Abstract
PURPOSE: This study aims to develop and validate an explainable machine learning model for predicting postoperative survival in patients with locally advanced gastric cancer (LAGC), optimizing predictive accuracy while ensuring clinical applicability to facilitate personalized prognostication for patients. METHODS: The study utilized data from 8616 LAGC patients who underwent gastrectomy (2004-2015) in the Surveillance, Epidemiology, and End Results (SEER) database for model development and validation, with external validation performed using 235 postoperative LAGC cases (2016-2022) from Maoming People's Hospital (Maoming, China). Five predictive models-Cox proportional hazards model (CoxPH), random survival forest (RSF), extreme gradient boosting (XGBoost), gradient boosting machine (GBM), and DeepSurv-were developed using the training set. Model performance was evaluated using the concordance index (C-index), area under the receiver operating characteristic curve (AUROC), and Brier score. Additionally, 1-, 3-, and 5-year receiver operating characteristic curves (ROC), calibration curves, and decision curve analysis (DCA) were employed for further assessment. The optimal model was interpreted using explainability tools such as SurvSHAP and SurvLIME. Finally, an interactive prediction tool was created to provide personalized survival evaluation for LAGC patients. RESULTS: RSF exhibited the highest predictive performance, with a C-index of 0.732 (95% CI: 0.720-0.745) in the validation set and 0.723 (95% CI: 0.696-0.755) in the external validation set. The 1-, 3-, and 5-year AUROCs were 0.771, 0.803, and 0.809 in the validation set, and 0.802, 0.711, and 0.721 in the external validation set. Explainability analysis identified lymph node ratio (LNR), AJCC stage, and age as the most influential prognostic factors. An interactive prediction tool was developed to provide individualized prognosis visualization. CONCLUSION: This study developed an RSF-based model to predict postoperative survival in LAGC patients, emphasizing the prognostic significance of LNR, AJCC stage, and age. The interactive prediction tool enhances clinical utility, facilitating personalized treatment decision-making for physicians.