/L). Urinary MDA concentrations were measured using the TBARS assay. Given that the raw data distribution deviated significantly from normality (Shapiro-Wilk W = 0.609, p < 0.001; skewness = 5.133; kurtosis = 40.486), a natural logarithmic transformation was applied. After transformation, the skewness and kurtosis of Ln_TBARS were −2.195 and 9.645, respectively. Normality was assumed based on the central limit theorem.²⁵ The natural log-transformed, creatinine-adjusted MDA concentration (hereafter referred to as Ln_TBARS) was used as the dependent variable in all analyses.

Independent variables

Forty-seven independent variables (89 including dummy variables) were selected across five domains: sociodemographic characteristics; residential environment; lifestyle; self-reported occupational exposure to hazardous agents; and residential area. The full variable list is provided in Supplementary Table 1. Height and weight, which overlapped with body mass index, and average alcohol consumption, which overlapped with alcohol consumption frequency, were excluded from multivariable analyses to reduce multicollinearity.

Data sources and measurement

Urinary MDA levels were assessed using the TBARS assay with the DG-TBA200 kit (DoGenBio, Seoul, Korea) per the manufacturer’s instructions, with butylated hydroxytoluene added as an oxidation inhibitor to prevent artifactual lipid peroxidation. Samples and MDA standards were reacted with thiobarbituric acid reagent (95°C–100°C, 60 minutes), cooled, and centrifuged (3,000 ×g, 10 minutes); supernatant absorbance was measured at 540 nm and converted to MDA concentration using a standard curve. Final values were standardized to urinary creatinine concentration and expressed as μmol/L.

Analytical quality control (QC) comprised a standardized sequence of reagent blank, calibration, QC-verified recalibration, and duplicate analysis of all samples. The method detection limit, derived from seven replicate measurements of a 0.5 μM standard addition, was 0.0214 μM in 2022 and 0.0200 μM in 2023–2024, indicating stable sensitivity, and the calibration curve showed excellent linearity throughout (r > 0.999). Intra- and inter-day accuracy (94.5%–102.7%) and precision (coefficient of variation [CV] 1.4%–2.4%) at low, medium, and high concentrations are summarized in Supplementary Table 2. Samples were processed in sequential batches in the order received by the laboratory, with technicians blinded to residential group assignment. Each batch included a calibration-verification standard and blank and spiked pooled urine QC samples; recovery remained within the acceptance range throughout (bias ≤2.4%; CV ≤3.9%) (Supplementary Table 3), and no batch required reanalysis.

Participant information was collected using a structured questionnaire adapted from the standard instrument used in the National Environmental Health Survey.

Statistical methods

Descriptive statistics were presented as mean ± standard deviation or frequency (%). Differences in Ln_TBARS according to categorical variables were assessed using independent-samples t-tests or one-way analysis of variance, followed by Scheffé’s post hoc test for multiple comparisons.

Traditional regression analysis is prone to overfitting and multicollinearity and has limitations in capturing nonlinear relationships.¹⁷^,²⁶^,²⁷ Therefore, two machine learning methods, namely, LASSO regression and random forest, were additionally applied.

For multiple linear regression, enter, forward, backward, and stepwise methods were applied to the full variable set. Model explanatory power was evaluated using the adjusted coefficient of determination (adjusted R²), and model fit was compared using the Akaike Information Criterion and Bayesian Information Criterion (Supplementary Table 4).

For LASSO regression, the optimal regularization parameter was determined via 10-fold cross-validation (seed = 42); λ_min (minimizing CV-MSE) and λ_1se (most parsimonious model within one SE of the minimum) were reported, and predictive performance was evaluated using CV-based R².

For random forest, 500 trees were generated (mtry = default one-third of input variables); approximately 63% of the sample was used for bootstrap construction, with the remaining 37% (out-of-bag, OOB) used to estimate prediction error, and performance was assessed using OOB-based pseudo R².

All statistical analyses were performed using R software, version 4.5.3 (R Foundation for Statistical Computing, Vienna, Austria); p-values of <0.05 were considered statistically significant for all analyses.

Ethics statement

The present study protocol was reviewed and approved by the Institutional Review Board of Ulsan University Hospital (IRB no. 2025-10-012). This study was conducted in accordance with the principles of the Declaration of Helsinki, and voluntary written informed consent was obtained from all subjects prior to enrollment. To ensure participant safety and data protection, all personal identification information was de-identified and handled with strict confidentiality.

RESULTS

Participants

Table 2 presents the general characteristics of the study participants. Mean age was 61.32 ± 14.13 years. Mean duration of residence and building age were 16.96 ± 13.99 and 23.12 ± 16.60 years, respectively. A total of 651 participants (71.6%) were women, 702 (77.2%) were married, and 96 (10.6%) were current smokers. The numbers of participants residing in groups A, B, and C were 324 (35.6%), 322 (35.4%), and 263 (28.9%), respectively. Mean Ln_TBARS values differed significantly across residential areas (F = 93.84, p < 0.001), with values of 1.100 ± 0.420 in group A, 0.845 ± 0.727 in group B, and 0.184 ± 1.050 in group C. Post hoc comparisons using Scheffé’s test revealed that Ln_TBARS levels were significantly higher in group A than in groups B and C and significantly higher in group B than in group C.

