Large language model-assisted study examines conflicts of interest in clinical trials published in clinical oncology journal

Conflicts of interest (COIs) between clinical trial investigators and the biopharmaceutical companies participating in the trials could later present bias in research when the findings are published by clinical trials’ contributing authors. Pedro Silberman, PhD, a cancer researcher and third-year medical student at The Ohio State University College of Medicine, says the need to analyze trends of bias is necessary to foster a better understanding of the clinical trial landscape. This formed the basis of a recent study, “Trends of Authors’ Conflicts of Interest in Clinical Trials Published in the Journal of Clinical Oncology: A Large Language Model-Assisted Longitudinal Study,” led by Jiasheng Wang, MD, assistant professor of Internal Medicine in the Ohio State College of Medicine. Dr. Wang’s experience using machine learning methods in basic and clinical hematology to develop treatment protocols was instrumental in this research.

They used large language models (LLMs) to parse large amounts of written text to analyze the prevalence and trends of COIs in oncology clinical trial studies published in the Journal of Oncology. Details of the study and findings include:

• Their use of OpenAI’s GPT-4o LLMs to search published trials between 2010-2025, grouping them by country of origin into three periods: 2010-2015, 2015-2020 and 2020-2025. Combed data included study name, authors, associated company if applicable, abstract and identified COIs.
• Analysis by GPT-4o demonstrated 95% accuracy when compared to select, manually curated data.
• 72% of publications with a medial product had at least one author match a COI with the biopharmaceutical company working with the study.
• COI prevalence increased from 70% in 2010-2015 to 77% in 2015-2020 and then decreased to 72% between 2020-2025.
• U.S.-led studies had a significantly higher COI prevalence than those from other geographical regions.
• Validation that LLMs such as GPT-4o can reliably process large amounts of scientific studies over long time periods and are useful in monitoring trends in biomedical research and COIs.

This study reveals that the LLM-based method provides an efficient solution for COI monitoring, promoting transparency in biomedical research.