AI may help provide clinical decision support for multidisciplinary care teams, according to findings from an evaluation focused on colorectal cancer cases. In the evaluation, the system received high performance ratings from clinicians for case summarization, management recommendations, and clinical trial matching.  

Findings from a blinded pilot evaluation of the agentic AI platform, called BEAGLE-MDT, involving synthetic cases were presented at the ESMO Congress 2026 (Abstract LBA18). 

Background

Presenting author Shazad Ashraf, DPhil, FRCS, of the Department of Colorectal Surgery at Queen Elizabeth Hospital in Birmingham, U.K., explained that clinicians and surgeons are overloaded with information, making it difficult to adequately prepare for multidisciplinary team meetings and make decisions.  

Additionally, missing data from various areas—including pathology, radiology, endoscopy, clinical history, molecular and genomic information, and trial eligibility—can delay team decisions.  

In an audit of 14,180 consecutive colorectal cancer multidisciplinary team cases at University Hospitals Birmingham, 8.2% were deferred. Sixty percent of these deferrals were due to missing histology or radiology information. Also, clinical trials were discussed in only 3.9% of cases, predominantly involving locally led studies.  

“We have a problem in our current multidisciplinary team process. There are inefficiencies. We’re in an area where there’s increasingly sophisticated, complex treatment decisions having to be made…. We need a better way,” Dr. Ashraf said. 

He noted that with advances in AI technology, AI agents could be used to summarize and collate large quantities of information, as well as identify gaps and potential clinical trials.  

Study and Model Methods

The researchers designed an auditable multi-agent AI platform to address some of these gaps at the point of multidisciplinary team decision-making. The platform comprised four linked functions and was updated to include additional features: a completeness check with an added deferral gate to flag missing information; evidence synthesis of guidelines and trial evidence relevant to each case; clinical trial matching based on patient summaries; and, finally, creation of a treatment plan in a standardized format for each patient.  

They ensured that each statement from the AI platform was linked to a source and auditable, that all outputs were reviewed by the multidisciplinary team before any action was taken, and that a survey portal was added.  

The prototype was evaluated in two stages. The first involved a pilot of 15 synthetic colorectal cancer cases, varying by disease site and stage. Five expert clinicians received the same information as the AI agents and were asked to rate the platform’s outputs across three domains: patient summarization, management recommendations, and trial matching. In the second stage, version 2 of the platform was evaluated by 15 multidisciplinary team members through a survey portal.  

Findings

Clinicians rated the AI platform highly across all three domains in the first phase of the evaluation. For patient summarization, BEAGLE-MDT received a score of 9.8 out of 10, with 64 of 70 total ratings being 10 out of 10. For management recommendations, the platform achieved a score of 9.2 out of 10, with 73% of all ratings being 10 out of 10. For trial matching, the platform received a score of 9.5 out of 10, with 84% of all ratings being 10 out of 10.  

In the second phase, the multidisciplinary team gave the platform an overall score of 9.0 out of 10 for patient summarization, with 92% of 145 total reviews receiving ratings of 8 to 10. Colorectal surgeons gave the platform the lowest rating, at 8.98 out of 10, whereas medical oncologists, nurses, and multidisciplinary team coordinators rated it 9.2, and pathologists gave it full marks.  

“The summarization capability of the agent was deemed outstanding in a large portion of cases,” Dr. Ashraf commented.  

With the addition of a deferral gate to flag missing information, the AI platform deferred 32 of 81 cases. Seventy-five percent of the multidisciplinary team members agreed with the deferrals.  

“We live in a system where there’s a huge number of patients being processed…and having a tool [that] increases efficiency is important,” Dr. Ashraf noted.  

For management plans in the second phase, the overall score was 8.7 out of 10, compared with 8.6 for trial matching. BEAGLE-MDT identified open U.K. clinical trials for 92% of cases, with 29% involving actionable treatment trials. The oncologists flagged four trials that had not been included, but two of these were no longer open or were about to close.  

The research team is currently refining the platform, including calibrating the deferral gate and improving trial identification. Going forward, the researchers plan to further evaluate BEAGLE-MDT in shadow mode, in which it runs alongside a live multidisciplinary team without influencing decisions, and in a prospective, multicenter study comparing its performance with the standard workflow in consecutive colorectal cancer cases. 

Insights

Invited discussant Vanja Miskovic, PhD, Assistant Professor of Politecnico di Milano and co-leader of the AI-ON-Lab of the IRCCS Istituto Nazionale Tumori, explained that generative AI is increasingly being used in oncology clinical workflows, although prospective evidence focusing on safety, clinical utility, workflow impact, and patient outcomes remains limited.  

Dr. Miskovic highlighted the strengths of the evaluation, including the synthetic-case pilot, the multiple agentic tasks, and the high performance of the platform. “I think the biggest strength of this multimodal agent is that it supports different steps of the workflow,” she commented.  

She also questioned what the high clinical ratings really mean. In terms of management ratings, does perceived quality translate to objective performance and clinical accuracy? In the evaluation, there was 80% agreement among human clinicians, raising the question of how clinically acceptable decisions can be defined when even experts disagree. The study findings did not reveal whether disagreements were due to clinically significant errors made by the AI or whether the agent simply chose a different option from the one the reviewer would have selected. Additionally, although reviews were conducted in an average of 6 minutes, that did not necessarily mean that the platform provided a workflow benefit or saved time.  

She expressed hope that the next steps in the research would answer some of these questions.  

DISCLOSURES: Dr. Ashraf reported that the BEAGLE-MDT engine and evaluation portal were designed and developed by Histofy, and that his partner holds shares in Histofy. Study authors who are also Histofy employees included Nasir Rajpoot, David Snead, and Adnan Khan. For full disclosures of the other study authors, visit ctimeetingtech.com/esmo2026. 

Dr. Miskovic has been an invited speaker for Novartis, Merck, and Daiichi Sankyo, and is a principal investigator for a local Pfizer grant.  

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.