1251 - Interpretable Modeling of Dysphagia Risk after Head and Neck Re-irradiation

Purpose/Objective(s): Dysphagia is a major dose-limiting toxicity following head and neck cancer (HNC) re-irradiation (re-RT). Prior clinical studies have associated injury of salivary glands and laryngeal structures with swallowing dysfunction; however, individualized risk prediction remains limited. In this study, we hypothesized that interpretable machine learning models incorporating cumulative dose metrics and clinical factors can accurately predict dysphagia after re-RT while identifying predictors consistent with established clinical experience.

Materials/Methods: A retrospective cohort of 64 patients with recurrent HNC treated with re-RT at a single institution was analyzed. All patients received IMRT for their initial course. For re-RT, 43 patients received SBRT and the remainder received IMRT. Dysphagia after re-RT was dichotomized as grade=2 versus grade<2 per according to CTCAE criteria. Clinical data included tumor stage, initial and re-RT prescription doses, interval between radiation courses, and baseline dysphagia status. Dosimetric features were derived from cumulative dose distributions and converted to EQD2. Feature similarity analysis was performed for pre-selection to reduce redundancy. Logistic regression (LR) was the primary interpretable model, with explainable boosting machine (EBM) and decision tree (DT) models evaluated for comparison. Performance was evaluated using five-fold cross-validation and area under the receiver operating characteristic curve (AUC).

Results: LR achieved the highest performance (AUC=0.78). Parotid gland volume and cumulative parotid EQD2 were dominant predictors, consistent with prior clinical evidence linking salivary gland injury xerostomia-associated dysphagia and swallowing impairment. Laryngeal dose and volume were also significant predictors, aligning with established reports identifying the larynx as a critical swallowing organ. Shorter treatment intervals and higher cumulative dose were associated with increased dysphagia risk. Comparison between LR to EBM and DT models showed that more complex models did not improve performance (AUC<0.70) likely reflecting the limited sample size.

Conclusion: Interpretable modeling achieved clinically meaningful predictive performance for dysphagia after HNC re-irradiation and identified key predictors that align with prior clinical evidence. These results support the clinical relevance of cumulative dose to salivary and laryngeal structures and highlight the potential of transparent models to guide patient selection, optimize organ-at-risk dose sparing, and demonstrate that clinically interpretable AI models can support prior knowledge–guided HNC re-irradiation planning.