Multiple regression analysis (stepwise)

The stepwise model was selected as the final model based on the highest adjusted R² (0.293) and included 13 variables (Table 3, Supplementary Table 5). Compared with group A, residents in groups B and C showed significantly lower Ln_TBARS values (B = −0.187, β = −0.106, p = 0.001 and B = −0.738, β = −0.395, p < 0.001). Age was positively associated with Ln_TBARS, with each 1-year increase corresponding to a mean increase of 0.019 units in Ln_TBARS (B = 0.019, β = 0.319, p < 0.001). Female sex was associated with significantly higher Ln_TBARS values than male (B = 0.286, β = 0.152, p < 0.001). Residing adjacent to a four-lane road was associated with significantly higher Ln_TBARS values than residing adjacent to a two-lane road (B = 0.143, β = 0.080, p = 0.016).

LASSO regression

Table 4 presents the LASSO regression analysis results for Ln_TBARS. The optimal regularization parameters determined using 10-fold cross-validation were λ_min = 0.046 (CV-MSE = 0.536) and λ_1se = 0.155 (CV-MSE = 0.574). Under the λ_min criterion, eight variables were selected, with an R² of 0.251. The variable with the largest coefficient was residence in group C (coefficient = −0.529), indicating a strong association between regional exposure level and oxidative stress. Under the more conservative λ_1se criterion, only group C (coefficient = −0.359) and age (coefficient = 0.009) were retained.

Random forest

The random forest model (ntree = 500, mtry = 29) yielded an OOB-MSE of 0.556 and a pseudo R² of 0.223. Table 5 presents the top 10 variables ranked by importance. Group C ranked first in variable importance (percentage increase in MSE [%IncMSE] = 17.07%), followed by duration of residence (%IncMSE = 14.40%) and age (%IncMSE = 12.67%).

Summary of analysis results: stepwise regression, LASSO regression, random forest

Table 6 provides an overview of the three analytical techniques. Three variables were identified as consistent predictors across all three models: group C, age, and female sex. Variables identified as significant in the two analytical methods included the number of adjacent road lanes (four-lane roads), residence in group B, and duration of residence.

DISCUSSION

Key results

Across all three analytical approaches, residential area emerged as the strongest predictor of oxidative stress, with higher levels among residents of areas closer to the industrial complexes. Age and female sex were also consistent predictors, with oxidative stress increasing with age and being higher among women. Duration of residence and proximity to four-lane roads were significant in two of the three models.

Interpretation

The most salient finding that emerged from this multilayered analysis was the identification of residential areas as a highly significant predictor of oxidative stress. Specifically, oxidative stress levels were higher in areas with higher exposure to industrial HAPs. This suggests that pollutants from industrial complexes extend beyond the immediate vicinity into the surrounding influence zone, producing differential associations with oxidative stress according to residential proximity.²⁸ This association is likely attributable to chronic exposure to air pollutants and industrial chemicals prevalent in areas adjacent to industrial complexes, which may promote oxidative stress.²⁹ Furthermore, among the regional factors, duration of residence was identified as the second most important variable in the random forest analysis, suggesting that cumulative exposure in cities such as Ulsan—where exposure to air pollutants and industrial chemicals is high—may impact oxidative stress.³⁰^,³¹ Residents living near four-lane roads exhibited significantly higher oxidative stress, whereas the opposite trend was observed near six or more lanes (Supplementary Table 5), suggesting a nonlinear relationship potentially explained by differences in noise-avoidance behavior, time spent at home, and roadside vegetation or noise barriers.³²–³⁴

Several observations support the validity of the exposure classification by area. First, oxidative stress followed the same monotonic ordering as the expected magnitude of industrial HAP exposure across all three groups, and all pairwise post hoc contrasts were statistically significant. Moreover, residential area was consistently the leading predictor across three analytically distinct modelling frameworks, including the conservative λ_1se LASSO criterion, which retained only two variables. Because the classification was fixed without reference to participants’ oxidative stress levels, any residual misclassification would be non-differential, and estimates would be biased toward the null.³⁵ Thus, the observed gradient is more likely attenuated than exaggerated.

The various statistical approaches had complementary strengths. Stepwise regression gave interpretable effect estimates but was vulnerable to multicollinearity and overfitting. LASSO addressed this by shrinking coefficients and reducing overfitting, isolating “residence in group C” and “age” as core predictors under the conservative λ_1se. Random forest went further by requiring no linearity assumption. This uncovered nonlinear predictors that the traditional models missed; namely, duration of residence. The variables that were consistently significant across all three methods (residential area, age, and female sex) can be regarded as particularly robust predictors of oxidative stress.

The overall explanatory power across the three models (0.223–0.293) is consistent with levels commonly observed in epidemiological studies investigating complex biological responses and multidimensional environmental exposures, likely reflecting the influence of unmeasured genetic, occupational, and lifestyle factors, as well as the inherent limitations of single-point urinary MDA measurements.³⁶^,³⁷

Comparison with previous studies

Our finding that residence near industrial complexes was the strongest predictor of oxidative stress reaffirms previous studies linking industrial and air pollution exposure to adverse health effects among nearby residents.²⁸^,²⁹^,³⁸^,³⁹ However, most focused on terminal outcomes such as mortality or disease incidence, with few examining oxidative-stress biomarkers in large general-population samples. This study extends that literature by directly measuring a validated biomarker in 909 residents, with residential area consistently emerging as the top-ranked predictor across three analytical approaches.

The positive correlation that we observed between age and oxidative stress is also consistent with previous research findings, which have provided mechanistic evidence, including mitochondrial dysfunction and reduced antioxidant capacity.⁴⁰^,⁴¹

In a notable divergence from previous findings, our study found higher levels of oxidative stress in women than in men, in contrast to reports that estrogen confers greater antioxidant protection and lower oxidative stress in females.⁴² Given the high proportion of women (71.62%) in our sample and the relatively advanced mean age of our participants, this result may have reflected the post-menopausal loss of estrogen-associated antioxidant protection suggested by Signorelli et al.⁴³

Several factors previously linked to oxidative stress were not significant predictors here. Self-reported occupational exposure to hazardous agents is widely reported to be associated with oxidative stress,⁴⁴^,⁴⁵ yet it was not retained in any model. Because exposure intensity and duration were not measured, the resulting non-differential misclassification would bias associations toward the null.³⁵ In addition, occupational exposure overlapped geographically with residential area, since those living near large industrial sources are more likely to work in them; its independent contribution may therefore have been diluted under multicollinearity.

