Scholarlink

<p>Artificial intelligence (AI) can support cancer care in many different ways. Classic, nongenerative AI is already established in oncology, as narrow applications (“one problem, one solution”) such as polyp and lung-nodule detection or automated differential blood counts are clearly measurable, prospectively testable, and validatable under current regulations. Large language models (LLMs) now allow for a&#xa0;completely new and broader usage of AI, as they offer substantial documentation relief (transcription, structuring) and real-time clinical decision-support systems (CDSS) that connect patient data to guidelines or studies or can match patients to clinical trials. Risks arise from model misreasoning (shortcut learning), hallucinations, and overconfirmation, as well as from shadow use, privacy/liability, vendor lock-in, de-skilling, and fairness. We outline pragmatic guardrails such as source-grounded recommendations; human-in-the-loop review; explicit uncertainty; alternatives; auditability; and safe, electronic health record (EHR)-integrated pathways.</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

Künstliche Intelligenz: Chancen und Grenzen

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Jan Clusmann,
  • Carolin V. Schneider,
  • Sebastian Foersch,
  • Daniel Truhn,
  • Jakob N. Kather

摘要

Artificial intelligence (AI) can support cancer care in many different ways. Classic, nongenerative AI is already established in oncology, as narrow applications (“one problem, one solution”) such as polyp and lung-nodule detection or automated differential blood counts are clearly measurable, prospectively testable, and validatable under current regulations. Large language models (LLMs) now allow for a completely new and broader usage of AI, as they offer substantial documentation relief (transcription, structuring) and real-time clinical decision-support systems (CDSS) that connect patient data to guidelines or studies or can match patients to clinical trials. Risks arise from model misreasoning (shortcut learning), hallucinations, and overconfirmation, as well as from shadow use, privacy/liability, vendor lock-in, de-skilling, and fairness. We outline pragmatic guardrails such as source-grounded recommendations; human-in-the-loop review; explicit uncertainty; alternatives; auditability; and safe, electronic health record (EHR)-integrated pathways.