User perceptions of machine learning models as decision support for colorectal cancer multidisciplinary team conferences (AID-SIM-2): a qualitative simulation study.

BACKGROUND: Multidisciplinary team (MDT) conferences are considered a cornerstone of decision-making in cancer diagnostics and care. However, the current literature has not demonstrated improved patient outcomes based on the decisions of the MDT conferences. AIM: We aimed to evaluate how four different decision-support modalities impacted the decision-making process and the internal discussions in the MDTs in a multicenter simulation study, with a focus on user perceptions. METHODS: Four colorectal cancer centers with MDTs participated. We performed four simulations in each center. Each simulation used a different decision-support tool: (1) Current standard, (2) Current standard plus a prediction model, (3) A structured data presentation tool, and (4) A structured data presentation tool plus the prediction model. Clinician- and model-estimated risks were compared, the treatment suggestions from each site were compared, questionnaires about user perceptions were conducted after Simulations 2, 3, and 4 using a Google Form link, and a semi-structured interview was conducted at each site after the last simulation. RESULTS: Similar distributions of risk groups between clinicians and models were found; however, distinct discrepancies in predictions arose, particularly with higher-risk patients, highlighting the need for standardization for more complex clinical cases. The primary perceived benefit of decision support was increased standardization of care, independent of the individual physicians' personal views. However, participants emphasized the necessity of clinician autonomy to overrule tool suggestions when identifying clinical nuances not captured by the model. CONCLUSIONS: The colorectal cancer MDTs expressed a positive view regarding the use of prediction models and other forms of decision-support in their workflow. While clinicians and prediction models had similar risk score distributions, they diverged in the assessment of specific individual patients.