Predicting Late Radiation-Induced Xerostomia with Parotid PET Biomarkers and Dose Metrics

Abstract

Xerostomia is a common long-term toxicity from head and neck radiotherapy that is only partially predicted by delivered dose to salivary glands. Associations between parotid gland imaging biomarkers and late radiation-induced xerostomia have recently been shown. The aim of this study was to assess predictive models for late xerostomia incorporating pre-treatment parotid PET biomarkers compared to a base model using parotid dose alone in a homogeneous group of patients treated for orophangeal cancer. 47 patients with p16+ oropharyngeal squamous cell carcinoma were treated on two prospective institutional protocols with 70 Gy IMRT and concurrent carboplatin/paclitaxel. The 90th percentile of standardized uptake value (SUV-P90) was generated from each parotid gland on pre and 3-month post-RT PET scans. Late xerostomia (12-months post-RT) was defined by the patient-reported University of Michigan Xerostomia Questionnaire (XQ), and parotid-specific salivary flow rate (SFR). XQ outcomes included both the summary score (XQ-sum12m) and the most variable eating-related item as determined by principal component analysis: frequency of sipping fluids to aid in swallowing food (XQ-sip12m). Univariate analysis was applied to dose and SUV-P90 features as predictors of late xerostomia using both continuous and discrete scales. Multivariate analysis was applied for dose only and combined dose and SUV-P90 using penalized logistic regression for feature selection and model prediction based on the LASSO technique. Mean parotid dose and SUV-P90 from both pre and post-RT PET scans were significantly associated with patient-reported xerostomia and stimulated SFR (Stim12m) but not unstimulated SFR. Associations were stronger for SUV-P90 from the contralateral parotid (SUV-P90CL) than the ipsilateral parotid, and from 3-month PET scans relative to pre-treatment scans. On multivariate analysis, the addition of pre-treatment SUV-P90CL and an interaction term for dose and SUV to the model significantly improved the prediction for late moderate or severe xerostomia compared to the base model, from AUC = 0.70 to 0.74 for XQ-sum12m and from 0.74 to 0.84 for XQ-sip12m. The improvements remained statistically significant upon cross-validation. The addition of pre-treatment parotid PET biomarkers significantly improved a predictive model for late patient-reported xerostomia over a model based on parotid dose alone. The improvement was greatest for the single eating-related item with the largest inter-patient variability (XQ-sip12m). Correlation between SUV-P90 from 3-month post-RT PET scans and late xerostomia was even higher than those from pre-treatment scans, suggesting that mid-treatment scans should be further explored for a potential role in adaptive radiotherapy.Abstract 121; Table 1SUV-P90CL (3m)Mean XQ-Sum12mMean XQ-Sip12mMean Stim12m< 1.551.37.50.171.5 - 2.029.04.40.392.0 - 2.521.92.80.48> 2.510.81.30.80 Open table in a new tab