Smoking, a well-established source of oxidative stress,⁴⁶ showed a significant bivariate association in the opposite direction, with lower Ln_TBARS among current smokers versus non-smokers (p = 0.036), although this was not retained in any multivariable model. This inversion most likely reflects confounding by sex and age, since current smoking was more frequent among men and younger participants, both of which are groups with lower oxidative stress; the low proportion of current smokers (10.56%) and the sensitivity of urinary MDA to recent rather than cumulative exposure may also have contributed.

Alcohol consumption, associated with elevated MDA through free-radical generation during ethanol metabolism, is most consistently demonstrated among heavy drinkers.⁴⁷ Its absence here may reflect dilution by stronger predictors and by collinearity with sex, given that women, who predominated in our sample, report lower drinking frequency.

Finally, vitamin supplementation has previously been reported to reduce oxidative stress. However, accumulating evidence from both epidemiological and interventional studies suggests that this may modify the effects of exposure to specific pollutants (e.g., heavy metals or particulate matter) rather than exerting an independent main effect.⁴⁸^,⁴⁹ As we assessed main effects only, this interaction-dependent protective association may not have been captured.

Limitations

This study has some limitations. First, as this was a cross-sectional study, it was not possible to establish a direct causal relationship between residential environment and oxidative stress. Second, most exposure-related information was derived through self-reported questionnaires, introducing the potential for information bias. Third, exposure was assessed at the area level rather than the individual level. Because administrative districts were the base unit for grouping, participants within a group differed in their actual distance from the complexes, and the distance ranges of groups A and B partially overlap; because only three discrete areas were sampled, the findings should not be extrapolated to the unsampled intermediate range. The concentrations used to characterize the groups were model-based estimates before the study period rather than personal measurements, and these relied on BTX as a partial indicator of a complex pollutant mixture. Although independent assessments spanning more than a decade have identified the same spatial ordering of pollutant burden across these areas,¹³^-¹⁶ the absolute concentrations during 2022–2024 may have differed. Personal factors such as activity patterns and ventilation were also not captured, likely introducing some degree of exposure misclassification. Studies using individual-level geographic information systems and directly measured personal exposure are needed to quantify exposure–response relationships. Fourth, our analytical batches were not randomized across residential groups. However, laboratory personnel were blinded to participants’ residential group during sample processing, and batch-level QC monitoring showed no evidence of systematic drift, reducing the likelihood of batch-related bias. Fifth, the disproportionately high representation of women and older adults in the study sample warrants caution when generalizing the findings to the broader adult population. Further studies using more representative samples are needed. Sixth, health status was captured only as single self-reported items that indicated whether the participant was currently under diagnosis or treatment for a chronic disease, cancer, or an acute illness. Chronic disease was included in the final model but was not statistically significant (p = 0.155). Specific conditions such as hypertension and information on medication use were not collected, so residual confounding by these factors cannot be excluded. Finally, although the TBARS assay has lower chemical specificity than F2-isoprostane analysis and may overestimate the degree of lipid peroxidation,¹⁰ F2-isoprostane analysis is expensive and requires complex sample pretreatment procedures, limiting its practicality for large-scale studies such as the present investigation.⁵⁰ Therefore, the TBARS assay remains a useful and cost-effective screening tool.⁸ In addition, the use of butylated hydroxytoluene during sample processing to prevent artifactual oxidation likely enhanced the reliability of the measurements.¹⁰

Generalizability

Although the areal classification has limitations in reflecting individual exposure, it identified a broad regional gradient that can guide future research. We hope causality will be established through cohort studies using directly measured exposure data, laying a foundation for future public health policies.

CONCLUSIONS

This study systematically analyzed the personal characteristics, socioeconomic factors, work-related exposure, residential area, and living environment associated with urinary MDA concentrations among Ulsan residents using stepwise multiple regression, LASSO regression, and random forest analysis.

First, residential proximity to industrial complexes and the resulting exposure to complex mixtures of industrial HAPs were the factors most strongly associated with oxidative stress, with levels highest in group A, followed by groups B and C.

Second, age and female sex were consistently identified as core predictors even after controlling for residential area.

Third, the number of adjacent road lanes and duration of residence were identified as significant predictors in two analytical models.

Abbreviations

BTX

benzene, toluene, and xylene

coefficient of variation

HAP

hazardous air pollutant

LASSO

least absolute shrinkage and selection operator

Ln_TBARS

log-transformed TBARS value

MDA

malondialdehyde

OOB

out-of-bag

quality control

ROS

reactive oxygen species

TBARS

thiobarbituric acid reactive substances

NOTES

Funding

This work was supported by the Ulsan Environmental Health Center funded by the Ministry of Climate, Energy and Environment and Ulsan Metropolitan City, Republic of Korea.

Competing interests

Jiho Lee, a contributing editor of the Annals of Occupational and Environmental Medicine, was not involved in the editorial evaluation or decision to publish this article. All remaining authors have declared no conflicts of interest.

