102 Enhancing the coverage of a multi-agent exposure assessment tool through the modelling of over 100,000 measurements

Introduction The INRS developed an online tool (https://www.inrs.fr/outil110) that provides descriptive statistics of concentrations for work situations defined by users using up to ten exposure variables. Statistics are calculated from measurements collected in French workplaces and recorded in the Colchic database. However, a given work situation may have only few measurements available for reporting reliable descriptive statistics. We explored the use of mixed-effects models to address this issue. Methodology We selected individual measurements collected between 2010- 2022. We defined a core model that contained all variables available in the tool: industry, occupation, task, and product associated with exposure (random effects), number of workers, general ventilation, local exhaust ventilation, exposure frequency, process, and sampling duration (fixed effects). We modeled each agent separately (minimum 100 measurements) after excluding highly correlated variables. Models were cross-validated and predicted concentrations were compared to those recorded in Colchic from January to July 2023 (minimum 10 measurements per agent). Results We fitted models to 134 agents with 101 to 11,819 measurements (median 405). The median cross-validation R2 across agents was 0.45 (interquartile interval 0.35-0.56). Predictive performance for 2023 concentrations by agent (n=48) was lower (median R2 0.10, interquartile interval 0.02-0.22; median relative bias -15%); coverage of the 80% prediction intervals was adequate (median 79%, interquartile interval 69-88%). Conclusions While the cross-validation showed promising results, the performance of the models to predict new concentrations was less satisfactory. Factors associated with poorer predictive performance should be investigated and remediated prior to integrating these models in the tool.