A machine learning approach to improve prediction in chemical exposure risk assessment

Exposure models play a crucial role in predicting chemical exposure in workplaces, offering an essential alternative to measurements, which are resource-intensive and time-consuming and sometimes not possible. Despite their widespread use and continuous development, significant challenges persist, including variability in predictions, limited model updates, and difficulties in accessing the required input data. In this study, we investigate how modern machine learning techniques can contribute to the improvement of exposure models by addressing these limitations. To overcome the frequent lack of data, we explore the use of synthetic datasets generated through existing exposure models. This approach allows for the study of relationships among variables, pattern extraction, and dimensionality reduction, facilitating model refinement without the need for complex model integration. When measurement data are available, supervised machine learning methods can be directly applied. As an example, deep neural networks trained on exposure data from a reference measurement dataset demonstrate substantial improvements compared to the Advanced Reach Tool (ART) predictions. However, this approach restricts the model to the determinants and exposure scenarios present in a single database, limiting its generalizability. To overcome this limitation, we propose a pipeline that integrates information from multiple sources into a unified predictive model. By learning a shared internal representation across datasets, we enable training a model that leverages all available data. Preliminary evaluations show successful integration, with predictive performance comparable to networks trained on individual datasets, and ongoing optimization to enhance accuracy through the unified representation. These preliminary findings highlight the potential of modern data driven approaches to enhance traditional exposure modelling and to support the development of more robust, accurate, and flexible predictive tools.