Abstract
OBJECTIVES: This study aimed to develop and validate machine learning (ML) models to predict survival following oesophagectomy in oesophageal squamous cell carcinoma (ESCC) patients using intratumoural and peritumoural radiomic features. METHODS: A retrospective analysis was conducted on ESCC patients with preoperative contrast-enhanced computed tomography who underwent oesophagectomy from June 2016 to January 2020. Patients were randomly assigned to training and test sets (8:2 ratio). Radiomic features were independently extracted from intratumoural and peritumoural regions. Cox regression, random survival forests (RSF), and gradient boosting decision tree (GBDT) were used for modelling. The performance of models was evaluated by discrimination and calibration. RESULTS: The study included 443 patients, 354 in the training set and 89 in the test set. Peritumoural radiomic features predominated in the final selection, with 14 of 17 selected features originating from peritumoural region. The optimal GBDT model (the integrated area under the curve [iAUC]: 0.854; the integrated Brier score [iBS]: 0.160) using dual-region radiomic and clinical features outperformed other models, with a 1-year time-dependent area under the curve (tAUC) of 0.712 (95% CI, 0.655-0.738) and 3-year tAUC of 0.733 (95% CI, 0.655-0.805). It effectively stratified patients into high- and low-risk groups (P < .001). CONCLUSIONS: ML models using intratumoural and peritumoural radiomic features showed potential for predicting postoperative survival in ESCC patients, with the optimal GBDT model demonstrating effective risk stratification.