Prediction of efficacy and prognosis of PD‑1/PD‑L1 inhibitor combination chemotherapy for gastric cancer using an AI model based on dual‑energy CT: A multicenter study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, the First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China; International Joint Laboratory of Medical Imaging of Henan Province, Zhengzhou 450052, Henan, China; Key Laboratory of Digestive Tumor Imaging of Henan Province, Zhengzhou 450052, Henan, China; Key Laboratory of CT Imaging of Henan Province, Zhengzhou, 450052, Henan, China.
- Henan Provincial Cancer Hospital, Zhengzhou, 450008, Henan, China.
- Henan Provincial Cancer Hospital, Zhengzhou, 450008, Henan, China. Electronic address: [email protected].
- Department of Radiology, the First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China; International Joint Laboratory of Medical Imaging of Henan Province, Zhengzhou 450052, Henan, China; Key Laboratory of Digestive Tumor Imaging of Henan Province, Zhengzhou 450052, Henan, China; Key Laboratory of CT Imaging of Henan Province, Zhengzhou, 450052, Henan, China. Electronic address: [email protected].
Abstract
The aim of this study was to evaluate the ability of dual-energy CT (DECT) quantitative parameters and radiomics to predict treatment efficacy and prognosis in gastric cancer patients receiving PD-1/PD-L1 inhibitor-based combination chemotherapy. This study enrolled a total of 344 patients with gastric adenocarcinoma, divided into training, validation, external-validation, and cross-platform validation sets. Based on the Immune Response Evaluation Criteria in Solid Tumors (iRECIST) for immunotherapy, patients were classified as responders or non-responders. Clinical features and qualitative and quantitative imaging biomarkers from conventional CT and DECT were evaluated. Five machine learning algorithms were used to integrate imaging biomarker models, independent clinical predictors, and DECT models. A logistic regression model was then constructed based on the integrated data. Model performance was assessed using receiver operating characteristic (ROC) analysis, calibration curves, decision curve analysis (DCA), and the IDI and NRI indices, with subgroup analysis performed to validate the model's generalizability. Kaplan-Meier (KM) survival analysis was conducted to evaluate the association between imaging biomarkers and progression-free survival (PFS) and overall survival (OS). The logistic regression model was identified as the optimal machine learning model, with AUCs of 0.827 (95% CI: 0.764-0.891), 0.803 (95% CI: 0.695-0.910), 0.775 (95% CI: 0.641-0.909), and 0.781 (95% CI: 0.614-0.948) across the four sets. No significant differences in AUC were observed between the training set and the other sets (DeLong test: P = 0.698, 0.488, and 0.646), indicating stable model performance. Calibration curves indicated good agreement between predicted and observed outcomes across all sets, with Hosmer-Lemeshow test P-values of 0.291, 0.162, 0.918, and 0.281 for the training, validation, test 1, and external test 2 sets, respectively. Brier scores (0.158-0.197) further confirmed high predictive accuracy. DCA showed positive net benefit across wide threshold ranges: 0.02-0.95 (training), 0.19-0.79 (validation), 0.04-0.66 (test 1), and 0.03-0.76 (external test 2). At a 50% threshold, net benefits were 0.160, 0.218, 0.120, and 0.314. Subgroup analyses confirmed robust generalizability (all P > 0.05). KM curves demonstrated that patients in the high-risk group had significantly worse OS and PFS compared with those in the low-risk group. Similarly, patients classified as non-responders by iRECIST had significantly poorer OS and PFS than responders (all P < 0.05, χ<sup>2</sup> = 3.98-13.67). The DECT-based radiomics model can predict treatment response and prognosis in gastric cancer patients receiving immunotherapy combined with chemotherapy and support individualized treatment planning and optimization of therapeutic strategies.