Development of a clinical prediction model for overall survival in locally advanced head and neck squamous cell carcinoma

Abstract To develop and validate a machine-learning–based prediction model for overall survival (OS) in locally advanced head and neck squamous cell carcinoma (LA-HNSCC) that integrates both surgery-based and chemoradiotherapy (CRT) cohorts within a single framework. A total of 1,005 LA-HNSCC patients receiving curative-intent surgery-based treatment or CRT were retrospectively assembled from a domestic multicenter cohort ( n = 678) and publicly available datasets from The Cancer Imaging Archive ( n = 327). Candidate predictors were pre-specified from routinely available clinical variables. Elastic net regularization with 10-fold cross-validation was applied for variable selection and weighting, followed by a multivariable Cox proportional hazards model. Elastic net selected 12 variables, of which p16 status, performance status, nodal stage, serum albumin, T stage, and age were most influential. The model showed good discrimination (C-index 0.73), with time-dependent AUCs of 0.78–0.73 across 1–5 years and good calibration at 1 and 3 years. Decision curve analysis confirmed clinical utility across relevant threshold probabilities in both cohorts. By relying on prognostic factors common to both surgery-based treatment and CRT, this interpretable model demonstrated stable performance and may support prognostic risk stratification applicable across treatment modalities in LA-HNSCC.