Exploratory machine learning analysis of factors associated with positive circumferential resection margins following transanal total mesorectal excision: a multicenter registry study

Technical Summary: Exploratory Machine Learning Analysis of Factors Associated with Positive Circumferential Resection Margins Following taTME

Problem Statement
Positive circumferential resection margin (CRM) is a critical oncological quality indicator in rectal cancer surgery, strongly associated with local recurrence and reduced survival. While transanal total mesorectal excision (taTME) was introduced to improve visualization and dissection in anatomically challenging cases (e.g., narrow pelvis, obesity), achieving negative CRM remains a primary surgical objective. Previous registry-based studies have utilized logistic regression to identify risk factors such as threatened mesorectal fascia (MRF) and advanced tumor stage. However, the incremental value of machine learning (ML) algorithms in modeling nonlinear relationships and interpreting complex variable interactions within real-world registry data requires further exploration. This study addresses the need to explore factors associated with CRM positivity after taTME using a large multicenter registry, comparing multiple ML algorithms to identify influential predictors.

Methodology

• Data Source: The study utilized data from the Chinese TaTME Registry Collaborative (CTRC), a nationwide web-based registry established in 2017. The cohort included 1,625 patients with rectal cancer who underwent primary taTME between November 2017 and July 2024 across 32 institutions.
• Outcome: The primary outcome was pathological CRM positivity, defined as the presence of tumor cells or malignant lymph nodes within ≤1 mm of the CRM. The observed positivity rate was 2.77% (45 cases).
• Predictors: Twenty-one variables were considered, including demographics (age, BMI, sex), preoperative imaging (tumor location, clinical T/N stage, threatened MRF, extramural vascular invasion), treatment history (neoadjuvant therapy), intraoperative factors (expert supervision, surgical approach, conversion, intersphincteric resection [ISR], intraoperative difficulties), and postoperative pathology (pathological T/N stage).
• Data Preprocessing: Missing data (11.69% overall) were handled using random forest-based multiple imputation. To prevent data leakage, imputation was performed after splitting the dataset into training (70%) and validation (30%) cohorts using stratified sampling to preserve the rare outcome distribution.
• Model Development: Nine supervised machine learning algorithms were evaluated: Logistic Regression (LR), Elastic Net (ENET), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Radial Support Vector Machine (RSVM), Multilayer Perceptron (MLP), LightGBM, and k-Nearest Neighbors (KNN). Hyperparameters were optimized via grid search with five-fold cross-validation.
• Evaluation: Performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, specificity, and other metrics. Model interpretation was performed using SHapley Additive exPlanations (SHAP) to quantify variable contributions.

Key Results

• Model Performance: In the validation cohort, the Radial Support Vector Machine (RSVM) model demonstrated the highest apparent AUC of 0.842 (95% CI: 0.696–0.989). Logistic Regression showed comparable discrimination with an AUC of 0.819. Other models, such as Random Forest and KNN, exhibited high training AUCs (>0.90) but lower validation performance, indicating overfitting in this rare-event setting.
• Variable Importance: SHAP analysis on the RSVM model identified the three most influential predictors of CRM positivity:
  1. Pathological N stage.
  2. Threatened mesorectal fascia (MRF) on preoperative MRI.
  3. Intersphincteric resection (ISR).
• Clinical Associations: Patients with CRM positivity were more likely to have higher clinical and pathological T/N stages, MRF involvement, and EMVI positivity on MRI. Surgeries performed without expert supervision were associated with higher CRM positivity, while ISR and robotic surgery showed higher rates of CRM positivity in univariate analysis.
• Inter-center Variation: Significant heterogeneity in CRM positivity rates was observed across the 32 participating centers (p < 0.001), though no significant correlation was found between center case volume and CRM positivity rates.

Significance and Claims
The authors position this study as an exploratory analysis rather than a definitive clinical prediction tool. The primary significance lies in:

1. Validation of Known Factors: Confirming within a large Chinese multicenter registry that clinically plausible factors (threatened MRF, advanced nodal stage, surgical complexity via ISR) are strongly associated with CRM positivity.
2. Methodological Exploration: Demonstrating the application of multiple ML algorithms and SHAP interpretation to a real-world taTME registry, highlighting the challenges of modeling rare outcomes (low event-per-predictor ratio) and the risk of overfitting.
3. Limitations and Scope: The authors explicitly state that the models included intraoperative and postoperative variables (e.g., pathological stage, ISR, intraoperative difficulties), rendering them unsuitable as preoperative decision-support tools. The low positive predictive value (PPV ~0.10 in validation) further limits immediate clinical utility for treatment escalation.
4. Future Directions: The study concludes that while ML can integrate complex variables, its value over standard multidisciplinary assessment remains unproven. Future research is required to develop models using exclusively preoperative variables, increase the number of CRM-positive events, and perform independent external validation before clinical application.

The paper does not claim to have discovered novel predictors but rather to have quantitatively integrated known risk factors within a large registry framework to better understand their relative contributions in the context of taTME.