Author contributions

Conceptualization: Sung C. Data curation: Sung C, Lee M, Oh I, Bang JH. Formal analysis: Sung C, Lee DY, Lee M. Investigation: Lee J, Bang JH, Oh I. Software: Sung C, Lee DY. Methodology: Sung C, Kang DY, Lee J, Sim CS. Validation: Lee J, Kang DY. Writing - original draft: Sung C. Writing - review & editing: Sung C, Lee DY, Lee J, Kang DY, Lee M, Sim CS, Oh I, Bang JH.

Acknowledgments

The authors declare that the study design and all statistical analyses were conducted independently, without AI assistance. In accordance with ICMJE guidelines, Claude Sonnet 5 (Anthropic, PBC, San Francisco, CA, USA) was used solely to assist with translation, enhance the content and restructure the human-written draft into an appropriate academic format in March–August 2026. All AI-assisted text was extensively revised by the authors to ensure academic accuracy and intended messaging. All references were manually identified and verified using OpenEvidence (OpenEvidence, Inc., Miami, FL, USA) April–August 2026. Professional proofreading and editing were provided by Enago (Crimson Interactive Inc., New York, NY, USA) April–August 2026. The authors accept full responsibility for the accuracy and integrity of the final manuscript.

SUPPLEMENTARY MATERIAL

Supplementary Table 1.

Independent variables examined in the analyses categorized by domain.

aoem-2026-38-e41_Supplementary-Table-1.pdf

Supplementary Table 2.

Intra-day and inter-day accuracy and precision of the TBARS assay across three concentration levels.

aoem-2026-38-e41_Supplementary-Table-2.pdf

Supplementary Table 3.

Batch-level QC monitoring results for the TBARS assay.

aoem-2026-38-e41_Supplementary-Table-3.pdf

Supplementary Table 4.

Comparison of model performance by entry method in multiple linear regression analysis.

aoem-2026-38-e41_Supplementary-Table-4.pdf

Supplementary Table 5.

Full results of stepwise multiple regression analysis for Ln_TBARS (all selected variables).

aoem-2026-38-e41_Supplementary-Table-5.pdf

Table 1.

Classification criteria and potential exposure to industrial HAPs of the three residential groups

Group^a	No. (%)	BTX level, median^b (IQR) (ppb)	Proximity to industrial complex^c (IC)	Potential exposure to HAPs from IC
A	324 (35.64)	1.48 (0.78–2.88)	Districts bordering the IC (0–2 km)	High likelihood of long-term, high-concentration exposure to industrial HAPs
B	322 (35.42)	0.92 (0.83–1.02)	Districts within the expected influence range of the IC (1–3 km)	Moderate average exposure, with possibility of short-term, high-concentration exposure depending on wind conditions and proximity to the IC
C	263 (28.93)	0.72 (0.58–0.82)	Districts beyond the influence range, without large-scale industrial emission sources (>10 km)	Low likelihood of high-concentration exposure

HAPs: hazardous air pollutants; BTX: sum of benzene, toluene, and xylene concentrations; IQR: interquartile range; IC: industrial complex.

^aGroups were classified based on the degree of potential exposure to industrial HAPs, determined jointly by the proximity of the district to large-scale national industrial complexes and by BTX concentrations, which served as reference indicators of industry-derived air pollution;

^bBTX concentrations were estimated using predictions from the Community Multiscale Air Quality model, a chemical transport model that simulates ambient air pollutant concentrations (reported in a previous study). The original 1-km model outputs were spatially interpolated to a 100-m resolution, and the interpolated BTX concentrations at each participant’s residential location were extracted;

^cAs participants were organized by administrative district before classification, the distances shown represent the range spanned by the constituent districts rather than individually measured distances; the ranges of groups A and B therefore partially overlap. The BTX concentrations are presented as reference indicators of the relative exposure ranking across groups and were not used as absolute exposure measures in the analyses.

Table 2.

General characteristics of study participants and mean Ln_TBARS values by residential area (n = 909)

Variable	No. (%) or mean ± SD	Ln_TBARS (mean ± SD)	p-value
Age (years)	61.32 ± 14.13	-	-
Duration of residence (years)	16.96 ± 13.99	-	-
Building age (years)	23.12 ± 16.60	-	-
Sex			<0.001
Male	258 (28.4)	0.514 ± 0.957
Female	651 (71.6)	0.836 ± 0.773
Occupation			<0.001
Professional/manager/office	71 (7.8)	0.301 ± 1.143
Service/sales	128 (14.1)	0.724 ± 0.894
Skilled/manual labor/agriculture	196 (21.6)	0.703 ± 0.904
Homemakers/students/unemployed	514 (56.5)	0.827 ± 0.728
Marital status			<0.001
Married	702 (77.2)	0.691 ± 0.824
Not married	207 (22.8)	0.924 ± 0.875
Current smoking			0.036
Yes	96 (10.6)	0.575 ± 0.811
No	813 (89.4)	0.765 ± 0.843
Residential area^a			<0.001
Group A	324 (35.6)	1.100 ± 0.420
Group B	322 (35.4)	0.845 ± 0.727
Group C	263 (28.9)	0.184 ± 1.050

p-values were calculated using independent-samples t-test or one-way analysis of variance.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance value; SD: standard deviation.

^aPost hoc comparisons using Scheffé’s test: group A > group C; group A > group B; group B > group C.

Table 3.

Key results of the stepwise multiple regression analysis for Ln_TBARS

Independent variable	Reference^a	B	Standard β	95% CI	SE	p-value
(Intercept)	-	−0.465	-	−0.762 to −0.169	0.151	0.002
Group C	Group A	−0.738	−0.395	−0.874 to −0.602	0.069	<0.001
Group B	Group A	−0.187	−0.106	−0.301 to −0.073	0.058	0.001
Age	-	0.019	0.319	0.015 to 0.023	0.002	<0.001
Female sex	Male	0.286	0.152	0.178 to 0.394	0.055	<0.001
Adjacent road lanes: 4-lane	2-lane	0.143	0.080	0.027 to 0.258	0.059	0.016

Adjusted R² = 0.293. Full variable results are provided in Supplementary Table 5.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration; B: unstandardized coefficient; CI: confidence interval; SE: standard error.

