Nasopharyngeal carcinoma (NPC) is a malignancy of the head and neck region with high prevalence in east and Southeast Asia, and its development is strongly linked to Epstein–Barr virus (EBV) infection (1, 2). Although intensity-modulated radiotherapy (IMRT) combined with chemotherapy has become the standard treatment for advanced-stage disease, recurrence and metastasis remain major clinical challenges (3). Plasma EBV DNA has emerged as a valuable biomarker for estimating tumor burden and predicting prognosis, however, its accuracy and consistency in different clinical scenarios are still debated (4, 5).

Conventional regression-based approaches may not fully uncover the intricate relationships between EBV levels and various prognostic factors. In contrast, artificial intelligence (AI) techniques, especially machine learning models, provide a powerful framework for handling complex datasets and capturing nonlinear associations (6–8). These models have shown increasing promise in cancer prognosis and treatment optimization (9, 10). Nonetheless, few studies have systematically evaluated the integration of EBV DNA into AI-based prediction tools for NPC.

In this study, we propose to develop a robust and interpretable AI model incorporating EBV DNA levels and clinical characteristics to enhance survival prediction in patients undergoing concurrent chemoradiotherapy. This model aims to assist clinicians in tailoring personalized treatment strategies and long-term monitoring plans.

In recent years, EBV DNA has emerged as a key biomarker in the management of NPC, particularly in regions with a high disease burden such as Southeast Asia (11–13). While its prognostic relevance is well established, translating EBV DNA levels into individualized survival predictions remains a challenge. In this study, we sought to address this gap by constructing AI–based models that integrate EBV DNA with standard clinical features to predict PFS in advanced NPC patients treated with concurrent CRT.

Unlike traditional statistical methods, which typically assume linear associations and fixed hazard ratios, AI algorithms are capable of modeling more complex, nonlinear interactions (14, 15). This advantage is particularly relevant in heterogeneous cancers such as NPC, which clinical behavior is influenced by both tumor burden and host factors (16). By incorporating machine learning techniques—namely Cox regression, LASSO, RSF, and GBM—we aimed to compare the predictive capacity of conventional and data-driven approaches (17–20).

Among the four models tested, the RSF algorithm demonstrated the strongest performance (21). It achieved the highest C-index in the training cohort and showed stable results in the validation set. Beyond overall accuracy, interpretability analyses—including SHAP values, time-dependent feature importance, and partial dependence plots—consistently highlighted EBV DNA as the most significant predictor of survival outcomes. These findings reinforce the notion that EBV DNA is not merely a passive biomarker, but an active contributor to risk stratification in AI-driven models.

The partial dependence plot for EBV DNA revealed a clear inflection point: patients with baseline EBV DNA ≥1500 copies/ml consistently showed markedly lower predicted survival probabilities. Interestingly, survival curves for patients with undetectable or low-level EBV DNA copies were similar to those of patients with EBV DNA > 0 and < 1500 copies/ml, suggesting a threshold effect. These observations are in line with previous clinical studies and further validate the value of integration of EBV DNA into automated risk prediction frameworks.

Another strength of this study lies in its focus on model interpretability. One of the major criticisms of AI in healthcare is the “black box” nature of many algorithms (22, 23). By applying tools such as SHAP and PDP, we made our models more transparent and clinically meaningful (24). These visualizations allow clinicians understanding not only that EBV matters, but how and when it exerts the greatest impact on patient outcomes.

In our study, the RSF model demonstrated superior performance compared to other machine learning algorithms. RSF is an ensemble learning method that is particularly well-suited for survival analysis due to its ability to handle both censored data and complex non-linear relationships between predictors and survival outcomes (25, 26). Unlike traditional survival models, RSF does not require the assumption of proportional hazards and can model interactions between variables more effectively. Additionally, RSF can handle high-dimensional datasets with many variables and is less prone to overfitting compared to other algorithms, making it a robust choice for survival prediction in clinical datasets like ours (27).

From a clinical standpoint, integrating EBV DNA into AI-based models offers more than statistical insight—it provides a practical tool for precision medicine (28–30). By accurately identifying patients at higher risk of disease progression, clinicians can consider early treatment intensification, closer surveillance, or inclusion in clinical trials. Conversely, patients with low-risk profiles may benefit from treatment de-escalation, reducing toxicity and preserving quality of life. The model’s transparency also enhances clinical confidence, making it more acceptable for integration into multidisciplinary tumor boards. As such, EBV-informed AI models may serve as decision-support systems that personalize management strategies and improve long-term outcomes in advanced NPC (31–33).

Although the current model demonstrates promising results, integrating other modalities such as radiomics, genomic alterations, and immune-related biomarkers could further improve its prognostic performance. Radiomics could provide detailed imaging features related to tumor phenotype, allowing for more accurate risk stratification. Genomic alterations, including mutations, copy number variations, and methylation patterns, could offer insights into the underlying molecular mechanisms of tumor progression and treatment resistance (34). Additionally, immune-related biomarkers, such as tumor-infiltrating lymphocytes or checkpoint markers, may help capture the immune response to therapy, which is often a critical factor in cancer prognosis (35). Future studies incorporating these multimodal data types, in combination with EBV DNA, would likely improve the model’s accuracy and utility, providing more comprehensive and personalized predictions for patients with advanced NPC.

Despite its strengths, this study has several limitations. First, the retrospective nature of the analysis, even though based on a multicenter cohort, may introduce selection and information bias. Second, while EBV DNA testing was performed uniformly within each center, inter-institutional variability in laboratory procedures cannot be completely ruled out. Third, although treatment protocols were broadly standardized (IMRT with platinum-based concurrent chemotherapy), subtle differences in chemotherapy regimens, supportive care, or radiation planning across centers may contribute to treatment heterogeneity, potentially confounding survival outcomes. Fourth, the model did not incorporate other potentially informative modalities such as radiomics, genomic alterations, or immune-related biomarkers, which could further enhance predictive accuracy. Finally, the absence of external validation limits the generalizability of the model. Future prospective studies should include external datasets and integrate multimodal information to validate and refine AI-based prognostic models for advanced NPC.

In conclusion, our study demonstrates that elevated EBV DNA levels are closely associated with poorer PFS in patients with advanced NPC. By integrating EBV DNA with advanced AI algorithms, particularly random survival forests, we were able to more effectively characterize its prognostic impact across the disease course. These findings highlight the potential of EBV-informed AI models as valuable tools for enhancing risk stratification and guiding personalized treatment decisions in clinical practice.