The present single-center retrospective cohort study included 118 patients with T3-4 NPC who received definitive IMRT at Guilin People's Hospital (Guilin, China) between March 2020 and June 2025. All patients had pathologically confirmed NPC and underwent pretreatment staging examinations, including contrast-enhanced magnetic resonance imaging (MRI) of the nasopharynx and neck, chest imaging, abdominal imaging and bone assessment using bone scintigraphy, skeletal MRI or positron emission tomography/computed tomography (CT), as clinically indicated. As the present study period spanned different TNM classification periods, all patients were retrospectively restaged according to the 9th edition of the American Joint Committee on Cancer/Union for International Cancer Control TNM classification system before statistical analysis using available pretreatment clinical, imaging and pathological information.

The inclusion criteria were as follows: i) Pathologically confirmed NPC; ii) T3-4 disease; iii) receipt of definitive IMRT; iv) availability of complete pretreatment imaging and radiotherapy planning data for GTVnx assessment; and v) available follow-up data for OS. The exclusion criteria were as follows: i) Distant metastasis at initial diagnosis; ii) previous head and neck radiotherapy or treatment for another malignancy; iii) concurrent active malignancy; and iv) missing key clinical, radiotherapy planning or follow-up data. The study protocol was approved by the Ethics Committee of Guilin People's Hospital (Guilin, China; approval no. 2025-316KY) and written informed consent for the use of medical data for research purposes was obtained from all patients at admission.

All patients received definitive IMRT. GTVnx was manually delineated by radiation oncologists based on pretreatment contrast-enhanced MRI, simulation CT images and clinical examination findings. The delineation included all radiologically and clinically visible primary nasopharyngeal lesions and was performed according to institutional target delineation protocols for NPC. All GTVnx contours were independently reviewed and confirmed by two senior radiation oncologists before radiotherapy planning to minimize inter-observer variability. The GTVnx volume was automatically calculated by the radiotherapy planning system and recorded in cm³. Cervical nodal target volumes, clinical target volumes and organs at risk were delineated according to the institutional head and neck radiotherapy protocol. All patients received platinum-based combined treatment, including induction chemotherapy, concurrent chemoradiotherapy, concurrent targeted therapy, adjuvant chemotherapy and maintenance chemotherapy when indicated. The specific treatment regimen was determined by the attending physician according to disease stage, performance status and organ function. None of the patients received immunotherapy during the study period.

Data on age, sex, histological subtype, T stage, N stage, clinical stage, pretreatment EBV DNA level, radiotherapy dose, GTVnx and follow-up outcomes were collected. The pretreatment plasma EBV DNA level was dichotomized using a cut-off value of 400 copies/ml, consistent with previous studies and commonly used clinical thresholds for risk stratification in NPC (18). OS was defined as the time from the start of radiotherapy to death from any cause. Patients who were alive at the last follow-up were censored.

All statistical analyses were performed using R software (version 4.5.0; R Foundation for Statistical Computing) and SPSS (version 25.0; IBM Corp.). Continuous variables were first assessed for normality. Normally distributed variables were presented as the mean ± standard deviation and were compared using the independent-samples t-test. Non-normally distributed variables were presented as the median (first quartile, third quartile) and were compared using the Mann-Whitney U-test. Categorical variables were summarized as counts and percentages and were compared using the χ² test or Fisher's exact test.

A 36-month time-dependent ROC curve was used to evaluate the ability of GTVnx to discriminate patients according to their OS outcomes. The 36-month time-point was selected based on the maturity of follow-up data as it approximated the median follow-up duration of the cohort (34.1 months) and represented an intermediate-term survival endpoint. The variability of the cut-off value was further assessed using bootstrap resampling with 1,000 repetitions, and the median bootstrap-derived cut-off value and 95% percentile interval were calculated. Patients were then divided into low- and high-GTVnx groups based on this cut-off. OS curves were generated using the Kaplan-Meier method and differences between groups were compared using the log-rank test. Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs).

Given the limited number of OS events (n=18), the events-per-variable ratio was below the commonly recommended threshold of 10 events per variable. Therefore, multivariable Cox regression models and incremental model performance analyses were considered exploratory and hypothesis-generating. Variables included in the multivariable models were selected a priori based on established clinical relevance and prognostic significance rather than solely on univariable statistical significance. The clinical model included T stage and pretreatment EBV DNA, and the incremental prognostic value of GTVnx was subsequently evaluated by adding GTVnx to this model. Considering the risk of model overfitting caused by limited outcome events, the N stage was not incorporated into the primary model and was examined separately in sensitivity analyses. Model performance was compared using the C-index, AIC and the 36-month time-dependent area under the curve (AUC). In addition, bootstrap resampling was used to estimate the CI of the difference in C-index between the clinical and extended models.

