A Scalable AI Framework for Automated TNM Staging of 51,242 Patients with Nasopharyngeal Carcinoma

Accurate tumor-node-metastasis (TNM) staging is essential for treatment decision-making, prognostic stratification, and outcomes modeling in oncology. However, real-world clinical databases often extend over decades and inherently encompass multiple successive staging systems, introducing substantial heterogeneity that undermines the stability and interpretability of survival analyses, prognostic modeling, and staging system evaluation. To address this challenge, we developed an automated framework that combines TNM staging criteria and staging-relevant anatomical knowledge with a large language model (LLM) and a second-stage reflection step to convert narrative radiology reports into unified AJCC/UICC 9th edition TNM staging. In an expert-annotated cohort of 340 NPC patients, we evaluated multiple LLMs and found that the GLM4.5-based framework achieved the highest accuracy for T (0.95), N (0.86), and M (0.99) classification. Based on this optimal performance, the framework was applied to a real-world cohort of 51,242 NPC patients between 2010 and 2025. A stratified random validation subset of 200 cases showed high concordance with the expert reference standard. In downstream utility analyses, the unified staging was associated with a higher C-index and more distinct survival-curve separation than existing electronic health record staging. Collectively, this study provides a scalable framework for unifying heterogeneous historical cancer staging data, supporting longitudinal analyses in large real-world oncology cohorts.