Large Language Models for Data Extraction in Toxicology: Implications and Lessons Learned from the Clinical Evidence Domain

A living review of automated data extraction methods for clinical systematic reviews identified 76 papers up to the end of 2022 describing unique extraction algorithms and assessed their methods and quality of reporting. The current in-progress update for 2024 observed a rapid increase in papers; it now includes 117 papers, of which 17 employ Large Language Models (LLMs) to automate data extraction from randomized controlled trials. In this commentary, we describe findings from the analysis of these LLM automation papers and discuss parallels to the field of automated data extraction in toxicology. We describe the evaluation strategies and reporting observed in LLM publications and focus on inconsistencies and potential pitfalls that may bias how the performance of LLMs is perceived by those likely to apply them to support systematic reviews. We then discuss potential applications of LLMs in evidence mapping and good practice in the reporting of LLM automation methods – based on a checklist and guideline developed during the 2023 Evidence Synthesis hackathon in Newcastle (UK).