Web-based explainable machine-learning tool for predicting five-year recurrence of colorectal cancer after curative...

The World Health Organization identifies colorectal cancer as the third-most diagnosed malignancy and second leading cause of cancer-related mortality worldwide. Up to 30% of patients relapse within 5 years postoperatively; however, conventional staging methods cannot reliably stratify individual risks, underscoring the need for precise, patient-centred decision-support tools.

We retrospectively analysed data on 1,789 colorectal cancer patients undergoing curative resection (2013-2023) from the Tri-Service General Hospital registry. Four tree-based machine learning algorithms were trained on demographic, tumour, immunohistochemical, and laboratory features. Patients were divided chronologically into training (January to September) and validation (October to December) sets. Feature importance was assessed using random forest impurity scores and Shapley additive explanations (SHAP). A web-based artificial intelligence-clinical decision support system (AI-CDSS) was developed to provide real-time, scenario-specific five-year recurrence-risk estimates.

Of the 1,789 patients, 406 (22.7%) experienced recurrence. The top 10 predictors accounted for approximately 43% of the total model importance. SHAP analysis confirmed that tumour burden, biological markers, treatment intensity, and host factors were key drivers of recurrence risk. On the validation set, all models achieved area under the receiver operating characteristic curve values of 0.83-0.84. The random forest-based system demonstrated 87% accuracy, 85% positive predictive value, 87% negative predictive value, and an F1 score of 0.64, and was consequently selected as the AI-CDSS engine.

The proposed AI-CDSS delivers personalised five-year recurrence-risk estimates for patients with colorectal cancer within 1 s via an intuitive web interface, facilitating evidence-based, patient-centred treatment decisions grounded in local population data.