Analysis of conflicts of interest in clinical trials published in JCO: A large language model-assisted longitudinal study.

Abstract

e23042 Background: Conflicts of interest (COIs) in oncology trials, particularly those arising from industry funding and author affiliations, may influence the objectivity of clinical trial findings. The manual identification of medical products and corresponding pharmaceutical companies, as well as author COIs from numerous publications is a time-consuming process. Large language models (LLMs) offer a promising solution to streamline this process by automating this process, facilitating comprehensive analyses of COI trends in oncology. Methods: We systematically collected abstracts and COI sections from clinical trials published in the Journal of Clinical Oncology (JCO) from Jan 2010 to Jan 2025. With the assistance of OpenAI GPT-4o and human oversight, we identified the primary medical product investigated in each trial along with its corresponding pharmaceutical company. Various types of COIs were analyzed. To assess temporal trends, the studies were categorized into three time periods: 2010–2015, 2015–2020, and 2020–2025. Results: A total of 2,219 clinical trials were identified. We found that 72.6% of the studies had at least one author with a COI related to the associated pharmaceutical company. Over the three study periods, COI rates were 70.0%, 77.0%, and 72.0%, respectively. Employment-related COIs remained relatively stable over time (31.3%, 38.3%, and 36.2%), as did consulting/advisory roles (52.5%, 61.5%, and 59.1%) and honoraria (43.8%, 49.4%, and 48.8%). US-led studies accounted for about half of the trials and were significantly more likely to report COIs with the associated company compared to studies from other regions (77.6 vs. 67.4%, p = 0.03). The proportion of US-led studies with employment-related COIs was 32.8%, 44.6%, and 41.2% across the three periods, while consulting/advisory COIs were reported at 54.0%, 65.8%, and 65.8%. Across all time periods, 61.9% of studies had a first or last author with a COI related to the associated company. Consulting/advisory COIs among first or last authors increased over time, from 18.4% in 2010–2015 to 41.3% in 2015–2020 and 46.4% in 2020–2025. Conclusions: LLMs can streamline clinical research by analyzing large text datasets. COIs with pharmaceutical companies are common, with one-third of studies involving authors employed by the company and half with authors in advisory roles. These trends have been consistent for 15 years, with US-led studies reporting more COIs. COIs are particularly prevalent among first or last authors. Prevalence of COIs in clinical trials published in JCO (2010–2025). All periods, %(n=2,219) 2010-2015, %(n=890) 2015-2020, %(n=605) 2020-2025, %(n=724) All Types of COIs 72.6 70.0 77.0 72.0 Employment 35.0 31.3 38.8 36.2 Advisory/Consultant 57.1 52.5 61.5 59.1 Speaker’s Bureau 28.7 N.A. 26.0 30.9 Honoraria 47.0 43.8 49.4 48.8