Establishing Pelvimetry-Based Machine Learning Models to Predict Surgical Difficulty For Laparoscopic Intersphincteric Resection: a retrospective cohort study

Background Intersphincteric resection (ISR) is a procedure aimed at preserving the anus during the treatment of ultra-low rectal cancer (ULRC). However, the absence of an effective predictive model for selectively identifying patients suitable for laparoscopic ISR (LISR) operations persists, primarily owing to factors related to pelvic anatomy. Methods The present study encompassed individuals diagnosed with ULRC, who underwent LISR between January 2017 and August 2022. These ULRC patients were stratified into difficult or non-difficult LISR groups using recognized and widely accepted scoring criteria. Following the identification of crucial variables, five machine learning (ML) models—Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), and the Random Forest (RF) algorithm—were employed to predict surgical difficulty for LISR. Ultimately, area Under the Curve (AUC) and other indices were utilized to evaluate the predictive performance of these ML models. Results In adherence to the inclusion and exclusion criteria, finally, 163 patients diagnosed with ULRC were included in the present study. Among these, 36 patients (22.1%) were categorized into the difficult ISR group, while the remaining 127 patients (77.9%) were classified as belonging to the non-difficult ISR group. Using Lasso regression and binary logistic regression analysis, nine variables were selected for constructing the ML model. To enhance the reliability and predictive accuracy of this study, we established two types of ML models, each incorporating either nine variables or all variables, respectively. One category preserved all variables, with RF yielding the best performance (accuracy: 0.878, PPV: 1, NPV: 0.867, sensitivity: 0.4, specificity: 1, AUC: 0.877), while the other category retained the screened 9 variables, with SVM demonstrating superior performance (accuracy: 0.857, PPV: 0.636, NPV: 0.921, sensitivity: 0.7, specificity: 0.897, AUC: 0.854). Conclusions The proposed ML model offers a dependable and precise approach to classify ULRC patients undergoing LISR using preoperative pelvimetry imaging data. These models can assist clinicians in better preparing for challenging ISR cases and optimizing treatment plans for individual ULRC patients.