Work

Predicting worker-level particulate matter exposure categories at a renovation site using multi-source monitoring and machine learning

Year: 2026

Type: article

Abstract: Construction activities are a major source of airborne particulate matter (PM), exposing workers to elevated concentrations of PM10, PM2.5, and PM1 that pose serious health risks. However, current monitoring practices focus primarily on ambient or fixed-site concentrations and often fail to capture the dynamic, task- and location-dependent nature of worker exposure. This study developed a high-resolution, multi-source dataset at an active renovation site by combining helmet-mounted sensors, fixed-point monitors, meteorological measurements, and detailed worker activity and location records. Using a machine-learning-based prediction framework, worker PM measurements were assigned to five ordered exposure categories, and four machine-learning models—deep neural network (DNN), random forest (RF), XGBoost, and LightGBM—were evaluated under the observed imbalanced category distributions. Results show that DNN provided the most balanced overall performance across the nine worker–PM prediction tasks, with its clearest and most consistent advantage observed for PM10. For PM10, DNN achieved Accuracy values of 0.44–0.51 and Macro F1 values of 0.40–0.46, compared with 0.17–0.22 and below 0.13, respectively, for the tree-based models. For PM1, high Accuracy coexisted with low Macro-level performance across all four models, reflecting the dominance of the Good category. Feature-importance analyses using PFI and SHAP consistently identified monitored worker activity and meteorological conditions as leading contributors to the model predictions, although their relative importance varied among workers. Fixed outdoor PM measurements generally made the smallest contribution. This study provides field-based evidence that high-frequency, multi-source data can support worker-level PM exposure classification and establishes a methodological foundation for the development of adaptive, worker-centred exposure monitoring and management systems.

Source: Journal of Cleaner Production

DOI: 10.1016/j.jclepro.2026.149221

Authors: Xingyue Fang, Yanquan Zhang, Ruidong Chang, Lishan Yang, Wei Emma Zhang, Jian Zuo

Institutions: Adelaide University, The University of Adelaide

Language: English

Topic: Air Quality and Health Impacts

Subfield: Health, Toxicology and Mutagenesis

Field: Environmental Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: hybrid

Estimated APC paid: $4,660