Correlation and Prediction of Modified Barium Swallowing Study with Patient-Reported Outcomes (PROs) for HPV Associated Oropharyngeal Squamous Cell Carcinoma

Abstract

<h3>Purpose/Objective(s)</h3> We combine data from three multi-site Phase II/III radiation therapy (RT) clinical trials aimed at treating HPV(+) oropharyngeal squamous cell carcinoma (OPSCC). Trials consisted of patients treated with standard of care (SOC) adjuvant (chemo) IMRT, de-escalated adjuvant RT and chemotherapy (DART), and mucosal sparing proton beam therapy (MSPBT). We provide models utilizing the EORTC-HN35 to predict scores for the modified barium swallowing impairment profiles (MBSImP). Validated, accurate models could provide significant clinical information while reducing patient personal, structural, financial, and emotional burden of treatment (BoT) by utilizing PROs in lieu of MBSImP. <h3>Materials/Methods</h3> In total, 496 paired HN35/ MBSImP surveys were collected at pre-RT and 12 months post-RT for 249 trial patients. Analyses included all paired data points. Correlation, backward-stepwise regression (OLS), regression tree, and random forest models were fit to assess the predictive power of utilizing HN35 scales for MBSImP scales of penetration/aspiration, oral, pharyngeal, and overall impairment profiles. Model validation and calibration was conducted utilizing bootstrap replicates. <h3>Results</h3> Of the 18 HN35 scales, 10 were found through correlation coefficients and regression models, to be associated with MBSImP: swallowing (SW), social eating (SO), social contact (SC), pain (PA), speech (SP), senses (SS), sexuality (SX), dry mouth (DR), coughing (CO), and opening mouth (OM). OLS for MBSImP scales varied in variable selection and prediction accuracy (Table 1), however, all models demonstrated a high level of accuracy (mean absolute error) and calibration (slope). Underprediction of none/mild impairment (<2) was most common, however, predictions of high impairment were most accurate. Regression trees were able to accurately predict high impairment (top 75^th percentile) for each MBSImP scale utilizing 7 or less HN35 scales for 84% of patients. In many cases, patients which would typically score low on the MBSImP scales can be correctly classified with four or fewer HN35 scales: SW, SX, SS, PA. <h3>Conclusion</h3> The ability to use patient reported measures (HN35) to correlate and accurately predict the outcomes from a functional imaging (MBSImP) could have significant impact on burden of treatment. Verification studies are planned to confirm the efficacy of these models prospectively.