External validation of a head & neck cancer outcome prediction model in radiotherapy patients: Evaluating generalisability in a geographically diverse population, with particular consideration of older patients.

Abstract

Outcome prediction models for Head and Neck Squamous Cell Carcinoma (HNSCC) may be useful for clinical decision support. Prior to clinical use, multiple independent external validations are important to demonstrate generalisability in different populations. This study aimed to externally validate a Danish HNSCC outcome prediction model in an Australian patient population, treated using different treatment guidelines and evaluate validation in older (70 + years) patients. Retrospective data from patients treated at six Australian centres were used to validate a multi-endpoint model predicting 3-year locoregional failure (LRF), distant metastases (DM) and death without evidence of disease (death NED). Model performance was assessed using Area Under Curve (AUC), Brier score, calibration plots and subdistribution hazard ratios for competing risks. Model validation was also undertaken in older patients (≥70 years) to examine model fairness in this subgroup. For the full cohort (n = 502), 3-year AUC values were 0.66 (95 % CI 0.60-0.73) for LRF, 0.73 (0.65-0.80) for DM and 0.73 (0.66-0.79) for Death NED. Similar values were obtained for the older subgroup. Subdistribution hazard ratios demonstrated the model could distinguish low, intermediate, and high-risk groups for LRF and DM. Calibration plots exhibited slight miscalibration in the high-risk groups for LRF and Death NED. The model demonstrated consistent performance in an Australian population compared to the Danish development cohort, and in older patients, highlighting its generalisability. It may serve as a practical tool to support personalised discussions around prognosis and treatment, particularly for older patients requiring treatment modifications.