Interpretable machine learning model for predicting operative difficulty in robotic total mesorectal excision for...

Precise preoperative prediction of surgical complexity in robot-assisted total mesorectal excision (R-TME) is essential for optimizing surgical strategies. The current study aimed to construct an interpretable machine learning (ML) model to anticipate operative difficulty in sphincter-preserving R-TME. Retrospective data from 449 patients diagnosed with mid-to-low rectal cancer undergoing R-TME at Center A were analyzed. The dataset was partitioned randomly into training (n = 314) and internal validation (n = 135) groups using a 7:3 ratio. Additionally, external validation was conducted with a prospective cohort (n = 100) from multiple centers. Operative difficulty was quantified using a scoring system ranging from 0 to 13. Feature selection was performed employing Least Absolute Shrinkage and Selection Operator (LASSO) regression, followed by evaluation of five ML algorithms. Model accuracy and robustness were measured using metrics including the Area Under the Curve (AUC), calibration plots, decision curve analysis (DCA), and supplementary indicators. Interpretability of the predictive model was enhanced using SHapley Additive exPlanations (SHAP). Critical predictive factors comprised BMI, neoadjuvant treatment status, clinical staging, tumor distance to the anal verge, interspinous diameter, lateral mesorectal width, posterior mesorectal thickness, and two specific pelvic angle measurements. Among evaluated ML methods, Gradient Boosting Machine (GBM) demonstrated superior performance, achieving an AUC of 0.874 in the training cohort and 0.835 in internal validation. Calibration plots and DCA indicated excellent robustness and significant clinical applicability of the GBM model. Furthermore, external validation presented an AUC of 0.809, confirming the model's generalizability. SHAP-based analysis delineated individual predictor impacts, facilitating the creation of an accessible online prediction instrument. This study successfully established and externally validated a transparent ML-based model to predict operative challenges in sphincter-preserving R-TME. This model can effectively aid surgeons in identifying anticipated difficulties prior to surgery, thereby enhancing clinical decision-making and improving surgical planning.