Multi-Modal Machine Learning for Evaluating the Predictive Value of Pelvimetric Measurements (Pelvimetry) for Anastomotic Leakage After Restorative Low Anterior Resection

Background/Objectives: Anastomotic leakage (AL) remains a major complication after restorative rectal cancer surgery, with accurate preoperative risk stratification posing a significant challenge. Pelvic measurements derived from magnetic resonance imaging (MRI) have been proposed as potential predictors of AL, but their clinical utility remains uncertain. Methods: This retrospective, multicenter cohort study analyzed rectal cancer patients undergoing restorative surgery between 2013 and 2021. Pelvic dimensions were assessed using MRI-based pelvimetry. Univariate and multivariate regression analyses identified independent risk factors for AL. Subsequently, machine Learning (ML) models—logistic regression, random forest classifier, and XGBoost—were developed to predict AL using preoperative clinical data alone and in combination with pelvimetry. Model performance was evaluated using F1 scores, with the area under the receiver operating characteristic (ROC-AUC) and precision–recall curves (AUC-PR) as primary metrics. Results: Among 487 patients, the overall AL rate was 14%. Multivariate regression analysis identified distance to the anorectal junction, pelvic inlet width, and interspinous distance as independent risk factors for AL (p < 0.05). The logistic regression model incorporating pelvimetry achieved the highest predictive performance, with a mean ROC-AUC of 0.70 ± 0.09 and AUC-PR of 0.32 ± 0.10. Although predictive models that included pelvic measurements demonstrated higher ROC-AUCs compared to those without pelvimetry, the improvement was not statistically significant. Conclusions: Pelvic dimensions, specifically pelvic inlet and interspinous distance, were independently associated with an increased risk of AL. While ML models incorporating pelvimetry showed only moderate predictive performance, these measurements should be considered in developing clinical prediction tools for AL to enhance preoperative risk stratification.