Analysis of Local Recurrence After Robotic-Assisted Total Mesorectal Excision (ALRITE): An International, Multicentre, Retrospective Cohort

Abstract

Background/objectives: Rectal cancer is a major global health issue with high morbidity and mortality rates. Local recurrence (LR) significantly impacts patient outcomes, decreasing survival rates and often necessitating extensive secondary treatments. While robot-assisted total mesorectal excision (R-TME) is becoming a preferred method for rectal cancer surgery due to its improved precision and visualisation, long-term data on LR and predictors of recurrence remain limited. This study aims to determine the 3-year LR rate following R-TME and to identify predictors of recurrence to enhance patient selection and the personalisation of treatment.

Methods: This retrospective international multicentre cohort study included 1039 consecutive rectal cancer patients who underwent R-TME between 2013 and 2020, with a minimum of 3 years of follow-up. Data from tertiary colorectal centres in the United Kingdom, the Netherlands, Spain, France, Italy, and Belgium were analysed. Potential predictors of LR were identified using backward elimination, and four machine learning models were evaluated for predicting LR.

Results: The 3-year LR rate was 3.8%. Significant predictors of LR included advanced clinical M-staging, length of the hospital stay, postoperative ileus, postoperative complications, pathological N-staging, the completeness of resection, and the resection margin distance. The eXtreme Gradient Boosting model performed best for LR prediction, with a final accuracy of 77.1% and an AUC of 0.76.

Conclusions: R-TME in high-volume centres achieves low 3-year LR rates, suggesting that robot-assisted surgery offers oncological safety and advantages in rectal cancer management. This study underscores the importance of surgical precision, patient selection, and standardised perioperative care, supporting further investment in robotic training to improve long-term patient outcomes.

Keywords: artificial intelligence; prediction models; rectal cancer; robot-assisted surgery; total mesorectal excision.

Figure 1

ROC curve for XGB (eXtreme Gradient Boosting) prediction model with all variables included and oversampling in both training-and-testing (500 iterations of bootstrapping) and validation phases. ROC curve and AUC are metrics for demonstrating performance of model. Abbreviations: XGB, eXtreme Gradient Boosting; ROC, Receiver Operator Characteristic; and AUC, Area Under the Curve.

Figure 2

SHapley Additive exPlanations (SHAP) is an explainable AI model for the general model explanation of the XGB (eXtreme Gradient Boosting) prediction model. The feature value demonstrates the value of that variable, with blue highlighting the lower end of the scale and red highlighting the higher end of the scale. The variables are named on the left, and the dots represent the SHAP value that each variable has on the total model output, with pooled dots showing a more uniform and significant contribution to the model.

Figure 3

Local Interpretable Model-Agnostic Explanations (LIME) is an explainable AI model for the local explanation of the XGB (eXtreme Gradient Boosting) prediction model. The left is the total prediction probability of the model, which should approximately correspond with the incidence of the output variable. The right shows an example of a random prediction and which variables contributed in which way to the model decision.