^aReference categories: residential area = group A; road lanes = 2-lane.

Table 4.

Results of LASSO regression analysis for Ln_TBARS

Variable^a	λ_min^b coefficient (R² = 0.251, n = 8)	λ_1se^c coefficient (R² = 0.199, n = 2)
Residential area (group C)	−0.529	−0.359
Age	0.014	0.009
Female sex	0.154	-
Alcohol consumption (1–2 times/week)	−0.059	-
Adjacent road lanes (6-lane or more)	−0.039	-
Residential area (group B)	−0.017	-
Educational attainment (college graduate or above)	−0.008	-
Duration of residence (years)	0.001	-

LASSO: least absolute shrinkage and selection operator; Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration.

^aOptimal regularization parameters were determined via 10-fold cross-validation (seed = 42);

^bλ_min, minimum cross-validation MSE criterion;

^cλ_1se, most parsimonious model within 1 standard error of the minimum MSE (conservative criterion).

Table 5.

Top 10 variables by importance in random forest analysis for Ln_TBARS

Rank	Variable	%IncMSE^a
1	Residential area (group C)	17.07
2	Duration of residence	14.40
3	Age	12.67
4	Marital status (not married)	7.02
5	Female sex	6.93
6	Adjacent road lanes (4-lane)	6.74
7	Income (≥6,000,000 KRW/month)	6.48
8	Residential building type (mid-rise apartment)	6.45
9	Building age	6.35
10	Smoking history (pack-years)	5.08

Model parameters: ntree = 500, mtry = 29; OOB-MSE = 0.556; Pseudo R² = 0.223.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration; KRW: Korean won.

^a%IncMSE, percentage increase in mean squared error upon permutation of the variable (out-of-bag); higher values indicate greater variable importance.

Table 6.

Summary of analysis results: stepwise regression, LASSO regression, and random forest

Item	Stepwise regression	LASSO (λ_min)	Random forest	No. of models^a
Model fit (adjusted R²/CV-R²/Pseudo R²)	0.293	0.251	0.223	-
No. of selected variables	13	8	89 (by importance)	-
Residential area (group C)	✓	✓	✓ (1st)^b	3
Age	✓	✓	✓ (3rd)	3
Female sex	✓	✓	✓ (5th)	3
Adjacent road lanes (4-lane)	✓	-	✓ (6th)	2
Residential area (group B)	✓	✓	-	2
Duration of residence	-	✓	✓ (2nd)	2

LASSO: least absolute shrinkage and selection operator; CV-R²: cross-validation-based R²; Pseudo R²: out-of-bag-based R².

^aNo. of models, number of analytical approaches in which the variable was selected. Values are presented for reference and are not directly comparable across methods due to differences in analytical properties;

^b✓, variable selected by the respective model. Numbers in parentheses indicate rank by %IncMSE in random forest.

