AI-Induced Occupational Health Assessment

While work plays a crucial role in our well-being, it also exposes us to various health risks. By linking subjects' job histories to exposure assessment tools (i.e., Job-Exposure Matrices, JEMs), large-scale cohort and case-control studies assess risks associated with jobs. Before JEMs can be applied, free-text job descriptions must be standardized, using occupational classification systems. This process, usually performed manually, is time-consuming, expensive, and requires specialized knowledge. To address these limitations, (semi-)automatic coding and Decision Support Systems (DSS) have been developed. These systems utilize string-similarity-based, machine-learning or hybrid architectures. Although some fully automatic coding systems approach or even match expert performance, their classification accuracy does not generalize well: it decreases when applied to out-of-distribution data, limiting their real-world applicability. To enable expert correction, which crucially improves the coding process' reliability, DSS are used. Pre-trained on vast amounts of text data, Large Language Models (LLM) could improve the (semi-)automatic or DSS' coding process, improving accuracy and generalizability. However, LLM's application in automatic occupational coding is unexplored. This chapter provides a comprehensive overview of occupational health assessment, focusing on the development and use of JEMs, the challenges of standardizing occupational information, and the current state-of-the-art in Automatic Occupational Coding (AOC). Subsequently, we explore the background of LLM and their potential applications in this field. We conclude with highlighting challenges and give an outlook for AOC.