Abstract
BACKGROUND: This study was to evaluate the predictors of xerostomia and Grade 3 xerostomia in locoregionally advanced nasopharyngeal carcinoma (NPC) patients receiving radical radiotherapy and establish prediction models for xerostomia and Grade 3 xerostomia based on the predictors. METHODS: Totally, 365 patients with locoregionally advanced NPC who underwent radical radiotherapy were randomly divided into the training set (n = 255) and the testing set (n = 110) at a ratio of 7:3. All variables were included in the least absolute shrinkage and selection operator regression to screen out the potential predictors for xerostomia as well as the Grade 3 xerostomia in locoregionally advanced NPC patients receiving radical radiotherapy. The random forest (RF), a decision tree classifier (DTC), and extreme-gradient boosting (XGB) models were constructed. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under the curve (AUC) and accuracy were analyzed to evaluate the predictive performance of the models. RESULTS: In the RF model for predicting xerostomia, the sensitivity was 1.000 (95%CI 1.000-1.000), the PPV was 0.990 (95%CI 0.975-1.000), the NPV was 1.000 (95%CI 1.000-1.000), the AUC was 0.999 (95%CI 0.997-1.000) and the accuracy was 0.992 (95%CI 0.981-1.000) in the training set. The sensitivity was 0.933 (95%CI 0.880-0.985), the PPV was 0.933 (95%CI 0.880-0.985), and the AUC was 0.915 (95%CI 0.860-0.970) in the testing set. Hypertension, age, total radiotherapy dose, dose at 50% of the left parotid volume, mean dose to right parotid gland, mean dose to oral cavity, and course of induction chemotherapy were important variables associated with the risk of xerostomia in locoregionally advanced NPC patients receiving radical radiotherapy. The AUC of DTC model for predicting xerostomia was 0.769 (95%CI 0.666-0.872) in the testing set. The AUC of the XGB model for predicting xerostomia was 0.834 (0.753-0.916) in the testing set. The RF model showed the good predictive ability with the AUC of 0.986 (95%CI 0.972-1.000) in the training set, and 0.766 (95%CI 0.626-0.905) in the testing set for identifying patients who at high risk of Grade 3 xerostomia in those with high risk of xerostomia. CONCLUSIONS: An RF model for predicting xerostomia in locoregionally advanced NPC patients receiving radical radiotherapy and an RF model for predicting Grade 3 xerostomia in those with high risk of xerostomia showed good predictive ability.