Interpretable AI for Stratifying Risk for Immunoradiotherapy for Locally Advanced Nasopharyngeal Carcinoma

Outcomes for patients with locally advanced nasopharyngeal carcinoma remain variable despite advances in chemoradiotherapy and the growing integration of immune checkpoint inhibitors into treatment paradigms. Although recent phase III trials have demonstrated improvements in event-free survival with immunotherapy, clinicians still lack practical tools to identify which patients are most likely to benefit from these approaches and which may require closer monitoring or alternative strategies.

In this context, Guili Cao, Professor of Oncology at First People's Hospital of Zigong, Zigong Medica Science Academy in Zigong, China, and colleagues, developed an interpretable AI model designed to predict survival and support risk-adapted treatment decision-making for patients with locally advanced nasopharyngeal carcinoma using routinely available clinical variables. The findings of their study were published in Frontiers in Oncology.

Model Methods

This multicenter, retrospective study included 249 patients with locally advanced nasopharyngeal carcinoma who were treated between 2018 and 2025. All patients received definitive chemoradiotherapy, with immunotherapy administered in a subset based on clinical judgment. The data set was divided into a training cohort (70%) and a validation cohort (30%). Investigators first performed univariable Cox regression to identify prognostic variables, ultimately selecting age, tumor grade, and nodal (N) stage for model development.

A Cox proportional hazards–based XGBoost model was then constructed to generate individualized survival risk scores. This machine-learning approach was selected for its ability to capture nonlinear relationships and interactions among variables without requiring prespecified assumptions. Patients were stratified into low- and high-risk groups using the median risk score as a cutoff. Model performance was assessed using time-dependent receiver operating characteristic analysis in the validation cohort.

To address a key barrier to clinical implementation—a lack of transparency—the investigators incorporated SHapley Additive exPlanations (SHAP) to provide both global and patient-level interpretability. This allowed visualization of how each variable contributed to predicted risk, enabling clinicians to better understand the drivers of model outputs.

Key Findings

The AI model demonstrated clinically meaningful predictive performance in the validation cohort, with area under the curve values of 0.784, 0.765, and 0.725 for 1-, 2-, and 3-year overall survival rates, respectively. Risk stratification based on the model clearly separated patients into prognostic groups, with significantly longer survival observed in the low-risk group compared with the high-risk group (P < .001), as illustrated in Kaplan-Meier analyses.

Age emerged as the most influential predictor of survival in the model, followed by N stage and tumor grade. SHAP analyses demonstrated that older age and more advanced nodal disease were consistently associated with higher predicted mortality risk, while lower-grade tumors were associated with more favorable outcomes. The relationship between age and risk was not strictly linear, with risk increasing more prominently beyond approximately 50 to 60 years of age. This finding underscores the advantage of machine-learning approaches over traditional linear models in capturing clinically relevant patterns.

The results also reinforce the clinical significance of nodal burden in nasopharyngeal carcinoma, with advanced N stage contributing meaningfully to poorer outcomes, while tumor grade remained an important marker of biologic aggressiveness. Notably, tumor stage did not emerge as a significant predictor in this cohort, likely reflecting the relatively narrow range of locally advanced disease included in the study.

From a clinical perspective, the model offers a potential framework for risk-adapted care. Patients identified as high risk may benefit from closer surveillance, more proactive management of treatment-related toxicities, and more individualized supportive care strategies. At the same time, the use of explainable AI provides transparency that may facilitate clinician trust and integration into routine practice.

According to the study authors, this work highlights the potential of interpretable AI to support precision oncology by combining robust prediction with clinically meaningful explanation for greater adoption in routine oncology workflows.

“This study demonstrates that an interpretable XGBoost-based AI model can provide clinically useful survival prediction and risk stratification for patients with locally advanced NPC undergoing immuno-chemoradiotherapy,” they said, adding that the addition of radiomics, genomic/transcriptomic, or immune biomarkers could further strengthen the model's utility for precision medicine.

DISCLOSURES: This study was supported by 2024 Annual Scientific Research Project Approval of the Health System in Zigong. For full disclosures of the study authors, visit frontiersin.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.