REFERENCES

1. Jomova K, Raptova R, Alomar SY, Alwasel SH, Nepovimova E, Kuca K, et al. Reactive oxygen species, toxicity, oxidative stress, and antioxidants: chronic diseases and aging. Arch Toxicol 2023;97(10):2499–574.Article PubMed PMC PDF
2. Sharifi-Rad M, Anil Kumar NV, Zucca P, Varoni EM, Dini L, Panzarini E, et al. Lifestyle, oxidative stress, and antioxidants: back and forth in the pathophysiology of chronic diseases. Front Physiol 2020;11:694.Article PubMed PMC
3. Valko M, Leibfritz D, Moncol J, Cronin MT, Mazur M, Telser J. Free radicals and antioxidants in normal physiological functions and human disease. Int J Biochem Cell Biol 2007;39(1):44–84.Article PubMed
4. Rani V, Deep G, Singh RK, Palle K, Yadav UC. Oxidative stress and metabolic disorders: pathogenesis and therapeutic strategies. Life Sci 2016;148:183–93.Article PubMed
5. Bryan S, Baregzay B, Spicer D, Singal PK, Khaper N. Redox-inflammatory synergy in the metabolic syndrome. Can J Physiol Pharmacol 2013;91(1):22–30.Article PubMed
6. Finkel T, Holbrook NJ. Oxidants, oxidative stress and the biology of ageing. Nature 2000;408(6809):239–47.Article PubMed PMC PDF
7. Griendling KK, Touyz RM, Zweier JL, Dikalov S, Chilian W, Chen YR, et al. Measurement of reactive oxygen species, reactive nitrogen species, and redox-dependent signaling in the cardiovascular system: a scientific statement from the American Heart Association. Circ Res 2016;119(5):e39–75.Article PubMed PMC
8. Aguilar Diaz De Leon J, Borges CR. Evaluation of oxidative stress in biological samples using the thiobarbituric acid reactive substances assay. J Vis Exp 2020;(159):e61122.Article PubMed PMC
9. Kil HN, Eom SY, Park JD, Kawamoto T, Kim YD, Kim H. A rapid method for estimating the levels of urinary thiobarbituric acid reactive substances for environmental epidemiologic survey. Toxicol Res 2014;30(1):7–11.Article PubMed PMC
10. Marrocco I, Altieri F, Peluso I. Measurement and clinical significance of biomarkers of oxidative stress in humans. Oxid Med Cell Longev 2017;2017:6501046.Article PubMed PMC PDF
11. Il'yasova D, Scarbrough P, Spasojevic I. Urinary biomarkers of oxidative status. Clin Chim Acta 2012;413(19-20):1446–53.Article PubMed PMC
12. Fustinoni S, Rossella F, Campo L, Mercadante R, Bertazzi PA. Urinary BTEX, MTBE and naphthalene as biomarkers to gain environmental exposure profiles of the general population. Sci Total Environ 2010;408(14):2840–9.Article PubMed
13. Sung JH, Oh I, Kim A, Lee J, Sim CS, Yoo C, et al. Environmental and body concentrations of heavy metals at sites near and distant from industrial complexes in Ulsan, Korea. J Korean Med Sci 2018;33(5):e33.Article PubMed PMC PDF
14. National Institute of Environmental Research. Evaluating Exposure to Environmental Pollutants and Health Effects of Inhabitants in Industrial Complex Areas: Ulsan–Onsan, 3rd Stage, 2nd Year. Incheon, Korea: National Institute of Environmental Research; 2019.
15. National Institute of Environmental Research. Evaluating Exposure to Environmental Pollutants and Health Effects of Inhabitants in Industrial Complex Areas: Ulsan–Onsan, 3rd Stage, 3rd Year. Incheon, Korea: National Institute of Environmental Research; 2020.
16. Lee SJ, Cho IG, Lee HY, Ju JT, Shin HJ, Choi SD. Development of a comprehensive air risk index and its application to high spatial-temporal health risk assessment in a large industrial city. Environ Pollut 2025;367:125545.Article PubMed
17. Yasin SA, Salih ZR. Oxidative stress biomarkers in gasoline station workers chronically exposed to heavy metals in Erbil City. J Appl Toxicol 2025;45(11):2400–11.Article PubMed
18. Watanabe S, Li YS, Kawasaki Y, Ootsuyama Y, Kawai K. Health examination results and work environment factors affecting urinary 8-hydroxy-2'-deoxyguanosine levels. J Occup Health 2021;63(1):e12210.Article PubMed PMC PDF
19. Atias D, Ashri S, Goldbourt U, Benyamini Y, Gilad-Bachrach R, Hasin T, et al. Machine learning in epidemiology: an introduction, comparison with traditional methods, and a case study of predicting extreme longevity. Ann Epidemiol 2025;110:23–33.Article PubMed
20. Tibshirani R. Regression shrinkage and selection via the LASSO. J R Stat Soc Series B Methodol 1996;58(1):267–88.Article PDF
21. Cutler DR, Edwards TC Jr, Beard KH, Cutler A, Hess KT, Gibson J, et al. Random forests for classification in ecology. Ecology 2007;88(11):2783–92.Article
22. Feng Y, Yang M, Chen H, Zhang K, Ran F, Chen Z, et al. Synergistic effects of environmental factors on benthic diversity: machine learning analysis. Water Res 2025;282:123789.Article
23. Byun D, Schere KL. Review of the governing equations, computational algorithms, and other components of the Models-3 Community Multiscale Air Quality (CMAQ) modeling system. Appl Mech Rev 2006;59(2):51–77.Article PDF
24. Bang JH, Oh I, Kim S, You S, Kim Y, Kwon HJ, et al. Modeling the effects of pollutant emissions from large industrial complexes on benzene, toluene, and xylene concentrations in urban areas. Environ Health Toxicol 2017;32:e2017022.Article PMC PDF
25. Lumley T, Diehr P, Emerson S, Chen L. The importance of the normality assumption in large public health data sets. Annu Rev Public Health 2002;23:151–69.Article PubMed PMC
26. Kim JH. Multicollinearity and misleading statistical results. Korean J Anesthesiol 2019;72(6):558–69.Article PubMed PMC PDF
27. Greenland S, Daniel R, Pearce N. Outcome modelling strategies in epidemiology: traditional methods and basic alternatives. Int J Epidemiol 2016;45(2):565–75.Article PubMed PMC
28. Chen CH, Yuan TH, Shie RH, Wu KY, Chan CC. Linking sources to early effects by profiling urine metabolome of residents living near oil refineries and coal-fired power plants. Environ Int 2017;102:87–96.Article PubMed
29. Landrigan PJ, Fuller R, Acosta NJ, Adeyi O, Arnold R, Basu NN, et al. The Lancet Commission on pollution and health. Lancet 2018;391(10119):462–512.Article PubMed
30. Seals R, Bartell SM, Steenland K. Accumulation and clearance of perfluorooctanoic acid (PFOA) in current and former residents of an exposed community. Environ Health Perspect 2011;119(1):119–24.Article PubMed PMC PDF
31. Tsai CY, Su CL, Wang YH, Wu SM, Liu WT, Hsu WH, et al. Impact of lifetime air pollution exposure patterns on the risk of chronic disease. Environ Res 2023;229:115957.Article PubMed
32. Rajagopalan S, Brauer M, Bhatnagar A, Bhatt DL, Brook JR, Huang W, et al. Personal-level protective actions against particulate matter air pollution exposure: a scientific statement from the American Heart Association. Circulation 2020;142(23):e411–31.Article PubMed
33. Lunderberg DM, Misztal PK, Liu Y, Arata C, Tian Y, Kristensen K, et al. High-resolution exposure assessment for volatile organic compounds in two California residences. Environ Sci Technol 2021;55(10):6740–51.Article PubMed
34. Greenwald R, Sarnat JA, Fuller CH. The impact of vegetative and solid roadway barriers on particulate matter concentration in urban settings. PLoS One 2024;19(1):e0296885.Article PubMed PMC
35. Blair A, Stewart P, Lubin JH, Forastiere F. Methodological issues regarding confounding and exposure misclassification in epidemiological studies of occupational exposures. Am J Ind Med 2007;50(3):199–207.Article PubMed
36. Baxter LK, Wright RJ, Paciorek CJ, Laden F, Suh HH, Levy JI. Effects of exposure measurement error in the analysis of health effects from traffic-related air pollution. J Expo Sci Environ Epidemiol 2010;20(1):101–11.Article PubMed PDF
37. Xu X, Du H, Lian Z. Discussion on regression analysis with small determination coefficient in human-environment researches. Indoor Air 2022;32(10):e13117.Article PubMed PDF
38. Kim H, Kim H, Lee JT. Spatial variation in lag structure in the short-term effects of air pollution on mortality in seven major South Korean cities, 2006-2013. Environ Int 2019;125:595–605.Article
39. Choi JY, Kim SY, Kim T, Lee C, Kim S, Chung HM. Ambient air pollution and the risk of neurological diseases in residential areas near multi-purposed industrial complexes of Korea: a population-based cohort study. Environ Res 2023;219:115058.Article PubMed
40. Zia A, Farkhondeh T, Pourbagher-Shahri AM, Samarghandian S. The roles of mitochondrial dysfunction and reactive oxygen species in aging and senescence. Curr Mol Med 2022;22(1):37–49.Article PubMed
41. Maldonado E, Morales-Pison S, Urbina F, Solari A. Aging hallmarks and the role of oxidative stress. Antioxidants (Basel) 2023;12(3):651.Article PubMed
42. Vina J, Gambini J, Lopez-Grueso R, Abdelaziz KM, Jove M, Borras C. Females live longer than males: role of oxidative stress. Curr Pharm Des 2011;17(36):3959–65.Article PubMed
43. Signorelli SS, Neri S, Sciacchitano S, Pino LD, Costa MP, Marchese G, et al. Behaviour of some indicators of oxidative stress in postmenopausal and fertile women. Maturitas 2006;53(1):77–82.Article PubMed
44. Khoshakhlagh AH, Ghobakhloo S, Vogel UB, Gruszecka-Kosowska A. Exposure to welding fumes and health effects: a systematic evaluation based on human biomonitoring analyses. Toxicol Lett 2026;418:111861.Article PubMed
45. Khan C, Rehman MY, Malik RN. Metal biomonitoring using fractioned dust to investigate urinary and oxidative stress biomarkers among occupationally exposed chromite mine workers. Environ Sci Pollut Res Int 2022;29(21):31164–79.Article PubMed PDF
46. Emma R, Caruso M, Campagna D, Pulvirenti R, Li Volti G. The impact of tobacco cigarettes, vaping products and tobacco heating products on oxidative stress. Antioxidants (Basel) 2022;11(9):1829.Article PubMed PMC
47. Moraes L, Dries SS, Seibert BS, Linden R, Perassolo MS. Evaluation of oxidative stress markers in ethanol users. Braz J Med Biol Res 2023;56:e12465.Article PubMed PMC
48. Lee E, Hong S, Kim YD, Lee DI, Eom SY. Evaluating the impact of airborne fine particulate matter and heavy metals on oxidative stress via vitamin supplementation. Toxics 2024;12(7):465.Article PubMed PMC
49. Romieu I, Castro-Giner F, Kunzli N, Sunyer J. Air pollution, oxidative stress and dietary supplementation: a review. Eur Respir J 2008;31(1):179–97.Article PubMed
50. Janicka M, Kot-Wasik A, Kot J, Namiesnik J. Isoprostanes-biomarkers of lipid peroxidation: their utility in evaluating oxidative stress and analysis. Int J Mol Sci 2010;11(11):4631–59.Article PubMed PMC

