# Interpretable AI for treatment decision-making in immunoradiotherapy of locally advanced nasopharyngeal carcinoma

## TL;DR

This study develops an interpretable AI model to predict survival outcomes in locally advanced nasopharyngeal carcinoma patients, helping guide treatment decisions.

## Contribution

The novel contribution is an explainable AI model using clinical variables for risk stratification in immunoradiotherapy of nasopharyngeal carcinoma.

## Key findings

- The XGBoost model effectively stratified patients into low- and high-risk groups with significantly different survival outcomes.
- Age, tumor grade, and N stage were identified as the strongest predictors of mortality risk using SHAP analysis.
- The model achieved AUCs of 0.784, 0.765, and 0.725 for predicting 1-, 2-, and 3-year overall survival in the validation cohort.

## Abstract

Survival remains heterogeneous in locally advancednasopharyngeal carcinoma (NPC) despite immunotherapy, highlighting the need for explainable artificial intelligence (AI) for risk-adapted care.

We retrospectively analyzed 249 patients with locally advanced NPC between 2018 and 2025. Patients were randomly split into a training cohort (70%) and a validation cohort (30%). A Cox–XGBoost survival modeling framework was developed using routinely available clinical variables to generate individualized risk scores and classify patients into low- and high-risk groups. Model discrimination was assessed using time-dependent ROC analysis. SHAP (SHapley Additive exPlanations) was applied to provide transparent, feature-level and patient-level interpretations of predicted risk.

Univariable Cox regression identified age, tumor grade, and N stage as significant prognostic factors. In the training cohort, the XGBoost-derived risk score robustly separated low- and high-risk groups, with significantly prolonged survival in the low-risk group (P < 0.001). In the validation cohort, the AUCs for predicting 1-, 2-, and 3-year OS were 0.784, 0.765, and 0.725, respectively. SHAP analyses consistently highlighted age as the strongest driver of predicted risk, followed by N stage and tumor grade; older age and advanced nodal disease were associated with higher predicted mortality risk.

An interpretable XGBoost-based survival model built from routine clinical variables provides clinically meaningful risk stratification for locally advanced NPC patients.