AI-augmented colorectal cancer MDT concordance modelling: a prospective evaluation of performance, interpretability, and provider-side budget impact

Abstract Purpose Colorectal multidisciplinary teams (MDTs) face increasing workload. We evaluated whether routinely collected colorectal MDT variables could support local MDT concordance modelling, interpretable feature attribution, and provider-side budget impact estimation. Methods Prospective observational cohort of 250 consecutive adults with newly diagnosed colorectal cancer discussed at a single UK district general hospital MDT (June 2023–June 2025). Six machine learning model families were trained sequentially on cumulative 50-patient blocks and tested on subsequent non-overlapping 50-patient blocks. Primary outcome was grouped MDT decision concordance. Final block performance, clinical pattern sensitivity, exploratory calibration, SHAP explanations, and provider-side opportunity value were assessed. Results Median age was 69.0 years; 136 (54.4%) were male; tumours were colonic in 166 (66.4%) and rectal/anal/rectosigmoid in 84 (33.6%). Surgery was recommended in 170 (68.0%) cases. Accuracy improved across blocks for all model families ( P ≤ 0.001). In the final hold-out block, random forest, gradient boosting, and neural network achieved 1.00 concordance (95% CI 0.929–1.000); logistic regression, support vector machine, and naïve Bayes achieved 0.92, 0.94, and 0.76, respectively. Clinical pattern sensitivity identified 25 unique core patterns; 24/25 (96.0%) mapped to a single MDT outcome. Leading SHAP-attributed features were tumour site, neoadjuvant status, differentiation/dysplasia grade, MRI–TNM descriptor, ASA grade, and M stage. At 95% concordance, modelled provider-side opportunity value was £26,575 per 100 cases at 50% adoption and £42,320 at full adoption. Conclusion Routinely collected colorectal MDT data supported high MDT decision concordance with clinically plausible explanations and potential opportunity value. These findings support further evaluation of AI as an MDT concordance and triage support tool, not an autonomous treatment decision system; multicentre validation is required.