Preoperative prediction of early recurrence in pancreatic cancer: A novel clinical-radiomics model.

Abstract

Early postoperative recurrence critically impacts pancreatic ductal adenocarcinoma prognosis, yet comprehensive preoperative prediction models remain underexplored. In this two-center retrospective study of 895 treatment-naïve PDAC patients who underwent direct resection (training n = 567; internal validation n = 241; external validation n = 87), we defined early recurrence as tumor relapse within 6 months of surgery. We first built a clinical model using logistic regression to select clinical variables and a radiomics model by applying LASSO regression to features extracted from preoperative CT images, then combined these into an integrated clinical-radiomics model via logistic regression. Of the 895 patients (64.4% male; mean age 64.4 ± 8.7 years), 213 (23.8%) experienced early recurrence. Four clinical variables (gender, CA125, radiologic N stage, adjuvant treatment) and 29 radiomics features were selected for the final model, which achieved area under the curve values of 0.862 (95% CI 0.828-0.896) in the training cohort, 0.843 (0.785-0.901) in internal validation, and 0.848 (0.748-0.949) in external validation-each outperforming either the clinical or radiomics model alone. Stratified analyses confirmed robustness across subgroups, and patients classified as high risk by the model had significantly shorter disease-free and overall survival (both p < .001). This clinical-radiomics model offers a preoperative tool to identify PDAC patients at high risk of early postoperative recurrence, thereby supporting personalized treatment planning beyond immediate surgery.