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

Area of Science:

• Surgical Oncology
• Medical Informatics
• Machine Learning in Medicine

Background:

• Precise prediction of surgical complexity in robot-assisted total mesorectal excision (R-TME) is crucial for optimizing surgical strategies.
• Current methods for assessing operative difficulty in R-TME are limited, necessitating advanced predictive tools.

Purpose of the Study:

• To construct an interpretable machine learning (ML) model for predicting operative difficulty in sphincter-preserving R-TME.
• To validate the model's performance using internal and external datasets.

Main Methods:

• Retrospective analysis of 449 R-TME cases, with data split into training and internal validation sets.
• External validation using a prospective cohort of 100 patients.
• Feature selection via LASSO regression, model evaluation using Gradient Boosting Machine (GBM), and interpretability analysis with SHAP.

Main Results:

• The GBM model achieved an AUC of 0.874 (training) and 0.835 (internal validation), with external validation yielding an AUC of 0.809.
• Key predictors included BMI, neoadjuvant treatment, clinical stage, tumor distance, pelvic dimensions, and mesorectal measurements.
• SHAP analysis provided insights into individual predictor impacts, enabling the development of an online prediction tool.

Conclusions:

• An interpretable and externally validated ML model (GBM) effectively predicts operative difficulty in sphincter-preserving R-TME.
• This tool can significantly aid surgeons in pre-operative decision-making and surgical planning.
• The model's transparency and generalizability enhance its clinical applicability for anticipating surgical challenges.