Pre-treatment prediction of EBV-DNA status in nasopharyngeal carcinoma using CE-T1WI radiomics combined with clinical features: a multi-sequence MRI radiomics study

Epstein-Barr virus (EBV) DNA is a critical biomarker for nasopharyngeal carcinoma (NPC) diagnosis, treatment response, and prognosis. Noninvasive pre-treatment prediction of EBV-DNA status using magnetic resonance imaging (MRI) radiomics may provide additional clinical value. We aimed to develop and validate an MRI radiomics-based model for predicting EBV-DNA status in NPC patients using a fully leakage-free, nested validation framework. A total of 277 NPC patients with complete multi-sequence MRI (T1WI, T2WI, and contrast-enhanced T1WI [CE-T1WI]) and EBV-DNA measurements were retrospectively enrolled from a public dataset, and 2553 radiomics features were extracted per patient. Feature stability was assessed with the intraclass correlation coefficient (ICC > = 0.75) in 30 patients with repeated tumor segmentation. The complete modeling pipeline—median imputation, z-score standardization, feature filtering, LASSO feature selection, penalized logistic regression fitting, and decision-threshold selection—was nested within repeated stratified 5-fold cross-validation (10 repeats), so that every outer test fold remained fully unseen. Multicollinearity among clinical variables was assessed using variance inflation factors (VIF), and a parsimonious clinical baseline (age, sex, N stage, VCA-IgA status) was prespecified. Robustness to scanner and voxel-geometry heterogeneity was evaluated using scanner-stratified analyses, isotropic 1-mm voxel resampling, and ComBat harmonization fitted within training folds only. Calibration was optimized with Platt recalibration within folds. Incremental value over the clinical baseline was assessed with the DeLong test, Brier score, and integrated discrimination improvement (IDI). Prognostic value was evaluated in 251 patients with available progression-free survival (PFS) using multivariable Cox regression with multiplicity control. The combined CE-T1WI radiomics–clinical model achieved a leakage-free area under the receiver operating characteristic curve (AUC) of 0.801 (95% CI: 0.742–0.854), significantly exceeding the parsimonious clinical baseline (AUC 0.628, 95% CI: 0.563–0.691; ΔAUC 0.173, 95% CI: 0.104–0.242; DeLong p < 0.001) and improving the Brier score from 0.196 to 0.178; IDI 0.052 (95% CI: 0.027–0.078, p < 0.001). It outperformed the single-sequence CE-T1WI (AUC 0.694), multi-sequence (AUC 0.671), T2WI (AUC 0.636), and T1WI (AUC 0.587) models. Performance was consistent across scanner subgroups (AUCs 0.762–0.812), with 1-mm isotropic resampling (AUC 0.792), and with ComBat harmonization (AUC 0.788). After Platt recalibration, the calibration slope was 1.00 (95% CI: 0.81–1.19), and decision curve analysis showed net benefit over threshold probabilities of approximately 0.10–0.55. The multi-sequence Rad-score was independently associated with PFS (adjusted hazard ratio per 1-SD increment 1.38, 95% CI: 1.05–1.82, p = 0.021). With leakage-free nested validation, CE-T1WI radiomics combined with a parsimonious clinical baseline provided moderate, scanner-robust discrimination of pre-treatment EBV-DNA status in NPC, with established incremental value over clinical variables alone. External validation in independent cohorts is still required before any clinical application.