Prediction of Dysphagia after Head and Neck Re-Irradiation Using Interpretable Machine Learning Models
Abstract
Purpose
Re-irradiation (re-RT) is an important treatment option for recurrent head and neck cancer (HNC); however, dysphagia remains a clinically significant toxicity. Accurate prediction models are needed to identify high-risk patients and support patient treatment management. This study develops and evaluates interpretable machine learning models to predict dysphagia in patients receiving second course of RT.
Methods
A retrospective cohort of 46 patients with recurrent HNC treated with re-irradiation was analyzed. Dysphagia following the second RT was dichotomized as grade ≥ 2 versus grade < 2 according to CTCAE criteria. Clinical features included tumor stage at re-RT, re-RT prescription, prior prescription, interval between the first and second courses, and the presence of dysphagia following the first radiation course. Dosimetric features were derived from cumulative dose distributions, with doses converted to equivalent dose in 2-Gy fractions (EQD2) to account for differences in fractionation. A logistic regression model (LR) was trained as the primary predictive model, with explainable boosting machine and decision tree models were trained for comparison. Model performance was evaluated using five-fold cross-validation and area under the receiver operating characteristic curve (AUC).
Results
LR achieved the best predictive performance among the evaluated models, with a cross-validated AUC of 0.755. Parotid gland volume was the strongest predictor of dysphagia after re-irradiation, with cumulative dose, laryngeal dose and volume, fractionation, parotid dose metrics, and treatment interval also contributing significantly. More complex models did not improve performance over LR in this cohort (AUC<0.700).
Conclusion
The results suggest that dysphagia following HNC re-RT is primarily driven by cumulative dose burden and injury to salivary and laryngeal structures, with parotid gland volume and laryngeal dose emerging as key predictive factors. The proposed interpretable LR model demonstrates potential for clinically meaningful risk stratification and underscores the value of transparent modeling approaches for toxicity prediction in re-RT settings.