“ChatGPT-4 mirrored our MDT in four of five colorectal-cancer cases.”
Tumour Boards (MDT) are swamped. Some UK centres now log more than 30 hours each month discussing colorectal cancer (CRC) cases. Yet large-language models (LLMs) such as ChatGPT could already shoulder a sizeable slice of that load.
Study snapshot
We retrospectively audited 100 consecutive CRC cases discussed by our lower-GI MDT (September-December 2023)1 Raw patient data including surgical notes, preoperative work up, TNM staging, comorbidities and imaging summaries were uploaded to ChatGPT‑4 without prior medical interpretation. The model returned a single “best” plan for the pre‑ and (where applicable) post‑therapeutic settings (Figure 1). After that the concordance with the German S3 guideline was manually assessed by our team (Table 1)2. In the pretherapeutic discussions, complete concordance was observed in 72.5 per cent, with partial concordance in 10.2 per cent and discordance in 17.3 per cent. For post-therapeutic discussions, the concordance increased to 82.8 per cent; 11.8 per cent of decisions displayed partial concordance and 5.4 per cent demonstrated discordance. Key discordance drivers were age > 77 years (OR 1.08 per year) (Figure 2), ASA ≥ III, rectal cancer, N1 stage on the TNM Classification.
Figure 1.
ChatGPT prompt.
Table 1.
Concordance Results
|
Decision point
|
Complete concordance
|
Partial concordance
|
Discordance
|
n
|
|
Pre‑therapeutic MDT
|
72.5 %
|
10.2 %
|
17.3 %
|
98*
|
|
Post‑therapeutic MDT
|
82.8 %
|
11.8 %
|
5.4 %
|
93†
|
*Two emergency ileus cases bypassed the pre‑therapeutic MDT. †Seven post‑therapeutic decisions were unnecessary (endoscopic R0, benign histology or palliative pathway).
In cases of discordance, the MDT was in alignment with the guidelines in up to 94.1% of the decisions (Figure 3). After removing the patients with risk factors for discordance from the cohort, concordance rates of up to 85.1% were observed, while the distribution of answers exhibited no significant deviation (χ2(2) = 0.77, 95% c.i., P = 0.679), signaling a more normal distribution of decisions.
Why does this matter?
It is evident that workload relief is a significant concern, as evidenced by the fact that a considerable number of MDTs are now encountering in excess of 500 cases per annum3. The allocation of MDT minutes to complex cases can be optimized by delegating the triaging of straightforward decisions to LLM. Moreover LLMs could be a crucial part in a bias-check process, which supports the need for reproducible and guideline-anchored decisionmaking. The purposes of such a bias check are to identify human outliers and stimulate discussion. The use of LLM-enhanced MDT decision could not only help move to more time and resource efficient decision-making in the fields of CRC management, but also help residents familiarize with the newest guidelines and thinking process for those patients.
Figure 2.
Age as risk factor for discordance.
How do these results compare?
Human auditors report 71–85 % adherence to NICE CRC guidance in UK national audits according to Harari et al (4). Our ChatGPT concordance (73–83 %) therefore sits within the current human benchmark (1). Similar single‑centre LLM studies show 65–86 % concordance in CRC, 60–75 % in breast and gastric boards, and superiority to a rival model in head‑and‑neck MDTs5-8. The above-mentioned literature results support the idea that the ChatGPT decision enhancement in CRC decision making could be safely deployed in the whole spectrum of MDTs, provided there is also human supervision.
Limitations & Safeguards
In our study, frailty and personal preferences of the patients were the so called “blind spots” of the ChatGPT. Nearly all discordant cases were frail elderly patients with ASA grade ≥ III. Weighing functional reserve, or patient wishes and considerations is thus a current limitation of the LLMs. The role of the notes of the physicians is of great importance because, given the well documented wishes of the patients, could lead to more accurate results.
The prompt engineering risk was addressed in our study by requesting a single "best" option, which enhanced clarity. Conversely, the presence of multiple options resulted in an escalation of partial concordance. It is imperative that the outputs of AI remain subject to audit and be endorsed by the MDT committee.
Figure 3.
Concordance of the decisions with the German S3 Guidelines.
Distribution (%) of the side that had the correct recommendation after partial agreement or disagreement before and after surgery.
Where next?
The integration of electronic health records (EHR) should be developed to facilitate the generation of one-click prompts. It is imperative to undertake meticulous refinement of regional pathways and frailty metrics with a view to enhancing AI performance in high-risk subgroups9.
An international Delphi Consensus should address the ethical and legal considerations that must be taken into account when AI is employed in the context of decision-making processes10. The subsequent phase of the project must entail multicentre "silent-member" trials, the objective of which is to measure meeting duration, decision quality and patient outcomes. These trials should be conducted in parallel with ChatGPT.
Take-home message
Provided EHR raw data, ChatGPT‑4 matched CRC MDT decisions in four of five cases, rivalling national human guideline‑adherence rates. The model is not ready to replace live MDT expertise, especially for frail patients, but it already looks useful as an efficiency booster and bias check.
Why not run ChatGPT on next week’s agenda and see which patients it would fast‑track?
References
Chatziisaak D, Burri P, Sparn M, Hahnloser D, Steffen T, Bischofberger S. Concordance of ChatGPT artificial intelligence decision-making in colorectal cancer multidisciplinary meetings: retrospective study. BJS Open. 2025;9(3). doi: 10.1093/bjsopen/zraf040
AWMF. German Guideline Program in Oncology (German Cancer Society, German Cancer Aid, AWMF): S3-Guideline Colorectal Cancer, long version 2.1, 2019. 2019.
Soukup T, Lamb B, Shah N, Morbi A, Bali A, Asher V, et al. Relationships between communication, time pressure, workload, task complexity, logistical issues and group composition in transdisciplinary teams: A prospective observational study across 822 cancer cases. Frontiers in Communication. 2020;5.
Harari D, Husk J, Lowe D, Wagg A. National audit of continence care: adherence to National Institute for Health and Clinical Excellence (NICE) guidance in older versus younger adults with faecal incontinence. Age Ageing. 2014;43(6):785-93.
Choo JM, Ryu HS, Kim JS, Cheong JY, Baek SJ, Kwak JM, et al. Conversational artificial intelligence (chatGPT™) in the management of complex colorectal cancer patients: early experience. ANZ J Surg. 2024;94(3):356-61.
Park YE, Chae H. The Fidelity of Artificial Intelligence to Multidisciplinary Tumor Board Recommendations for Patients with Gastric Cancer: A Retrospective Study. J Gastrointest Cancer. 2024;55(1):365-72.
Lorenzi A, Pugliese G, Maniaci A, Lechien JR, Allevi F, Boscolo-Rizzo P, et al. Reliability of large language models for advanced head and neck malignancies management: a comparison between ChatGPT 4 and Gemini Advanced. Eur Arch Otorhinolaryngol. 2024;281(9):5001-6.
Li H, Huang J, Liu K, Liu J, Liu Q, Zhou Z, et al. ChatGPT-4o outperforms gemini advanced in assisting multidisciplinary decision-making for advanced gastric cancer. Eur J Surg Oncol. 2025;51(8):110096.
Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56.
Hantel A, Clancy DD, Kehl KL, Marron JM, Van Allen EM, Abel GA. A Process Framework for Ethically Deploying Artificial Intelligence in Oncology. J Clin Oncol. 2022;40(34):3907-11.



