Real-world findings show moderate-to-high concordance between AI-generated and human experts decisions
As multidisciplinary tumour boards (MDT) face increasing complexity due to a time-consuming, heterogeneous operational and burdensome process, artificial intelligence (AI)-assisted support may help streamline case preparation, improve consistency, and identify guideline-concordant treatment options.
The new potential role of AI emerges from findings of a prospective blinded concordance study in which 106 patient cases across 21 tumour sites were evaluated in parallel by a MDT at the Basavatarakam Indo-American Cancer Hospital & Research Institute, Hyderabad, India, and ChatGPT® (GPT-4 or GPT-5), reporting moderate-to-high concordance between AI-generated recommendations and MDT decisions (ESMO Real World Data and Digital Oncology, 2026; 11).
The primary endpoint was concordance between AI-generated recommendations and MDT decisions for the main clinical question under consideration. Using a predefined three-point scoring system, investigators reported a mean concordance score (MCS) of 1.42 out of a maximum of 2, corresponding to approximately 71% of maximum possible agreement.
Overall, 54.7% of primary clinical queries demonstrated full concordance, while 33% showed partial concordance and only 12.3% were completely discordant. In partially concordant cases, AI-generated recommendations were deemed preferable more often than MDT recommendations (19.8% versus 13.2%). The highest level of agreement was observed for treatment-intent decisions, with an MCS of 1.74 and full concordance in 84.6% of cases. Radiotherapy recommendations also showed strong alignment (MCS 1.68), followed by surgical decision-making (MCS 1.55). By contrast, systemic therapy selection (MCS 1.40) and diagnostic recommendations (MCS 1.39) showed lower concordance, reflecting the greater complexity and contextual variability inherent to these domains.
According to Dr Rodrigo Dienstmann of the Oncology Data Science (ODysSey) Group at Vall d’Hebron Institute of Oncology, Barcelona, Spain, and Editor-in-Chief of the ESMO Real World Data and Digital Oncology, the study clearly exemplifies that “AI is beginning to evolve from a passive information retrieval tool toward an active participant in clinical reasoning, while still requiring physician oversight, governance frameworks, and rigorous prospective validation,” he wrote in an accompanying editorial in the ‘Artificial Intelligence in Clinical Oncology’ Special Issue, where the results were published.
In the same issue, a perspective article authored by oncology experts from several institutions described how to design a natural language processing (NLP)-assisted workflow to facilitate and enhance the MDT process by using small NLP, optimised generative NLP and powerful foundational generative NLP, which include large language models (LLMs) (ESMO Real World Data and Digital Oncology, 2026; 11).
Small NLP models, typically processing up to 500 word, can accurately perform targeted information extraction tasks such as named entity recognition, question answering, relation extraction, and classification, while requiring relatively modest computational resources, straightforward validation, and infrastructure often already available within healthcare institutions, such as pathology or next generation sequencing reports. Optimised generative NLP models can process larger volumes of text (up to 5,000 words), enabling them to integrate and structure information from diverse clinical data sources while providing greater contextual understanding and flexibility than small NLP models, including medical text summarization. However, they come with increased computational requirements and more rigorous validation. For complex clinical workflows, foundational generative NLP models and emerging agentic AI systems can autonomously coordinate specialised tools, synthesise and aggregate information across large and diverse datasets, and support comprehensive case analysis beyond simple data extraction. However, their deployment in clinical practice remains limited by concerns around reliability, transparency, and accountability, necessitating extensive prompt engineering, robust validation, and expert human review.
As Dienstmann highlighted in his editorial, oncology is now entering a new era in the adoption of AI-based solutions, “characterised not by proof-of-concept demonstrations alone, but by prospective validation, real-world implementation, and measurable impact on clinical workflows.”
