The rise of large language models represents not just a technological leap, but also a paradigm shift in how healthcare professionals will interact with vast amounts of medical knowledge. ESMO President Fabrice André emphasized the organization’s commitment to ensuring that innovation in this area translates into tangible benefits for patients while offering workable solutions for healthcare providers. The ELCAP framework allows for a nuanced approach, acknowledging the varied contexts in which AI language models might be applied—whether they be aimed at patients, clinicians, or healthcare institutions.

Fundamentally, ELCAP is structured around three distinct categories that cater to user-specific needs and contexts. The first, Type 1, is tailored for patient-facing applications. These include chatbots designed for education and symptom management, which are intended to complement traditional clinical care. However, they operate under a stringent supervision protocol, ensuring that there is a clear pathway for escalation in serious cases. This careful balancing act aims to protect patient data while providing them with immediate access to information tailored to their needs.

Type 2 addresses tools intended for healthcare professionals, focusing on decision support systems, clinical documentation, and necessary translations. The recommendations stipulate that these instruments undergo formal validation to ensure their reliability in clinical circumstances. Moreover, transparency about the limitations of these models is critical, establishing a framework where human accountability is at the forefront of clinical decision-making.

The third category, Type 3, pertains to institutional systems integrated with electronic health records. These systems can simplify processes such as data extraction, create automated summaries, and facilitate matching patients with clinical trials. ELCAP emphasizes that these systems must not only be tested prior to deployment but also continuously monitored for bias and performance shifts. The guidance highlights the importance of institutional governance, stressing that any change in data source or process necessitates re-validation to ensure ongoing compliance with safety protocols.

As ELCAP points out, the quality of the output produced by these AI systems is fundamentally linked to the quality of the input data. Incomplete clinical documentation or vague patient queries could result in erroneous or misleading responses, reinforcing the necessity for vigilant supervision and clear escalation routes for addressing issues that arise. The guidance acts as both a roadmap for navigating potential pitfalls and a springboard for the responsible application of AI tools within the healthcare setting.

Miriam Koopman, who chairs ESMO’s Real World Data & Digital Health Task Force and contributed to the paper, reinforced that the effectiveness of language models is highly dependent on the context of their use. By categorizing applications based on their audience—patients, clinicians, and institutions—expectations can be appropriately aligned. This structured approach is designed to protect patients, ensure validated tools for clinicians, and maintain governance in institutional settings.

ELCAP emphasizes the role of assistive large language models, which are meant to support clinicians rather than supplant their expertise. By providing essential information or drafting preliminary content, these systems are set to enhance clinical workflows and decision-making processes. Deputy Chair Jakob N. Kather, also co-author of the study, noted that while current models offer promising enhancements to patient care, guidance must evolve to address autonomous AI models capable of initiating actions without direct human input, as these present unique safety, regulatory, and ethical challenges.

Looking forward, the foundation of trust in AI-driven cancer care hinges not just on the technology itself, but also on the establishment of shared standards across varying applications. André’s concluding remarks stressed that the integration of algorithms in oncology must go hand-in-hand with maintaining trust in clinical judgment. ELCAP serves as a crucial step in outlining how language models can be harnessed to improve the quality, equity, and efficiency of cancer care, all while safeguarding the integrity of medical decisions.

The development of ELCAP was an extensive process, undertaken by a diverse international panel comprised of experts in oncology, AI, biostatistics, digital health, ethics, and patient advocacy. This collaborative effort took place between November 2024 and February 2025, and exemplifies the commitment to an interdisciplinary approach in addressing the complexities of AI integration into health practices.

In summary, the ESMO’s ELCAP framework signals a pioneering approach to the adaptation of AI in oncology, setting the stage for future innovations while embedding essential safeguards to uphold patient welfare and clinical integrity. As this guidance takes root in clinical practice, it is poised to facilitate a transformative evolution in the delivery of cancer care, solidifying the place of AI as a valuable ally in the ongoing battle against cancer.

Subject of Research: Integration of Large Language Models in Oncology
Article Title: ESMO Guidance on the Use of Large Language Models in Clinical Practice (ELCAP)
News Publication Date: 20 October 2025
Web References: https://www.annalsofoncology.org/article/%20S0923-7534(25)04698-8%20/fulltext
References: E.Y.T. Wong et al. Annals of Oncology. doi: 10.1016/j.annonc.2025.09.001
Image Credits: Not provided

Keywords

AI, Oncology, Large Language Models, Clinical Practice, Patient Safety, Clinical Decision-Making

Circulating tumor DNA refers to tiny fragments of cancer-derived DNA that circulate freely in a patient’s bloodstream. Its detection after surgery indicates minimal residual disease (MRD), a state where microscopic tumor cells persist and could eventually drive cancer relapse. Traditionally, clinicians lacked robust tools to identify MRD, leading to a one-size-fits-all approach in post-operative treatment. The IMvigor011 trial utilized a highly personalized ctDNA assay, with blood samples screened every six weeks for up to a year following surgery. This rigorous monitoring enabled researchers to classify patients into ctDNA-positive or ctDNA-negative groups, guiding targeted immunotherapeutic intervention.

