Risk stratification of long-term distant metastasis in nasopharyngeal carcinoma: a retrospective study of 7,045 patients diagnosed from 2012 to 2022

Abstract

Distant metastasis (DM) remains the leading cause of treatment failure in nasopharyngeal carcinoma (NPC), severely compromising long-term survival. Despite advances in screening and multimodal therapies, a considerable proportion of patients still develop DM after standard treatment. Therefore, early identification of individuals at high risk of metastasis is essential for optimizing surveillance strategies and guiding personalized therapeutic interventions. However, existing prognostic models are often limited by small sample sizes, incomplete clinical variables, and insufficient interpretability, which restrict their clinical applicability. Our objective was to construct and validate an explainable machine learning model for early DM risk prediction in NPC by integrating routinely available clinical and hematological indicators. We retrospectively analyzed clinical data of 7,045 NPC patients between 2012 and 2022. Demographic characteristics, TNM staging, pathological subtypes, and hematological parameters were collected. Missing data were handled using multiple imputation, and continuous variables were standardized. Feature selection was performed using variance inflation factor and univariate Cox regression with concordance index ranking. Six survival models were developed to predict distant metastasis-free survival (DMFS), including the Component-wise Gradient Boosting (CGB), Cox proportional hazards (CoxPH), Least Absolute Shrinkage and Selection Operator (LASSO)-Cox, Extra Survival Trees (EST), Gradient Boosting Survival Analysis (GBSA), and Random Survival Forest (RSF). Model performance was assessed through Harrell’s concordance statistic, Time-dependent receiver operating characteristic curve, and Time-dependent Brier score across 1–5 years. Calibration curves and decision curve analysis were further applied. Model interpretability was enhanced using Shapley additive explanations (SHAP), and a nomogram was developed for individualized prediction. Among all algorithms, the CoxPH model demonstrated the best overall performance and generalizability. In the test cohort, it achieved a C-index of 0.810 (95% CI: 0.774–0.843), and an average Time-AUC of 0.824 (95% CI: 0.782–0.858), with favorable calibration and low Brier scores. Risk-stratification analysis revealed significant separation between high- and low-risk groups in both cohorts (all P < 0.0001). SHAP analysis highlighted N stage, Post-RT EBV-DNA, and T stage as the strongest contributors to DM risk. A visualized nomogram enabled intuitive estimation of 1-5-year DMFS probabilities, facilitating clinical implementation. Leveraging a large single-center cohort, we developed and validated an interpretable ML framework capable of predicting DM risk in NPC. The model supports timely identification of high-risk patients and provides a practical tool for personalized decision-making. Further multicenter prospective studies are warranted to validate its generalizability and clinical utility.