Machine learning prognostication in nasopharyngeal carcinoma: a european multicentre analysis of survival and risk of second malignancy.

INTRODUCTION: Nasopharyngeal carcinoma (NPC) is rare in Europe, and emerging data suggest poorer outcomes in Caucasian patients compared with Asian populations, highlighting the need for region-specific prognostic tools. Inflammation-based biomarkers and artificial intelligence show promise for risk stratification and prediction of survival and second primary cancers (SPCs). MATERIALS AND METHODS: We conducted a retrospective multicentre study including 405 NPC patients from six European institutions. Demographic, clinicopathological, and haematologic inflammatory markers were collected, and machine learning algorithms were developed to predict 5-year OS and SPC occurrence. Multiple train-test splitting strategies and machine learning (ML) classifiers were evaluated. Models were tested both with and without systemic inflammatory ratios to assess their added prognostic value. RESULTS: The median age was 52 years, 91.6% of patients were classified as White/European ancestry, and 77.3% received chemoradiotherapy. Five-year OS was 66.6%, while 12.8% developed SPC. The Random Forest classifier achieved the best performance for OS prediction (accuracy 0.74; AUC 0.66) using the complete feature set, while SPC prediction reached an accuracy of 0.80 (AUC 0.74). Exclusion of inflammatory markers resulted in a consistent decline in accuracy across all models. Feature-importance analysis highlighted inflammatory ratios among the strongest predictors for both OS and SPC. The present study was reported according to TRIPOD+AI reporting guidelines. CONCLUSIONS: This study presents the first machine-learning prognostic models for nasopharyngeal carcinoma derived from a predominantly Caucasian European multicentre cohort. Systemic inflammatory markers modestly improved overall survival prediction and substantially enhanced second primary cancer risk estimation. The resulting models are transparent, cost-effective, and support the potential benefit of prognostic assessment through machine learning in non-endemic settings.