Published 27 June 2026 • By Dr. Megan Tranter
For most of its history, the practice of industrial hygiene has run on sparse data. A handful of personal air samples, collected over a shift or two, were stretched to represent the exposures of an entire similar-exposure group across months of variable production. That scarcity shaped everything from sampling strategy to the statistical conventions we use to declare a workplace acceptable. Machine learning is now changing the arithmetic. By extracting structure from large, messy, high-frequency data streams, these methods promise exposure estimates with spatial and temporal resolution that traditional compliance sampling cannot match. They also introduce new ways to be confidently wrong. This post looks at where the science actually stands.
What You’ll Learn
- How supervised and unsupervised machine learning are applied to exposure modeling and prediction.
- How computer vision is used to monitor PPE compliance and at-risk behavior on worksites.
- How wearable and low-cost sensor networks generate the data these models depend on.
- Why data quality, calibration drift, and algorithmic bias can quietly undermine an otherwise sophisticated model.
- What the recent peer-reviewed literature demonstrates, and what it does not yet support.
Introduction
Artificial intelligence is an umbrella term for computational systems that perform tasks normally requiring human judgment, while machine learning is the subset of those systems that improve their performance by learning patterns from data rather than following explicitly programmed rules. In industrial hygiene, the relevant techniques range from regression-based predictive models and random forests to convolutional neural networks for image analysis. None of these tools replaces professional judgment or the hierarchy of controls. What they offer is a way to interpolate, predict, and flag exposures across the gaps that conventional air sampling strategies inevitably leave behind.
Machine Learning for Exposure Modeling and Prediction
The most mature application is statistical exposure prediction. Determinants-of-exposure models have existed for decades, but machine learning broadens the range of inputs and relationships a model can capture. Random forests, gradient-boosted trees, and neural networks can ingest production rate, ventilation status, task code, temperature, humidity, and time of day, then predict airborne concentration at locations and moments that were never directly sampled. Patton and colleagues (2022) demonstrated this concretely, pairing a low-cost sensor network with probabilistic machine learning to produce hazard maps across a workplace. Their estimates improved on traditional compliance sampling and aligned with the AIHA exposure-rating framework, allowing limited hygiene resources to be directed to areas with the highest modeled risk. This connects directly to structured decision approaches such as control banding, where a model’s predicted exposure category can drive control selection.
A second use is to fill and smooth sparse datasets. Bayesian and probabilistic models quantify uncertainty rather than producing a single point estimate, which matters because an industrial hygienist needs to know not just the predicted 95th percentile of exposure but the confidence around it. The honest framing is that these models are interpolation and extrapolation engines: they are only as trustworthy as the determinants they were trained on, and they extrapolate poorly to processes, materials, or conditions absent from the training data.
Computer Vision for PPE and Behavior
Computer vision applies deep neural networks to images and video so that a system can detect whether a worker is wearing a hard hat, respirator, hearing protection, or high-visibility clothing in real time. Object-detection architectures, such as the YOLO family, have been trained on construction imagery to flag non-compliance, and recent work has developed monitoring frameworks intended to integrate directly into a site safety workflow (Lee et al., 2023; Ferdous and Ahsan, 2022). The appeal is continuous observation: a camera does not get fatigued, and it can audit thousands of person-hours that a walkthrough never reaches.
The limitations are equally real. Detection accuracy degrades under poor lighting, occlusion, unusual camera angles, and PPE styles not present in the training set. A model that confirms a respirator is present cannot confirm that the cartridge is correct, that it is donned with a proper seal, or that the wearer has been fit-tested. Vision systems also raise legitimate worker-privacy and surveillance concerns that must be addressed transparently rather than treated as an afterthought. Used well, computer vision is a leading-indicator tool that surfaces lapses for human follow-up, not a substitute for a respiratory protection program.
Wearables and Sensor Networks
Machine learning is only as good as its data, and wearables are increasingly the source. Low-cost direct-reading sensors for particulate matter, gases, noise, and physiological strain can be worn or distributed across a facility to generate the high-frequency streams that models require. Wearable photometers and electrochemical sensors, paired with location data, let a model reconstruct an exposure map rather than a single time-weighted average. These same data streams feed into the broader exposome approach to cumulative exposure.
The catch is calibration. Low-cost sensors drift, respond to interferents, and are sensitive to humidity and temperature, so raw readings cannot be treated as reference-grade. Patton and colleagues (2022) addressed this explicitly, building drift management and calibration into their pipeline rather than assuming sensor output was accurate. The professional lesson is that a sensor network is an instrument system requiring validation against reference methods, not a set of plug-and-play gadgets.
