Click here to

Session: Emerging Imaging, Therapy, and Dosimetry Solutions I [Return to Session]

Machine-Learning Modeling of Patient-Reported Swallowing Dysfunction Using Dose to the Epiglottis and Pharyngeal Constrictors

MD Pepin*, S Anaya, RL Foote, YI Garces, H Emily, DW Mundy, MA Neben-wittich, DM Routman, S Catherine, SC Lester, DJ Ma, S Shiraishi, Mayo Clinic Rochester, Rochester, MN


SU-J-202-6 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Room 202

Purpose: The purpose of this study was to construct a machine-learning model to predict acute patient-reported swallowing dysfunction following head and neck (HN) radiotherapy.

Methods: Modeling was performed on a retrospective cohort of 136 HN patients treated at a single institution from 2015–2020. Swallowing toxicity was assessed using patient-reported outcomes gathered with the EORTC QLQ-H&N35 questionnaire before and after treatment; patients were stratified into high (N=72) and low (N=64) post-treatment scores. The data was split into training (N=108) and test (N=28) sets stratified by post-treatment score and gender. Modeling features included clinical variables and dosimetric metrics for the epiglottis and pharyngeal constrictors; the epiglottis was specifically contoured and considered due to its physiological involvement with swallowing. Regularized logistic regression (LR), support vector machine (SVM), and soft-voting classifier (VC) models were trained using iterative stratified 5-fold cross validation (CV) and assessed using the area under the receiver-operator curve (AUC). An in-silico nomogram was created by scaling the DVHs of the epiglottis and pharyngeal constrictors and then re-predicting the swallowing toxicity probability with the SVM model.

Results: The training, test, and CV AUCs were 0.79, 0.69, and 0.70±0.10 for the LR, 0.80, 0.67, and 0.70±0.10 for the SVM, and 0.79, 0.71, and 0.71±0.10 for the VC. The nomograms showed patient-specific behavior with different behavior for clinically similar patients. The relative importance of the epiglottis and pharyngeal constrictors was demonstrated in the nomogram; the predicted probability changed more with a change in the constrictor vs. the epiglottis DVH.

Conclusion: Multiple models were trained and evaluated showing comparable performance and agreement between the testing dataset and CV AUC values. The nomogram developed demonstrates how the model can be used during treatment planning to assess multiple plans and assist the clinician in making decisions in the era of personalized medicine.


Radiation Therapy, Radiation Effects, Modeling


TH- Response Assessment: Modeling: Machine Learning

Contact Email