To assess the robustness of the findings, several sensitivity analyses were performed as follows: i) Further adjustment for N stage based on the main model; ii) replacement of T stage with overall clinical stage; iii) inclusion of GTVnx as a continuous variable in the model, with the HR calculated per 10 cm³ increase; and iv) use of a conventional ROC-derived alternative cut-off to repeat the Kaplan-Meier survival analysis. A two-sided P<0.05 was considered statistically significant.

A total of 118 patients with T3-4 NPC were included in the present study. The median age was 53 years (range, 23–75 years), and the cohort included 84 male and 34 female patients (Table I). Based on the 78 cm³ cut-off determined by the 36-month time-dependent ROC analysis (Fig. 1), patients were divided into the low-GTVnx group (GTVnx ≤78 cm³; n=80) and the high-GTVnx group (GTVnx >78 cm³; n=38). Compared with the low-GTVnx group, patients in the high-GTVnx group had significantly higher proportions of male patients, T4 disease, stage III disease and pretreatment EBV DNA ≥400 copies/ml. There were no statistically significant differences between the two groups in terms of age, age category, N stage, histological subtype, follow-up duration, loss to follow-up rate, concurrent targeted therapy, adjuvant chemotherapy or capecitabine/S-1 maintenance chemotherapy (all P>0.05). The baseline characteristics are shown in Table I.

The median (25th, 75th percentile) follow-up time for the overall cohort was 34.1 (20.4, 49.3) months. There was no statistically significant difference in median follow-up time between the low- and high-GTVnx groups. During follow-up, 18 deaths occurred (15.3%). The proportion of OS events was significantly higher in the high-GTVnx group than in the low-GTVnx group (31.6 vs. 7.5%; P=0.002). The 3-year OS rate was lower in the high-GTVnx group than in the low-GTVnx group (74.1 vs. 95.1%; P<0.001).

GTVnx showed a right-skewed distribution in the overall cohort, with a median value of 60.5 cm³ and first and third quartiles of 39.2 and 93.9 cm³, respectively. Time-dependent ROC analysis for 3-year OS showed that GTVnx had discriminatory ability for predicting 3-year OS, with an AUC of 0.709. Based on the Youden index, the optimal cut-off value for GTVnx was 78 cm³, with a sensitivity of 69.9% and a specificity of 72.9%. The distribution of GTVnx and the time-dependent ROC curve are shown in Fig. 1. The bootstrap assessment of cut-off variability is presented in Table SI.

Univariable Cox regression analysis showed that stage III disease, EBV DNA ≥400 copies/ml and high GTVnx were significantly associated with poorer OS. Multivariable Cox regression analysis further showed that, after adjustment for T stage and EBV DNA, high GTVnx remained independently associated with unfavorable OS (HR, 4.375; 95% CI, 1.474–12.986; P=0.008). The results of the univariable and multivariable Cox analyses are shown in Table II.

In the same 118 patients with 18 deaths, the extended model incorporating GTVnx showed improved predictive performance compared with the clinical model including only T stage and EBV DNA. The C-index increased from 0.759 to 0.817, with a bootstrap-estimated difference in C-index of 0.052 (95% CI, −0.009–0.127). The 36-month AUC increased from 0.748 to 0.787, and the AIC decreased from 135.151 to 133.085 after adding GTVnx, with a ΔAIC of −2.066 (Table III).

After further adjustment for N stage based on the main model, the association between GTVnx and poorer OS remained consistent. Replacing T stage with overall clinical stage did not materially change the results. When GTVnx was included as a continuous variable, each 10 cm³ increase in GTVnx remained associated with an increased risk of death. In addition, after using a conventional ROC-derived cut-off of 106 cm³ as an alternative threshold for regrouping, the high-GTVnx group still showed a trend toward poorer OS. Detailed results are provided in Table SII and Fig. S1. Overall, the sensitivity analyses supported the robustness of the main findings.

