Does delivered OAR dose improve prediction of late toxicity in head & neck cancer patients?

Abstract

Purpose or Objective Delivered radiation dose to H&N OARs (DA) may be different from planned dose (DP), and adapting treatment plans mid-treatment can correct for this. However, few studies directly assess whether ART improves clinical outcomes, and None link delivered OAR dose to side effects. The purpose of this study was to establish whether planned or delivered OAR dose better predicts toxicity events, thereby providing clinical data to support ART in HNC. Material and Methods 198 HNC patients were treated with standard protocols on TomoTherapy units with daily MVCT-IG. Toxicity data were prospectively collected at baseline and 12 months. OARs manually segmented on planning CT scans were: bilateral parotid and submandibular glands, superior and middle pharyngeal constrictor muscles, supraglottic larynx and oral cavity. Contours were propagated to daily MVCTs with a specifically trained and validated open source DIR algorithm (Elastix). Daily dose was calculated with an inhouse ray-tracing algorithm (CheckTomo), and accumulated to estimate delivered DVH. Planned dose was recalculated with CheckTomo for direct comparison. Toxicity data for the following endpoints were binarised for analysis: CTCAEv4.03 – xerostomia, salivary duct inflammation and dysphagia, EORTC H&N35 - dry mouth (Q41), sticky saliva (Q42), taste disturbance (Q44), LENT/SOM – xerostomia subjective, xerostomia management. Forward stepwise variable selection by likelihood ratio fed statistically significant baseline and mean OAR DP variables (from univariate analysis) into logistic regression based multivariate NTCP models, which were internally validated/calibrated with 1000-fold bootstrapping and10-fold cross-validation. The process was repeated with mean OAR DA metrics, and DP and DAbased model performance was compared with Nagelkerke R2, -2 log-likelihood, discrimination slope (DS), and ROC curve AUC statistics. Conclusion All models were statistically significant predictors of toxicity. Interestingly, respective DA models had superior Nagelkerke R2 statistics for 6/8 endpoints, 6/8 ROC curve AUCs, and 8/8 discrimination slopes, but differences were minimal. This may be of interest for future NTCP model development, but supports the notion that dose deltas are clinically insignificant in most patients, and careful case selection is required for optimised ART.