ESTRO 2024 - Abstract Book

S1275

Clinical - Head & neck

ESTRO 2024

For the swallowing domain, model performance was good for both dysphagia (AUC refit =0.65 CI95%[0.59-0.72]) and aspiration (AUC refit =0.72 [0.63-0.82]) in the US cohort. Refitting the regression coefficients showed a larger impact of dose to the middle pharyngeal constrictor muscle and baseline toxicity score in the swallowing domain for patients in the US cohort. For the salivary domain, model performance was good for xerostomia (AUC refit =0.62 [0.55-0.68]) and sticky saliva (AUC refit =0.63 [0.56-0.71]) in the US cohort. Nonetheless, model performance for taste was low (AUC refit =0.54 [0.48 0.61]. For xerostomia and sticky saliva, refitted coefficients demonstrated similar relations between predictors as in the developed models. However, in the refitted taste model the contribution of coefficients for oral cavity and parotid gland dose was smaller than in the original model. For the general domain, model performance of fatigue was good (AUC refit =0.74 [0.68-0.80]) in the US cohort and without a refit it outperformed the Dutch validation cohort (AUC validation =0.73 vs. 0.70). Refitted coefficients demonstrated the same relation between the predictors as in the original model.

Conclusion:

Despite many differences between the development cohort and the validation cohort, all models exhibited good performance except for the taste model. This demonstrates that taste might be more complex to predict than the relationship captured in this model. Even after a refit of the models, predictors within most models displayed the