Figure & Data

REFERENCES

Citations

Citations to this article as recorded by

Cite

CITE: export Copy Download Format; Close

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

RIS — For EndNote, ProCite, RefWorks, and most other reference management software
BibTeX — For JabRef, BibDesk, and other BibTeX-specific software

Include:

Citation for the content below
Citation and abstract for the content below

Personal and environmental determinants of oxidative stress among the residents of Ulsan, South Korea: a combined approach of statistical and machine learning models

Korean Journal of Occupational and Environmental Medicine. 2026;e41 Published online September 1, 2026

DOI: https://doi.org/10.35371/aoem.2026.38.e41

XML Download

Personal and environmental determinants of oxidative stress among the residents of Ulsan, South Korea: a combined approach of statistical and machine learning models

Group^a	No. (%)	BTX level, median^b (IQR) (ppb)	Proximity to industrial complex^c (IC)	Potential exposure to HAPs from IC
A	324 (35.64)	1.48 (0.78–2.88)	Districts bordering the IC (0–2 km)	High likelihood of long-term, high-concentration exposure to industrial HAPs
B	322 (35.42)	0.92 (0.83–1.02)	Districts within the expected influence range of the IC (1–3 km)	Moderate average exposure, with possibility of short-term, high-concentration exposure depending on wind conditions and proximity to the IC
C	263 (28.93)	0.72 (0.58–0.82)	Districts beyond the influence range, without large-scale industrial emission sources (>10 km)	Low likelihood of high-concentration exposure

