Scholarlink

<p>Publicly available trial matching tools can improve the access to therapeutic innovations, but errors may expose to over-solicitation and disappointment. We performed a pragmatic non-interventional prospective evaluation on sequential patients at the Molecular Tumor Board of Centre Leon Berard. During 10 weeks in 2024, we analysed 157 patients with four clinical trial matching tools from the 19 screened: Klineo, ScreenAct, Trialing and DigitalECMT. Each patient had 2.19 trials proposed on average, and 38% had no trials suggested. The mean performances were precision = 0.33, recall = 0.32, AP@3 = 0.45, and NDCG@3 = 0.34. Using all the tools can increase to 26% the clinical trial options. The most frequent error concerned the type of gene variants required by the selection criteria. We showed that using a Large Language Model on the patients’ molecular reports could improve the performance by up to 5%. We recommend that experts supervise the results and we advocate for improved technologies.</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

A prospective pragmatic evaluation of automatic trial matching tools in a molecular tumor board

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年1月27日

查看全文
  • Lilia Gueguen,
  • Louise Olgiati,
  • Clément Brutti-Mairesse,
  • Alric Sans,
  • Vincent Le Texier,
  • Loic Verlingue

摘要

Publicly available trial matching tools can improve the access to therapeutic innovations, but errors may expose to over-solicitation and disappointment. We performed a pragmatic non-interventional prospective evaluation on sequential patients at the Molecular Tumor Board of Centre Leon Berard. During 10 weeks in 2024, we analysed 157 patients with four clinical trial matching tools from the 19 screened: Klineo, ScreenAct, Trialing and DigitalECMT. Each patient had 2.19 trials proposed on average, and 38% had no trials suggested. The mean performances were precision = 0.33, recall = 0.32, AP@3 = 0.45, and NDCG@3 = 0.34. Using all the tools can increase to 26% the clinical trial options. The most frequent error concerned the type of gene variants required by the selection criteria. We showed that using a Large Language Model on the patients’ molecular reports could improve the performance by up to 5%. We recommend that experts supervise the results and we advocate for improved technologies.