Discussion
The present study examined whether GTVnx could provide additional prognostic information in patients with T3-4 NPC. The main finding was that patients with a larger GTVnx had poorer OS, and this association persisted after adjustment for T stage and pretreatment EBV DNA. When GTVnx was added to a clinical model based on these conventional factors, the C-index, AIC and time-dependent AUC all changed in a favorable direction, although the absolute improvement was modest. Similar results were obtained across sensitivity analyses using alternative model specifications and variable definitions. These findings suggest that GTVnx may help refine prognostic assessment within the T3-4 subgroup, although further validation is required before clinical application.

TNM staging remains the foundation for prognostic assessment and treatment decision-making in NPC (4). However, T stage is primarily based on the anatomic extent of tumor invasion and does not directly quantify tumor burden (19). In clinical practice, patients classified as T3-4 may have substantially different primary tumor volumes. Certain tumors are assigned to an advanced T category due to the involvement of critical anatomic structures, even when the measurable primary volume is relatively limited. By contrast, other tumors within the same T category may show more extensive infiltration and a much larger volume (19,20). As a continuous volumetric parameter routinely derived from radiotherapy planning systems, GTVnx may capture part of this heterogeneity (19). The association between larger GTVnx and poorer OS suggests that tumor volume provides additional information regarding disease burden.

Previous studies have linked primary tumor volume with locoregional control, distant metastasis and survival outcomes in NPC (9,21–23). Most of these studies evaluated mixed-stage cohorts or the broader locoregionally advanced population (9,23). The present study stands out by focusing specifically on patients with T3-4 disease, a subgroup already considered anatomically high-risk. The present findings indicate that tumor volume may still distinguish prognostic differences within this advanced T-stage category (24,25). It may serve as a simple quantitative adjunct to conventional staging, particularly when patients with the same T stage show markedly different primary tumor burdens.

The incremental predictive value of GTVnx should also be interpreted with caution (26–28). Although adding GTVnx to the clinical model resulted in an improvement in model performance, the magnitude of incremental improvement was modest. The bootstrap CI for the difference in C-index included zero, suggesting uncertainty regarding the magnitude of improvement. This may be partly because the clinical model already included established prognostic factors, such as T stage and pretreatment EBV DNA (18). Although N stage showed a clinically meaningful HR, it was not included in the primary model due to the limited number of OS events and the potential risk of overfitting. However, sensitivity analysis incorporating N stage demonstrated consistent results, suggesting that the prognostic association of GTVnx remained consistent. In addition, GTVnx and T stage are not completely independent, as larger tumors are more likely to involve adjacent anatomic structures. These overlapping features may limit the additional contribution of GTVnx when it is combined with established clinical variables. Therefore, GTVnx should be viewed as supplementary risk information rather than as a stand-alone basis for treatment selection. For patients with a larger GTVnx, clinical interpretation should still integrate T stage, N stage, EBV DNA, treatment response and the general condition of the patient (29,30).

Several limitations should be acknowledged. First, this was a single-center retrospective study, and selection bias and unmeasured confounding cannot be excluded. Second, the relatively small number of OS events may have limited the statistical power and stability of multivariable modeling. Therefore, the prognostic value of GTVnx requires further evaluation. Third, the 78 cm³ cut-off identified by time-dependent ROC analysis was derived from the present cohort and should be considered an exploratory threshold rather than a generally applicable clinical threshold. Although bootstrap analysis showed good agreement between the original cut-off and the median bootstrap-derived cut-off, the relatively wide percentile interval indicated uncertainty regarding the precision of the cut-off estimate. Therefore, this cut-off requires validation in larger independent cohorts before clinical application. Fourth, the present study focused on OS and did not systematically assess progression-free survival, locoregional recurrence-free survival or distant metastasis-free survival. Finally, dynamic EBV DNA, response to induction chemotherapy, radiomic features and molecular biomarkers were not included. Larger multicenter studies with external validation are needed to confirm the stability of GTVnx and to determine how it can be integrated with other prognostic factors.

In summary, in patients with T3-4 NPC, GTVnx was associated with poorer OS and may provide supplementary prognostic information beyond T stage and pretreatment EBV DNA. The 78 cm³ cut-off derived from time-dependent ROC analysis helped identify patients at higher risk, and adding GTVnx to the clinical model modestly improved model performance. GTVnx should be regarded as an auxiliary quantitative parameter rather than an alternative to TNM staging. Further multicenter studies with larger sample sizes and external validation are needed to confirm these findings.