Variable	No. (%) or mean ± SD	Ln_TBARS (mean ± SD)	p-value
Age (years)	61.32 ± 14.13	-	-
Duration of residence (years)	16.96 ± 13.99	-	-
Building age (years)	23.12 ± 16.60	-	-
Sex			<0.001
Male	258 (28.4)	0.514 ± 0.957
Female	651 (71.6)	0.836 ± 0.773
Occupation			<0.001
Professional/manager/office	71 (7.8)	0.301 ± 1.143
Service/sales	128 (14.1)	0.724 ± 0.894
Skilled/manual labor/agriculture	196 (21.6)	0.703 ± 0.904
Homemakers/students/unemployed	514 (56.5)	0.827 ± 0.728
Marital status			<0.001
Married	702 (77.2)	0.691 ± 0.824
Not married	207 (22.8)	0.924 ± 0.875
Current smoking			0.036
Yes	96 (10.6)	0.575 ± 0.811
No	813 (89.4)	0.765 ± 0.843
Residential area^a			<0.001
Group A	324 (35.6)	1.100 ± 0.420
Group B	322 (35.4)	0.845 ± 0.727
Group C	263 (28.9)	0.184 ± 1.050

Independent variable	Reference^a	B	Standard β	95% CI	SE	p-value
(Intercept)	-	−0.465	-	−0.762 to −0.169	0.151	0.002
Group C	Group A	−0.738	−0.395	−0.874 to −0.602	0.069	<0.001
Group B	Group A	−0.187	−0.106	−0.301 to −0.073	0.058	0.001
Age	-	0.019	0.319	0.015 to 0.023	0.002	<0.001
Female sex	Male	0.286	0.152	0.178 to 0.394	0.055	<0.001
Adjacent road lanes: 4-lane	2-lane	0.143	0.080	0.027 to 0.258	0.059	0.016

Variable^a	λ_min^b coefficient (R² = 0.251, n = 8)	λ_1se^c coefficient (R² = 0.199, n = 2)
Residential area (group C)	−0.529	−0.359
Age	0.014	0.009
Female sex	0.154	-
Alcohol consumption (1–2 times/week)	−0.059	-
Adjacent road lanes (6-lane or more)	−0.039	-
Residential area (group B)	−0.017	-
Educational attainment (college graduate or above)	−0.008	-
Duration of residence (years)	0.001	-

Rank	Variable	%IncMSE^a
1	Residential area (group C)	17.07
2	Duration of residence	14.40
3	Age	12.67
4	Marital status (not married)	7.02
5	Female sex	6.93
6	Adjacent road lanes (4-lane)	6.74
7	Income (≥6,000,000 KRW/month)	6.48
8	Residential building type (mid-rise apartment)	6.45
9	Building age	6.35
10	Smoking history (pack-years)	5.08

Item	Stepwise regression	LASSO (λ_min)	Random forest	No. of models^a
Model fit (adjusted R²/CV-R²/Pseudo R²)	0.293	0.251	0.223	-
No. of selected variables	13	8	89 (by importance)	-
Residential area (group C)	✓	✓	✓ (1st)^b	3
Age	✓	✓	✓ (3rd)	3
Female sex	✓	✓	✓ (5th)	3
Adjacent road lanes (4-lane)	✓	-	✓ (6th)	2
Residential area (group B)	✓	✓	-	2
Duration of residence	-	✓	✓ (2nd)	2

Table 1. Classification criteria and potential exposure to industrial HAPs of the three residential groups

HAPs: hazardous air pollutants; BTX: sum of benzene, toluene, and xylene concentrations; IQR: interquartile range; IC: industrial complex.

Groups were classified based on the degree of potential exposure to industrial HAPs, determined jointly by the proximity of the district to large-scale national industrial complexes and by BTX concentrations, which served as reference indicators of industry-derived air pollution;

BTX concentrations were estimated using predictions from the Community Multiscale Air Quality model, a chemical transport model that simulates ambient air pollutant concentrations (reported in a previous study). The original 1-km model outputs were spatially interpolated to a 100-m resolution, and the interpolated BTX concentrations at each participant’s residential location were extracted;

As participants were organized by administrative district before classification, the distances shown represent the range spanned by the constituent districts rather than individually measured distances; the ranges of groups A and B therefore partially overlap. The BTX concentrations are presented as reference indicators of the relative exposure ranking across groups and were not used as absolute exposure measures in the analyses.

Table 2. General characteristics of study participants and mean Ln_TBARS values by residential area (n = 909)

p-values were calculated using independent-samples t-test or one-way analysis of variance.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance value; SD: standard deviation.

Post hoc comparisons using Scheffé’s test: group A > group C; group A > group B; group B > group C.

Table 3. Key results of the stepwise multiple regression analysis for Ln_TBARS

Adjusted R² = 0.293. Full variable results are provided in Supplementary Table 5.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration; B: unstandardized coefficient; CI: confidence interval; SE: standard error.

Reference categories: residential area = group A; road lanes = 2-lane.

Table 4. Results of LASSO regression analysis for Ln_TBARS

LASSO: least absolute shrinkage and selection operator; Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration.

Optimal regularization parameters were determined via 10-fold cross-validation (seed = 42);

λ_min, minimum cross-validation MSE criterion;

λ_1se, most parsimonious model within 1 standard error of the minimum MSE (conservative criterion).

Table 5. Top 10 variables by importance in random forest analysis for Ln_TBARS

Model parameters: ntree = 500, mtry = 29; OOB-MSE = 0.556; Pseudo R² = 0.223.

Ln_TBARS: natural log-transformed thiobarbituric acid reactive substance concentration; KRW: Korean won.

%IncMSE, percentage increase in mean squared error upon permutation of the variable (out-of-bag); higher values indicate greater variable importance.

Table 6. Summary of analysis results: stepwise regression, LASSO regression, and random forest

LASSO: least absolute shrinkage and selection operator; CV-R²: cross-validation-based R²; Pseudo R²: out-of-bag-based R².

No. of models, number of analytical approaches in which the variable was selected. Values are presented for reference and are not directly comparable across methods due to differences in analytical properties;

✓, variable selected by the respective model. Numbers in parentheses indicate rank by %IncMSE in random forest.