Abstract Objectives The incidence and mortality rates of gastric cancer remain notably high on a global scale. Early prediction of recurrence is essential for tailoring personalized treatment plans. This study intends to stratify the risk of resectable gastric cancer via the establishment of a five-year recurrence prediction model for gastric cancer, thereby more effectively guiding postoperative follow-up and treatment strategies. Methods In this retrospective study, a total of 785 patients with gastric cancer at two hospitals were included. We developed multiple recurrence prediction models by integrating radiomics features, deep features, and clinical parameters through deep learning techniques. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, true positive rate, true negative rate, positive predictive value, and negative predictive value. Results The comprehensive model, which integrated radiomic features, deep features, and clinical features, achieved AUCs of 0.96 and 0.83 in the training and internal validation sets, respectively, and 0.77 in the external test set. Feature importance analysis showed that variables such as Age, pN stage, HER2 status, radiomic risk score (RS-Rad), and deep learning risk score (RS-DL) were highly significant across multiple models. The primary clinical advantage of the proposed model lies in its very high negative predictive value (NPV>92 %), which allows for the accurate identification of patients at low risk of recurrence. Conclusions This model integrating radiomic features, deep features, and clinical variables may serve as a potentially useful tool for stratifying patients at relatively low risk of gastric cancer recurrence, with acceptable predictive performance. It could provide a supportive reference for the potential de-escalation of unnecessary treatment and the optimization of follow-up resource allocation in clinical practice.
Tong et al. (Thu,) studied this question.