Scholarlink

<p>It is difficult to ensure clinical trial outcomes are defined completely in registrations, and to check outcome changes between registration and posting results. We developed a large language model (LLM)-based approach to evaluate outcome definitions and changes on ClinicalTrials.gov. Our LLM-based approach accurately identified incomplete outcomes in prospective trial registrations (sensitivity, 0.91 [95% confidence interval {CI}, 0.87–0.94]; positive predictive value [PPV], 0.98 [95% CI, 0.97–1.00]). Comparing prospective registrations with posted results, it correctly identified 96.1% of missing and 97.9% of added outcomes. It identified all outcomes with a priority change (e.g., from primary to secondary) and 99.2% without one. Sensitivity and PPV for identifying any outcome change were 0.97 (95% CI, 0.95–0.99) and 0.95 (95% CI, 0.89–0.98), respectively. Estimated costs per trial were $0.13 for o3-mini, $0.30 for GPT-4o, and $1.80 for o1. The accuracy of our approach suggests it could be used to improve registration quality and to detect outcome changes at large scale and at low cost.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An automated approach to improve clinical trial registration and to identify outcome changes on ClinicalTrials.gov

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Xiangji Ying,
  • Kiran Ninan,
  • Jean-Pierre Oberste,
  • Colby J. Vorland,
  • Tianjing Li,
  • Andrew W. Brown,
  • Joe D. Menke,
  • Riaz Qureshi,
  • Nicholas J. DeVito,
  • Matthew J. Page,
  • Joanne E. McKenzie,
  • Ian J. Saldanha,
  • Sirui Zhang,
  • Nancy J. Butcher,
  • Martin Offringa,
  • Jamie Cummins,
  • Halil Kilicoglu,
  • Evan Mayo-Wilson

摘要

It is difficult to ensure clinical trial outcomes are defined completely in registrations, and to check outcome changes between registration and posting results. We developed a large language model (LLM)-based approach to evaluate outcome definitions and changes on ClinicalTrials.gov. Our LLM-based approach accurately identified incomplete outcomes in prospective trial registrations (sensitivity, 0.91 [95% confidence interval {CI}, 0.87–0.94]; positive predictive value [PPV], 0.98 [95% CI, 0.97–1.00]). Comparing prospective registrations with posted results, it correctly identified 96.1% of missing and 97.9% of added outcomes. It identified all outcomes with a priority change (e.g., from primary to secondary) and 99.2% without one. Sensitivity and PPV for identifying any outcome change were 0.97 (95% CI, 0.95–0.99) and 0.95 (95% CI, 0.89–0.98), respectively. Estimated costs per trial were $0.13 for o3-mini, $0.30 for GPT-4o, and $1.80 for o1. The accuracy of our approach suggests it could be used to improve registration quality and to detect outcome changes at large scale and at low cost.