Large language model-based screening of substances and their composition from safety data sheets for high-resolution chemical exposure assessment

BACKGROUND: SDSs provide information on chemical substances in professional and consumer products. Large language models (LLMs) offer a rapid approach to screening chemical information from SDSs. OBJECTIVE: This study aimed to validate the performance of LLMs in accurately extracting substance information from SDSs of products. METHODS: Chemical information was extracted from the SDSs of cleaning products using the LLMs ChatGPT-4o and Gemini 2.5 Pro. The performance of the LLMs was evaluated against manually extracted data using precision, sensitivity, and F1 score. RESULTS: A total of 301 substance-composition combinations across 59 products were included in the validation. The Gemini 2.5 Pro model showed a higher F1 score (1.00) than ChatGPT-4o (0.94). SIGNIFICANCE: LLMs enable high-throughput chemical information extraction from SDSs, reducing the burden of manual screening and supporting large-scale combined-exposure assessments. IMPACT STATEMENT: The accurate identification of chemical substances in professional and consumer products is a major challenge in assessing complex chemical exposures in exposure science and epidemiology. This study validated the use of Large Language Models (LLMs) as a novel, rapid, and highly accurate method of extracting high-throughput chemical data from multilingual Safety Data Sheets. Our findings illustrated that LLMs could be effectively used to overcome the labor-intensive and time-consuming limitations of manual data screening. This approach allows high-throughput analysis, enabling comprehensive combined-exposure assessments in large-scale epidemiological and exposure assessment studies. Additionally, as LLMs can extract data in different languages, they can facilitate international research collaboration.