Modeling Patient-Reported Swallowing Outcomes With Dose to the Epiglottis

Abstract

Swallowing is a complex physiologic function and care is needed for its preservation when head and neck (HN) cancer patients receive radiation therapy. The larynx, and especially the epiglottis, is one involved organ. This study's purpose was to investigate whether radiation dose to the epiglottis is more predictive of acute swallowing toxicity than dose to the rest of the larynx or the total larynx.A cohort of 91 HN patients with no prior HN radiation treated at a single institution form 2013-2020 with either protons or photons was studied retrospectively. Patients were selected based on primary-site diagnosis code (nasopharynx, oropharynx, tonsil, and base of tongue) and prescription dose (≥60 Gy). Swallowing outcomes were assessed with patient-reported outcomes (PROs), using the EORTC QLQ H&N35 survey, that were collected prior to (baseline) and following treatment (within 2 weeks of finishing). The post-treatment swallowing-specific PRO scores were used to evaluate acute swallowing toxicity, and patients were stratified into those with high (score > 50) and low (score≤50) responses. The epiglottis, the arytenoid cartilages, and the non-epiglottic larynx were contoured by a single individual (SA); the three substructures were then combined for a total larynx structure. 16 patients with either a larynx disfigured by their disease or high baseline swallowing score were removed from the analysis. A univariate analysis was first performed using a Wilcoxon rank-sum test on the mean doses to the epiglottis, total larynx, and non-epiglottic larynx structures. Next, a random-forest (RF) machine-learning model was trained using 29 patient-specific features with demographic and clinical information (including primary site, chemotherapy, smoking, bi-/unilateral, and HPV/p16), baseline PRO scores, and dose metrics to the total larynx and epiglottis. The model was trained on a 60-patient training set using nested 10-fold cross-validation and then assessed on a 15-patient test set using the area under the receiver-operator curve (AUC).The univariate analyses for the total and non-epiglottic larynx mean doses showed no statistical differences between the two groups (P = 0.56 and 0.86, respectively). That of the epiglottic mean dose, however, showed a significant difference (P = 0.026) with 39.2 [32.7-52.9] Gy and 47.2 [41.1-57.2] Gy for the low- and high-response groups, respectively. The RF had AUC = 0.71 ± 0.22 and 0.70 on the training and test sets, respectively, with the epiglottis volume receiving 30 Gy identified as the most important feature.The univariate analysis and RF model indicate that the dose to the epiglottis is more sensitive for predicting patient-reported acute swallowing toxicity than either the doses to the rest of the larynx or the full larynx. These results suggest that consideration be given to the epiglottic dose during treatment planning.