Sarah Daniels

It has until recently been difficult to explore how workers might behave with degrees of freedom operating within constraints. With Generative AI, it is now possible to treat a dataset as representative of a set of constraints (for example, socioeconomic group, geography, job exposures and activities, age), and then to construct simulated workers which behave with degrees of freedom producing health outcomes. As an epidemiological analysis, this reveals varied distributions of possible paths of behaviour of simulated workers within different kinds of constraint: some constraints produce little variation in simulated worker action, whereas other constraints produce much greater variation. Here we present analysis of this variation using geographical, job exposure matrices, and socioeconomic variables, examining how adverse health outcomes are weighted towards those conditions which provide the least freedom of worker choice. Drawing on the 36-year longitudinal THOR dataset, we show how similar constraints (particularly those around place and socioeconomics) persist across generations, producing similar patterns of restricted freedom. 

Our AI-driven Digital Worker Twins are created using generative AI with data from 24000 THORGP cases (alongside geographical and socioeconomic data). They present opportunities to target specific constraints which can be shown to produce poor health outcomes, and to explore ways of increasing degrees of freedom for workers. This use of health surveillance data as AI training allows for new kinds of interventions to be explored for possible mitigation of occupational health risks.  As an example of our approach, a key area for investigation concerns the rising prevalence of occupational stress. Using our approach, the degrees of freedom of action appears to be a key indicator of risk of stress, where geographical and socio-economic constraints impinge on worker behaviour. This suggests that constraint modelling using Digital Worker Twins might offer new insights into causes of occupational stress beyond traditional exposure-based modelling.

Objective: Manually coding job descriptions is time-consuming, expensive, and requires expert knowledge. To mitigate these issues, a Decision Support System (DSS) can be used. A DSS partly automates the coding process providing users with suggestions that can be manually corrected, improving both efficiency and reliability. With OPERAS, we present a customizable DSS that: 1) integrates highly accurate classification models for automatic coding, 2) provides confidence scores to flag codes needing expert review, and 3) provides an interface for manual correction. Moreover, its effectiveness with expert coders is evaluated in a real-world setting. 

Methods and Materials: Using Machine Learning and Natural Language Processing techniques, OPERAS’ classification models were developed and evaluated with 812,522 expert-coded job descriptions for the Professions et Catégories Socioprofessionnelles (PCS)2003, Nomenclature d’Activités Française (NAF)2008, International Standard Classifications of Occupation (ISCO)-88, and ISCO-68. Subsequently, 5 expert coders proficient in PCS2003 and NAF2008 tested OPERAS in a real-world setting. Each expert coded 2 subsets of job descriptions from the CONSTANCES cohort, both manually and using OPERAS. Then, we assessed coding time, inter-coder reliability, and usability. 

Results: OPERAS’ classification models showed inter-coder reliability ranging 0.57-0.78 (Cohen’s Kappa) on the least aggregated coding levels, yielding an exposure assessment accuracy of 75.0% to 98.4%. Using OPERAS for job coding showed to be >1.5x faster than manual coding, with median coding times of respectively 38 and 60.8 seconds per case. Inter-coder reliability ranged from 0.61–0.70 with OPERAS compared to 0.56–0.61 manually for PCS; for NAF, this was 0.38–0.61 with the DSS versus 0.34–0.61 manually. 

Conclusion: OPERAS showed high accuracy in occupational classification and exposure assessment, while being reliable, and significantly reducing workload. Consequently, OPERAS facilitates large-scale, harmonized job coding in occupational health research. 

Samuel Keer, Thomas Culling, Judge Crozier, Dave McLean, Tracey Whaanga, Soo Cheng, Jeroen Douwes

Objective: Since the introduction of engineered stone for kitchen benchtops, silicosis cases among fabricators have been reported in several countries. This is the first study to describe RCS exposure levels in NZ engineered stone workers and evaluate the effectiveness of interventions by comparing pre- and post-intervention exposures.

Material and Methods: We compared pre- and post-intervention RCS levels in six engineered stone fabrication workshops. Exposure monitoring involved personal full-shift respirable dust sampling using X-ray diffraction to determine RCS. Most baseline measurements were collected in 2019, prior to the intervention date (which ranged from 2019-2021), which involved the adoption of wet methods, automation, and regulator inspection visits. We collected information on interventions implemented after the baseline visit, identified opportunities for improvements, and repeated exposure monitoring in 2022 and 2024. Exposure ratios, representing the proportional difference in RCS levels relative to baseline, were calculated using multi-level linear regression.

Results: At baseline, the geometric mean (GM) for RCS for all samples (n=90) was 33.8µg/m3 (GSD:2.8) with 59% of samples above the NZ Workplace Exposure Standard (NZ-WES) of 25µg/m3. The highest levels were for benchtop fabricators: GM: 57.3µg/m3 (GSD:2.9) with 87% of samples above the NZ-WES. In 2024, the post-intervention results indicated a substantial reduction in RCS levels of 70-75% for all roles and 77% overall. The ER for pre vs post was 0.39 (0.30-0.50; p<.0001). The GM was 7.1µg/m3 (GSD:2.5) for all samples (n=87) and 10µg/m3 (GSD:2.7) for fabricators. However, 5% and 7% of samples, respectively, were still above the NZ-WES.

Conclusion: The pre-intervention results indicate that workers were frequently exposed to high RCS levels. Following industry-wide and regulator efforts to reduce exposure, levels reduced significantly but 7% were still above exposure limits. Thus, although relatively effective, further interventions are required to ensure that exposures will be below the WES for all workers.

Federico Fassio, Francesca Sellaro, Simona Villani, Enrico Oddone

Objective: To evaluate the relationship between environmental temperature, air pollution, and the frequency of work-related injury visits to the emergency department (ED) in Pavia (Italy) over a nine-year period, with a focus on thermal stress (WBGT index) and air pollution.

Material and Methods: This time-series study analyzed ED admissions due to occupational injuries from January 2014 to December 2022, excluding events related to violence, traffic, or animal/insect encounters. Meteorological data and pollutant concentrations (PM10, PM2.5, NO2, SO2, O3) were retrieved from local monitoring stations. Missing environmental data was imputed using Kalman smoothing. The Wet Bulb Globe Temperature (WBGT) was calculated and included as exposure variable. Distributed lag nonlinear models with cross-basis matrices were applied to account for delayed and non-linear effects, using negative binomial regression adjusted for seasonality, holidays, and long-term temporal trends. The goodness of fit was evaluated by means of Akaike information criterion (AIC).

Results: A total of 13,841 ED visits were recorded over 3,287 days. WBGT showed a significant positive association with injury rates especially at lower and higher values (at 0.5°C, compared to the median value 12.5°C:RR=1.17; 95% CI=1.09–1.24. At 25°C:RR=1.09; 95%CI=1.02; 1.16). Among air pollutants, NO2 was the only compound significantly associated with increased injury risk (at 50μg/m3, compared to the median value 24.3μg/m3:RR=1.18; 95% CI=1.07–1.30). In general, PM10, PM2.5, SO2, and O3 did not exhibit significant and consistent associations. Inclusion of WBGT covariate in all models with air pollutants improved model fit, as indicated by better goodness of fit (lower AIC values).

Conclusion: Thermal stress, as measured by WBGT, and ambient NO2 concentrations appear to be significantly associated with increased risk of occupational injuries. These findings underscore the need for integrating climate and air quality monitoring into occupational health policies, particularly in the context of climate change and increasing urban heat exposure.