Data Quality, Bias, and Validation
Every machine learning result inherits the strengths and the flaws of its training data. If sampling historically over-represented certain shifts, tasks, or worker groups, a model trained on that record will reproduce and even amplify those gaps, systematically underestimating exposures for the people who were never measured. This is algorithmic bias in a safety-critical setting, and it disproportionately affects the most vulnerable workers. Models can also overfit, learning quirks of the training facility that do not generalize, and they can present spurious precision through a confident-looking number with no honest uncertainty attached.
Sound practice mirrors good exposure science. Training and validation data must be separated so performance is judged on data the model has never seen. Predictions should be checked against reference-method sampling, not against other models. Uncertainty should be reported, ideally as a distribution rather than a point estimate. And the determinants driving a prediction should be interpretable enough that a hygienist can sanity-check them against process knowledge. A model that cannot explain why it predicts a high exposure is difficult to defend in an enforcement or litigation context. Machine learning augments the professional; it does not absolve the professional of judgment.
Summary
Machine learning provides industrial hygiene tools to predict exposures where sampling is sparse, to continuously monitor PPE compliance, and to make sense of high-frequency wearable data. The peer-reviewed evidence supports real gains in spatial and temporal resolution, particularly when sensor calibration and uncertainty are handled rigorously. The same evidence warns that biased training data, calibration drift, and overconfident outputs can produce results that appear authoritative while misrepresenting the true risk. These are decision-support tools that strengthen, but never replace, professional judgment and the hierarchy of controls.
Helpful Resources
- NIOSH: National Institute for Occupational Safety and Health
- American Industrial Hygiene Association (AIHA)
- Occupational Safety and Health Administration (OSHA)
- Related reading: Air Sampling Strategies and AIHA Exposure Assessment and Control Banding.
Bibliography
American Industrial Hygiene Association. (2015). A strategy for assessing and managing occupational exposures (4th ed.). AIHA Press.
Ferdous, M., & Ahsan, S. M. M. (2022). PPE detector: A YOLO-based architecture to detect personal protective equipment (PPE) for construction sites. PeerJ Computer Science, 8, e999. https://doi.org/10.7717/peerj-cs.999
Goede, H., Kuijpers, E., Krone, T., le Feber, M., Franken, R., Fransman, W., Duyzer, J., & Pronk, A. (2021). Future prospects of occupational exposure modelling of substances in the context of time-resolved sensor data. Annals of Work Exposures and Health, 65(3), 246-254. https://doi.org/10.1093/annweh/wxaa102
Lee, Y.-R., Jung, S.-H., Kang, K.-S., Ryu, H.-C., & Ryu, H.-G. (2023). Deep learning-based framework for monitoring wearing personal protective equipment on construction sites. Journal of Computational Design and Engineering, 10(2), 905-917. https://doi.org/10.1093/jcde/qwad019
National Institute for Occupational Safety and Health. (2023). Artificial intelligence and the future of work. U.S. Department of Health and Human Services, Centers for Disease Control and Prevention. https://www.cdc.gov/niosh/
Patton, A. N., Medvedovsky, K., Zuidema, C., Peters, T. M., & Koehler, K. (2022). Probabilistic machine learning with low-cost sensor networks for occupational exposure assessment and industrial hygiene decision making. Annals of Work Exposures and Health, 66(5), 580-590. https://doi.org/10.1093/annweh/wxab105
Peters, T. M., Anthony, T. R., Taylor, C., Altmaier, R., Anderson, K., & O’Shaughnessy, P. T. (2012). Distribution of particle and gas concentrations in swine gestation confined animal feeding operations. The Annals of Occupational Hygiene, 56(9), 1080-1090. https://doi.org/10.1093/annhyg/mes050
Sarkar, S., & Maiti, J. (2020). Machine learning in occupational accident analysis: A review using science mapping approach with citation network analysis. Safety Science, 131, 104900. https://doi.org/10.1016/j.ssci.2020.104900
Tixier, A. J.-P., Hallowell, M. R., Rajagopalan, B., & Bowman, D. (2016). Application of machine learning to construction injury prediction. Automation in Construction, 69, 102-114. https://doi.org/10.1016/j.autcon.2016.05.016
Zuidema, C., Schumacher, C. S., Austin, E., Carvlin, G., Larson, T. V., Spalt, E. W., Zusman, M., Gassett, A. J., Seto, E., Kaufman, J. D., & Sheppard, L. (2021). Deployment, calibration, and cross-validation of low-cost electrochemical sensors for carbon monoxide, nitrogen oxides, and ozone for an epidemiological study. Sensors, 21(12), 4214. https://doi.org/10.3390/s21124214
Zuidema, C., Stebounova, L. V., Sousan, S., Gray, A., Stroh, O., Thomas, G., Peters, T., & Koehler, K. (2020). Estimating personal exposures from a multi-hazard sensor network. Journal of Exposure Science & Environmental Epidemiology, 30(6), 1013-1022. https://doi.org/10.1038/s41370-019-0146-1