Procedure-specific prediction of surgical difficulty in laparoscopic and robotic right hemicolectomy: Interpretable machine learning models.
Authors
Affiliations (4)
Affiliations (4)
- Northern Jiangsu People's Hospital, Clinical Teaching Hospital of Medical School, Nanjing University, Yangzhou, China; Northern Jiangsu People's Hospital of Jiangsu Province, China; General Surgery Institute of Yangzhou, Yangzhou University, Yangzhou, 225001, China; Yangzhou Key Laboratory of Basic and Clinical Transformation of Digestive and Metabolic Diseases, China.
- Department of Gastrointestinal Surgery, The First Affiliated Hospital of Jinan University, Guangzhou, Guangdong, China.
- Northern Jiangsu People's Hospital of Jiangsu Province, China; General Surgery Institute of Yangzhou, Yangzhou University, Yangzhou, 225001, China; Yangzhou Key Laboratory of Basic and Clinical Transformation of Digestive and Metabolic Diseases, China; Northern Jiangsu People's Hospital Affiliated to Yangzhou University, China.
- Northern Jiangsu People's Hospital, Clinical Teaching Hospital of Medical School, Nanjing University, Yangzhou, China; Northern Jiangsu People's Hospital of Jiangsu Province, China; General Surgery Institute of Yangzhou, Yangzhou University, Yangzhou, 225001, China; Yangzhou Key Laboratory of Basic and Clinical Transformation of Digestive and Metabolic Diseases, China. Electronic address: [email protected].
Abstract
Laparoscopic (L-RHC) and robotic (R-RHC) right hemicolectomy are standard treatments for colon cancer, but procedure-specific prediction of surgical difficulty remains limited. This study developed interpretable machine learning (ML) models to predict surgical difficulty for both procedures and to support individualized surgical approach selection. Patients with right-sided colon adenocarcinoma who underwent L-RHC or R-RHC between Jan. 2019 and Dec. 2025 were retrospectively analyzed. The two procedures were treated as independent cohorts and randomly split into development and validation sets (7:3). Preoperative clinical variables and CT-derived anatomical metrics were collected. In each development set, LASSO regression was used for feature selection. Nine ML algorithms were trained with 10-fold cross-validation, and a weighted soft-voting ensemble of the five best-performing models was built for each cohort. SHAP was used for interpretation, and web-based calculators were developed. In the L-RHC cohort (n = 419), selected predictors included adiposity-related metrics, Henle's trunk type, presence of the right colonic artery, and high plasma triglycerides. In the R-RHC cohort (n = 215), selected predictors included prior abdominal surgery, adiposity-related metrics, and high plasma triglycerides. In internal validation, the ensemble models achieved AUCs of 0.919 for L-RHC and 0.901 for R-RHC. SHAP showed that adiposity-related metrics were key contributors in both cohorts, while vascular anatomy was more important in L-RHC and prior abdominal surgery in R-RHC. Procedure-specific ensemble ML models using routine clinical and CT-derived variables predicted surgical difficulty in L-RHC and R-RHC with good discrimination. The accompanying web-based calculators may support individualized risk assessment, preoperative planning, and surgical approach selection.