Advancing occupational exposure models: insights from a case study

Our plain-English summary of Advancing occupational exposure models: insights from a case study, by R. Franken, I. Ottenbros, C. Ge, et al (Annals of Work Exposures and Health, 2026).

The study

• Authors: R. Franken, I. Ottenbros, C. Ge, W. Fransman, T. Krone
• Journal: Annals of Work Exposures and Health
• Published: 2026;70(3)
• Read it: PubMed · Publisher · Free full text

Abstract

Currently used models in occupational exposure assessment are mostly mechanistic-based and their performance varies. The usage of more advanced modeling approaches such as machine learning (ML) or hybrid modeling approaches such as Bayesian Networks could improve accuracy of exposure predictions. Within this study, a case study was conducted where 5 different ML or hybrid modeling approaches were used to redevelop the "asbestos removal exposure assessment tool," an existing mechanistic exposure model, and their performance was compared. Multiple Linear Regression, Random Forest, Gradient Boosting Machines, Bayesian Network, and Neural Network models were developed using the same dataset and model determinants used to develop the mechanistic model. Random Forest and Gradient Boosting Machines performed best with regards to accuracy, followed by Bayesian Network, Multiple Linear Regression, the original model, and Neural Network. Such models show a promise for the development of more accurate models, but their limitations need to be considered before they can be implemented in a regulatory context. For these models, a trade-off exists where transparency in decision-making is traded against accuracy. Hybrid models such as Bayesian Networks might be a solution for this trade-off as expert knowledge and data-driven approaches are combined in a transparent model.

© The Author(s) 2026. Published by Oxford University Press on behalf of the British Occupational Hygiene Society.

In plain English

This study compared computer models estimating airborne asbestos exposure during asbestos removal (abatement). The authors rebuilt an existing model, the asbestos removal exposure assessment tool (AREAT), using five other methods: multiple linear regression, Random Forest, Gradient Boosting Machines, a Bayesian Network and a Neural Network. These were trained on the 370 personal measurements from Dutch abatement tasks that had been used to calibrate AREAT. They used the same inputs, such as chrysotile and amphibole content and ventilation. All six models were then tested against 270 other measurements. Fit statistics were calculated on log-transformed predictions and measurements. Random Forest had the highest correlation with the measurements (0.89, against 0.71 for AREAT). The Bayesian Network had the smallest average absolute error on that log scale. The Neural Network's correlation (0.68) was lower than AREAT's. AREAT's estimates were within a factor of 2 of the measurement 79% of the time. For the other models this ranged from 67% (linear regression) to 85% (Neural Network). The authors concluded that combining expert knowledge with data-driven methods shows promise, but that the limitations of these models need consideration before regulatory use.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Modeling case study: an existing mechanistic asbestos exposure model (AREAT) was redeveloped with five machine learning or hybrid methods, and the models' predictions were compared against a separate validation dataset
• Who was studied: Training dataset of 370 personal asbestos fiber exposure measurements from asbestos abatement tasks in the Netherlands, collected over past decades. A separate validation dataset held 270 personal asbestos exposure measurements from asbestos abatement activities.
• Exposure: Airborne asbestos fibers during asbestos abatement, analysed by SEM/EDXA under ISO 14966 and counted as fibers longer than 5 µm, 0.2 to 3 µm wide. Model inputs included product type, task, state of the material, local control measures, ventilation, room volume, and amphibole and chrysotile content of the material.
• Main results:
  • All fit measures were calculated on log-transformed observed and estimated fiber counts. On the validation data, the original AREAT model had Pearson correlation 0.71, RMSE 2.13, MAE 1.80, bias 0.65 and R2 0.51. A positive bias indicates a tendency to overestimate.
  • Random Forest had Pearson correlation 0.89, RMSE 1.43, MAE 1.15, bias 0.27 and R2 0.79 (log-transformed data). These were the highest correlation and R2 and the lowest RMSE in the table. The authors reported that Random Forest and Gradient Boosting Machines performed best on RMSE.
  • The Bayesian Network had Pearson 0.78, RMSE 1.71, MAE 1.1, bias -0.14 and R2 0.61 (log-transformed data). The authors described it as generally closest to the observed values, but outperformed by Gradient Boosting Machines and Random Forest when outliers are weighted more strongly.
  • On the log-transformed data, Gradient Boosting Machines had Pearson 0.79, RMSE 1.60, MAE 1.16, bias 0.40 and R2 0.62. Multiple linear regression had Pearson 0.74, RMSE 2.04, MAE 1.6, bias -0.15 and R2 0.55. The Neural Network had Pearson 0.68, RMSE 1.90, MAE 1.41, bias -0.14 and R2 0.46.
  • In 95% of estimations, AREAT was off from the observations by a factor of 24 or less. The corresponding factors were 16.8 for linear regression, 14.5 for Random Forest, 24.5 for Gradient Boosting Machines, 76.3 for the Neural Network and 50.1 for the Bayesian Network.
  • Estimates were less than a factor of 2 off the observation 79% of the time for AREAT. The figures were 67% for linear regression, 84% for Random Forest, 84% for Gradient Boosting Machines, 85% for the Neural Network and 81% for the Bayesian Network.
  • The linear regression model linked higher air exchange rates with higher fiber concentrations. The authors suggested this may be because higher ventilation rates typically indicate work inside enclosed containments. The Random Forest, Gradient Boosting Machines and Neural Network models identified air exchange rate as the most important predictor.
• The authors' conclusion: The authors concluded that the case study shows promise for advanced statistical techniques that combine expert knowledge with data-driven methods to improve existing or develop new occupational exposure models that are explainable and more accurate.
• Limitations the authors state:
  • Several models do not provide a confidence interval, so confidence intervals could not be compared across methods.
  • The limitations of these models need to be considered before they can be used in a regulatory context. The authors also describe a trade-off between transparency and accuracy for the black-box methods.
  • Neural Networks need a large, high-quality training dataset, which the authors note is hard to obtain in exposure measurement.

Topics

Medical terms in this study, explained in our Lexicon: Occupational Exposure.

Related studies

• Advancing Treatment Outcomes for Peritoneal Surface Malignancies in Low- and Middle-Income Countries: Insights from the First Multicenter Study in North Africa — A. Souadka, H. Habbat, et al. (2025)
• Predicting Mesothelioma Using Artificial Intelligence: A Scoping Review of Common Models and Applications — M. Ram, M.R. Afrash, et al. (2025)
• Prediction of Radiation Pneumonitis With Machine Learning in Stage III Lung Cancer: A Pilot Study — M. Yakar, D. Etiz, et al. (2021)
• Narciclasine as a Novel Treatment for Lung Cancer and Malignant Pleural Mesothelioma: Insights from 3D Tumor Spheroid Models — S. Jiang, S. Berkcan, et al. (2025)
• Cancer Risk Prediction Using Machine Learning for Supporting Early Cancer Diagnosis in Symptomatic Patients: A Systematic Review of Model Types — F. Pennisi, S. Borlini, et al. (2025)
• On the proportional hazards model for occupational and environmental case-control analyses — H. Gauvin, A. Lacourt, et al. (2013)

Browse more research in Annals of Work Exposures and Health, or return to medical research.