8274418 SOCcerNET: computer-based occupation coding that facilitates re-coding occupations to a unified system when pooling studies

Objective Pooling studies increases the study power to evaluate occupational risk factors, but this often requires recoding them into a unified occupation coding system. Administrative crosswalks between classification systems may not directly exist and often have one-to-many relationships between codes. In this study, we describe a novel tool, SOCcerNET, that can use both the original free-text job description and occupation codes from other systems or versions to code jobs into US SOC2010.Methods SOCcerNET uses the free-text job title and job tasks and, if available, any previously assigned occupation code to classify English-language job descriptions to US SOC2010. Currently, SOCcerNET can utilize previously assigned US SOC1980 codes, Canadian NOC2011 and NOC2016 codes, and ISCO1988 codes. SOCcerNET incorporates administrative crosswalks and uses small language models that first converts (embeds) the free-text information to numbers; it then classifies using a dense classification neural network and provides a score for each of the 840 codes. We validated SOCcerNET using 11,943 jobs coded to SOC1980 (Set 1), 428 jobs coded to NOC2011 (Set 2), and 1500 jobs coded to ISCO1988 (Set 3); all sets were expert-coded to SOC2010.Results Compared to using only job title and task, incorporating the previous occupation coding improved SOCcerNET’s agreement (based on highest scoring code) with expert codes from 56.3% to 69.9% for Set 1, from 68.8% to 74.4% for Set 2, and from 44.7% to 46.5% for Set 3.Conclusion Using previously assigned occupation codes improved the ability of SOCcerNET to correctly identify the expert-assigned code versus using only the free-text responses, though the improvement was better for SOC1980 and NOC2011/2016 than for ISCO1988. In addition, the SOCcerNET score can help triage jobs for expert review. With additional training data, SOCcerNET can be extended to other coding systems.Funding This work is funded by the Intramural Research Program of the US National Cancer Institute, Division of Cancer Epidemiology and Genetics.