Artificial intelligence exceeds humans in epidemiological job coding

Abstract

Background Work circumstances can substantially negatively impact health. To explore this, large occupational cohorts of free-text job descriptions are manually coded and linked to exposure. Although several automatic coding tools have been developed, accurate exposure assessment is only feasible with human intervention. Methods We developed OPERAS, a customizable decision support system for epidemiological job coding. Using 812,522 entries, we developed and tested classification models for the Professions et Catégories Socioprofessionnelles (PCS)2003, Nomenclature d’Activités Française (NAF)2008, International Standard Classifications of Occupation (ISCO)-88, and ISCO-68. Each code comes with an estimated correctness measure to identify instances potentially requiring expert review. Here, OPERAS’ decision support enables an increase in efficiency and accuracy of the coding process through code suggestions. Using the Formaldehyde, Silica, ALOHA, and DOM job-exposure matrices, we assessed the classification models’ exposure assessment accuracy. Results We show that, using expert-coded job descriptions as gold standard, OPERAS realized a 0.66–0.84, 0.62–0.81, 0.60–0.79, and 0.57–0.78 inter-coder reliability (in Cohen’s Kappa) on the first, second, third, and fourth coding levels, respectively. These exceed the respective inter-coder reliability of expert coders ranging 0.59–0.76, 0.56–0.71, 0.46–0.63, 0.40–0.56 on the same levels, enabling a 75.0–98.4% exposure assessment accuracy and an estimated 19.7–55.7% minimum workload reduction. Conclusions OPERAS secures a high degree of accuracy in occupational classification and exposure assessment of free-text job descriptions, substantially reducing workload. As such, OPERAS significantly outperforms both expert coders and other current coding tools. This enables large-scale, efficient, and effective exposure assessment securing healthy work conditions.

Verbal behavior and the future of social science.

“…We are aware of several labs around the world that are currently exploring AI's emerging potential to infer a given person's psychological traits and states from verbal behavioras well as its provided rationale-in ways that are virtually indistinguishable from expert human coders when given the same instructions (see, e.g., Gilardi et al, 2023;Giorgi et al, 2023;Kuzman et al, 2023;Langezaal et al, 2023; see also Chim et al, 2024). Such studies will almost certainly prove symbiotic, with humans providing (and receiving) new insights into gaps in theory, research, and practice when it comes to exploring the complex interplay of personal, social, and environmental factors in human psychology.…”