Integrating radiomic features and clinical variables from cervical lymph node CT to predict treatment response in advanced-stage nasopharyngeal carcinoma

Abstract

To develop a combined radiomic and clinical model based on pretreatment contrast-enhanced computed tomography (CECT) images of metastatic cervical lymph nodes(CLNs) for predicting complete response (CR) versus non-CR in patients with N2–N3M0 nasopharyngeal carcinoma (NPC). This retrospective study included 493 patients with N2–N3M0 NPC, who were stratified by treatment response and divided into a development cohort ( n  = 345) and a held-out validation cohort ( n  = 148). Radiomic features were extracted from manually delineated metastatic CLNs on pretreatment CECT images. The clinical model was constructed using ridge logistic regression. Following reproducibility assessment and least absolute shrinkage and selection operator–based(LASSO) feature selection, six radiomics classifiers were compared. The combined model integrated clinical and radiomics predictions using logistic stacking. Model development and hyperparameter tuning were conducted using nested stratified five-fold cross-validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), bootstrap confidence intervals, and DeLong tests. Gaussian naive Bayes was selected as the radiomics classifier based on the development nested out-of-fold AUC. In the held-out validation cohort, the clinical, radiomics, and combined models achieved AUCs of 0.695 (95% confidence interval [CI], 0.612–0.776), 0.805 (95% CI, 0.733–0.873), and 0.806 (95% CI, 0.729–0.872), respectively. The radiomics and combined models outperformed the clinical model (Holm-adjusted p  = 0.027 and 0.018, respectively), whereas the combined model did not outperform the radiomics model ( p  = 0.831). Pretreatment CECT radiomics of metastatic CLNs provided better discrimination of CR versus non-CR than clinical variables alone. Combining clinical and radiomics predictions produced no material improvement over the radiomics model, warranting further external validation.