Interpretable machine learning model to predict surgical difficulty in laparoscopic resection for rectal cancer

• DOI: 10.3389/fonc.2024.1337219
• Publication Date: 2024-02-06T17:50:09Z
• Authors: Miao Yu; Zihan Yuan; Ruijie Li; Bo Shi; Daiwei Wan; Xiaoqiang Dong
• Keywords: Lasso; Gradient boosting; Boosting

Background Laparoscopic total mesorectal excision (LaTME) is standard surgical methods for rectal cancer, and LaTME operation a challenging procedure. This study intended to use machine learning develop validate prediction models difficulty of in patients with cancer compare these models’ performance. Methods We retrospectively collected the preoperative clinical MRI pelvimetry parameter who underwent laparoscopic resection from 2017 2022. The was defined according scoring criteria reported by Escal. Patients were randomly divided into training group (80%) test (20%). selected independent influencing features using least absolute shrinkage selection operator (LASSO) multivariate logistic regression method. Adopt synthetic minority oversampling technique (SMOTE) alleviate class imbalance problem. Six model developed: light gradient boosting (LGBM); categorical (CatBoost); extreme boost (XGBoost), (LR); random forests (RF); multilayer perceptron (MLP). area under receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity F1 score used evaluate performance model. Shapley Additive Explanations (SHAP) analysis provided interpretation best Further decision (DCA) manifestations Results A 626 included. LASSO shows that tumor height, prognostic nutrition index (PNI), pelvic inlet, outlet, sacrococcygeal distance, fat angle 5 (the between apex sacral lower edge pubic bone) are predictor variables In addition, correlation heatmap there no significant seven variables. When predicting surgery, XGBoost performed among six (AUROC=0.855). Based on results, also superior, feature importance height most important variable factors. Conclusions developed an predict surgery. can help clinicians quickly accurately surgery adopt individualized methods.