Radiogenomic biomarkers for identifying surgery-sparing candidates in locally advanced gastric or gastro-esophageal junction cancer treated with perioperative immunochemotherapy: A discovery and validation study.

Abstract

4058 Background: Perioperative immunochemotherapy has markedly increased pathologic complete remission (pCR) rates in patients with locally advanced gastric cancer or gastroesophageal junction cancer (LAGC/GEJC), raising the possibility of surgery-sparing strategies in selected patients. However, reliable non-invasive tools to identify pCR preoperatively are lacking. This study aimed to develop and validate a radiogenomic model integrating CT radiomics and peripheral blood single nucleotide polymorphism (SNP) data to predict pCR. Methods: This retrospective multicenter study included 642 patients with LAGC/GEJC from 14 institutions who received perioperative immunochemotherapy. Patients were assigned to a primary cohort (n = 442; training n = 309, internal validation n = 133) and an external validation cohort (n = 200). Radiomic features were extracted from preoperative CT images, and SNPs were genotyped from peripheral blood. Feature selection was performed using LASSO, generating radiomic (Rad-score) and genomic (Gen-score) signatures. A combined radiogenomic model incorporating Rad-score, Gen-score, and clinical variables was developed using multivariable logistic regression. Model performance was evaluated using AUC and decision curve analysis (DCA). Results: Fourteen radiomic features and eight SNPs were selected. The radiogenomic model consistently outperformed single-modality models across cohorts. In the training set, the combined model achieved an AUC of 0.915 (95% CI 0.880–0.950), exceeding Rad-score (AUC 0.832) and Gen-score (AUC 0.798; both P < 0.001). In the internal validation set, the AUC was 0.882 (95% CI 0.825–0.939), and in the external validation set, 0.852 (95% CI 0.795–0.909), remaining superior to either modality alone. DCA demonstrated greater net clinical benefit of the combined model. At the optimal cutoff, 91.5% of true pCR cases were correctly identified in the validation cohort. Conclusions: A radiogenomic model integrating CT radiomics and peripheral blood SNP signatures enables accurate, non-invasive prediction of pCR in LAGC/GEJC patients undergoing perioperative immunochemotherapy and may support identification of candidates for surgery-sparing management.