Aims: Outcome prediction models may offer a precision approach for clinical decision-making. Prior to clinical use, independent external validation is important to demonstrate accuracy and generalisability across populations. We validated a Danish Head & Neck cancer outcome prediction model in an Australian population.
Methods: Retrospective data from 502 patients treated at six centres were used to validate the multi-endpoint model, predicting 3year locoregional failure (LRF), distant metastases (DM) and death without evidence of disease (DeathNED). Model performance was assessed using Area-Under-Curve (AUC), calibration plots and sub-distribution hazard ratios for competing risks. Model validation was also undertaken in patients ≥70yrs to examine model fairness in this subgroup.
Results: For the full cohort, 3yr AUC values were 0.66 (LRF), 0.73 (DM) and 0.74 (DeathNED), with similar values for the older subgroup. Sub-distribution hazard ratios demonstrated the model could distinguish low (ref), intermediate, and high-risk groups for LRF (p<0.001) (HR 1.92; 95% CI:1.19-3.09; 4.69; 95% CI:2.79-7.88, respectively) and DM (p<0.001) (HR 2.34; 95% CI:1.07-5.14; 7.19; 95% CI:3.45-15.0). For DeathNED, significant separation was observed between low and high-risk groups (HR 2.40, 95% CI: 1.15-5.02), but not between low and intermediate-risk groups (HR 0.66; 95% CI:0.40-1.08). Calibration plots exhibited good calibration for LRF and DM, with slight miscalibration observed in the high-risk group for LRF (over-prediction at risks >30%). For DeathNED, miscalibration was observed, with under-prediction at risks >25%. For the older subgroup, over-prediction at higher risk for LRF was observed, but the model was well-calibrated for DM and DeathNED.
Conclusions: The model demonstrated consistent performance in an Australian cohort relative to the Danish development cohort, highlighting its generalisability. It may support personalised discussions around treatment decision-making, particularly for older patients who may require treatment modifications due to frailty, comorbidities or personal preferences. It represents a step toward clinically deployable decision-support tools in radiation oncology.