The European Society for Medical Oncology (ESMO) has released its first set of recommendations for the use of artificial intelligence (AI) large language models in oncology practice, called the ESMO Guidance on the Use of Large Language Models in Clinical Practice (ELCAP). The guidance was published in Annals of Oncology.
“ESMO’s priority is to ensure that innovation translates into measurable benefit for patients and workable solutions for clinicians. With ELCAP, ESMO provides a pragmatic, oncology‑specific framework that embraces AI while upholding clinical responsibility, transparency, and robust data protection,” said Fabrice André, MD, PhD, President of ESMO.
Development
ELCAP was developed between November 2024 and February 2025 by the 20-member international ESMO Real World Data & Digital Health Task Force panel. The panel consisted of multispecialty experts across oncology, AI, biostatistics, digital health, ethics, and also included patient perspectives.
The panel defined three categories for large language model use in oncology:
- Type 1: Patient-facing applications, such as chat bots for patient education or lifestyle trackers
- Type 2: Health-care professional–facing tools, such as clinical decision support tools
- Type 3: Background institutional systems integrated with electronic health records, such as background alert systems
They developed 23 consensus statements for each type of model as practical guidance for day-to-day practice.
Guidances
The panel and the guidances recognize that the reliability of each model's output depends on how correct and complete the input data are, and that gaps in documentation or queries, for example, can lead to inaccurate or misleading information. This emphasizes the need for supervision of all AI model use in oncology practice.
“ELCAP recognizes that the value of language models depends on who is using them,” stated ELCAP co-author Miriam Koopman, MD, PhD, Chair of the ESMO Real World Data & Digital Health Task Force, who is also a Professor of Medical Oncology at the University Medical Center, Utrecht, the Netherlands. “By distinguishing patient‑facing, clinician‑facing, and background institutional systems, we set expectations for each context: supervised pathways for patients, validated and transparent tools for clinicians, and continuously monitored, well-governed systems embedded in the electronic health records.”
The guidances focus only on assistive models that operate with human oversight, as well as in alignment with clinical and ethical standards, to support clinicians with beneficial information or by drafting content, rather than on agentic AI that can take independent actions. “These systems are designed to enhance—and not replace—clinical workflows and decision-making,” added Jakob N. Kather, Deputy Chair of the ESMO Real World Data & Digital Health Task Force and co-author of the study. “At the same time, the guidance acknowledges the rapid emergence of autonomous, or ‘agentic’ AI models capable of initiating actions without direct prompts, which raises distinct safety, regulatory, and ethical challenges and will require dedicated future guidance.”
Additionally, the recommendations address issues around data privacy, algorithmic bias, regulatory compliance, and the risks of unsupervised AI use. The recommendations also urge that all type 2 and 3 large language models undergo systematic validation and continuous monitoring.
“Responsible use of AI in oncology requires shared standards as much as it requires algorithms; ELCAP sets out how to deploy language models in ways that improve the quality, equity and efficiency of cancer care, without compromising trust in clinical judgement”, Dr. André concluded.
Disclosure: For full disclosures of the study authors, visit annalsofoncology.org.
