Early toxicity prediction using longitudinal cone-beam CT delta-radiomics in HPV-positive oropharyngeal cancer.

Introduction: Long-term survival in human papillomavirus-positive (HPV+) oropharyngeal squamous cell carcinoma (OPSCC) has shifted clinical focus toward minimizing treatment-related toxicities, particularly dysphagia, which significantly impacts quality of life. This study investigated whether longitudinal cone-beam computed tomography (CBCT)-based delta-radiomics of the supraglottic region can provide early indication of 6-12 month post-radiotherapy dysphagia. Methods: Forty-three patients with HPV+ OPSCC treated with definitive radiotherapy or chemoradiotherapy underwent daily CBCT imaging. Forty-three radiomic features were extracted from the supraglottic region of interest, volume-adjusted where appropriate, averaged weekly, and transformed using a ratio-to-baseline formulation anchored to the first treatment-day CBCT. Spearman correlation filtering (|ρ| > 0.8) reduced redundancy prior to Random Forest ranking using out-of-bag permutation importance. Logistic regression models were constructed using incrementally increasing feature subsets across individual weeks and grouped multi-week CBCT temporal windows. Model performance was evaluated using repeated random two-thirds resampling with 1000 iterations, and discrimination was quantified by the mean area under the receiver operating characteristic curve with 95% percentile confidence intervals. Results: The study cohort included 43 patients with HPV+ OPSCC treated with definitive radiotherapy. Predictive signal emerged early during treatment. The Weeks 1-4 model achieved a mean area under the curve of 0.779 (95% confidence interval: 0.714-0.812) using three features, outperforming the Week 2 model (0.728) and demonstrating comparable discrimination to the Weeks 1-6 model (0.795) with greater parsimony. Global skewness was consistently the highest-ranked predictor across temporal windows. Conclusion: Supraglottic CBCT-based delta-radiomics demonstrates an early signal associated with post-radiotherapy dysphagia in HPV+ OPSCC. A ratio-to-baseline longitudinal framework combined with weekly aggregation enabled robust modeling while preserving interpretability. Given the limited cohort size and absence of external validation, these findings should be interpreted as exploratory and hypothesis-generating.