Machine-learning CT radiomics for prognostication in unresectable pancreatic cancer

Background: We aimed to develop an interpretable radiomics-clinical model to predict overall survival (OS) in unresectable pancreatic cancer (PC). Methods: In this retrospective cohort, 202 patients with unresectable PC were enrolled. A total of 1,130 radiomics features were extracted from a region of interest encompassing the largest primary lesion using 3D-Slicer. Least absolute shrinkage and selection operator (LASSO)-selected features associated with OS were used to construct a radiomics risk score (RS). Independent clinical predictors were identified through stepwise Cox regression. A nomogram integrating RS with independent clinical predictors was built. Results: Median OS (mOS) for the entire cohort was 20.3 months. From 1,130 baseline CT radiomics features, LASSO retained 12 prognostic descriptors, which were linearly combined to compute a radiomics RS. Stepwise Cox regression identified age, sex, and CA19-9 as independent clinical predictors. A nomogram integrating RS with these variables was constructed in the training set. In the validation set, the area under the receiver operating characteristic curve (AUC) reached 0.804, 0.812, and 0.794 for 1-, 2-, and 3-year OS, respectively. Conclusion: An interpretable radiomics-clinical nomogram provided accurate survival prediction in unresectable pancreatic cancer.