Atezolizumab functions by blocking PD-L1, a protein frequently overexpressed on cancer cells that suppresses the immune system’s ability to recognize and attack tumors. By inhibiting this checkpoint, atezolizumab effectively unmasks cancer cells, allowing T cells to mount an immune response. While previous studies, including the IMvigor010 trial, tested atezolizumab in unselected MIBC patients post-surgery, they failed to demonstrate a clear overall survival benefit. Retrospective analyses suggested that this lack of effect was due to the inclusion of patients without residual disease who were unlikely to benefit from immunotherapy, highlighting the need for better patient stratification.

In IMvigor011, 800 patients with no clinical evidence of disease following surgery were enrolled and subjected to personalized ctDNA testing every six weeks. Approximately 250 patients who tested positive for ctDNA were randomized to receive either atezolizumab or placebo in a 2:1 ratio. Strikingly, patients receiving atezolizumab demonstrated a 36% reduction in the risk of disease recurrence compared to placebo. More impressively, the risk of death was reduced by 41% among ctDNA-positive patients receiving the immunotherapy, a landmark finding in the context of adjuvant treatments for MIBC.

Another vital insight from the trial was that ctDNA screening captured patients with residual disease regardless of when ctDNA positivity emerged—from immediately post-surgery or during subsequent surveillance within the first year. This dynamic ability to identify MRD highlights the utility of ctDNA as a real-time biomarker, refining treatment decisions dynamically and enabling clinicians to escalate or withhold therapy based on evolving risk profiles.

Equally important was the observation that ctDNA-negative patients, who did not receive immunotherapy, experienced excellent outcomes. Approximately 89% remained disease-free and over 90% were alive at a median follow-up of 21.8 months without additional treatment. This finding confirms that ctDNA negativity reliably identifies patients with a low risk of recurrence, creating an opportunity to avoid overtreatment and the associated financial and physical burdens.

The absence of new or unexpected adverse effects in the atezolizumab-treated cohort reinforces the safety of this approach when guided by ctDNA stratification. Given the immune-related toxicities known for checkpoint inhibitors, such selective treatment minimizes unnecessary exposure among those unlikely to benefit. This targeted methodology exemplifies personalized medicine’s promise by matching treatment intensity with individual patient biology.

Dr. Joaquim Bellmunt, co-principal investigator and director of the Bladder Cancer Center at Dana-Farber, emphasized the clinical significance: “This is the first adjuvant immunotherapy trial that has demonstrated a survival benefit for patients selected by ctDNA testing. It marks a pivotal step towards precision oncology where therapeutic decisions are no longer ‘one size fits all’ but are tailored to the molecular fingerprints of residual disease.”

The implications for regulatory frameworks and clinical guidelines are profound. Regulatory agencies are currently evaluating whether ctDNA-guided use of atezolizumab should become the new standard of care for MIBC patients after surgery. Adoption of such biomarkers into routine practice could redefine oncological workflows by embedding minimally invasive blood-based diagnostics as decision-making tools for adjuvant therapies.

This study was funded by F. Hoffmann-La Roche Ltd, with collaboration from Natera, a leader in ctDNA assay development. The partnership underscores the critical role of industry-scientific collaboration in rapidly translating molecular diagnostics into clinical impact.

Dana-Farber Cancer Institute, known for its integrative approach to cancer treatment and research, continues to pioneer innovations that bridge laboratory discoveries with patient care. Its involvement in trials like IMvigor011 reinforces its mission to reduce the global cancer burden through scientific inquiry and compassionate, evidence-based care.

In summary, the IMvigor011 trial charts a new course in bladder cancer therapy by harnessing the precision of ctDNA to focus immunotherapy on patients most likely to benefit. This approach offers hope for improved survival while preserving quality of life, setting a precedent for similar strategies in other malignancies where minimal residual disease detection and targeted therapy can intersect to optimize outcomes.

Subject of Research: Muscle-invasive bladder cancer, circulating tumor DNA-guided immunotherapy

Article Title: ctDNA-Guided Adjuvant Atezolizumab in Muscle-Invasive Bladder Cancer

News Publication Date: 20-Oct-2025

Web References:
ESMO 2025 Congress Presentation
New England Journal of Medicine Article

References:
IMvigor011 Phase 3 Clinical Trial Data, Dana-Farber Cancer Institute et al., NEJM, 2025

Image Credits: Dana-Farber Cancer Institute

Keywords: Cancer, Muscle-invasive bladder cancer, Circulating tumor DNA, Immunotherapy, Atezolizumab, Minimal residual disease, Precision oncology