br Endpoints br The following toxicity endpoints were define
The following toxicity endpoints were defined at 6 months from treatment completion: (1) dysphagia grade 2; (2) dysphagia grade 3; (3) xerostomia grade 2; (4) salivary duct AR-13324 grade 2; and (5) tube feeding depen-dence. Salivary duct inflammation toxicity was graded as such: grade 1 Z slightly thickened saliva and slightly altered taste; grade 2 Z thick, ropy, sticky saliva, markedly altered taste, alteration in diet indicated, secretion-induced symptoms, and limiting instrumental activities of daily living; grade 3 Z acute salivary gland necrosis, severe secretion-induced symptoms, tube feeding indicated, limiting self-care activities of daily living, and disabling;
grade 4 Z life-threatening consequences with urgent intervention indicated; and grade 5 Z death.
The 6-month endpoint was chosen, given that this was the time point for which the largest amount of toxicity data existed for all patients. All toxicity endpoints were collected and documented prospectively.
Patients who had one of the endpoints already at baseline were excluded from the analyses regarding that particular endpoint. For each endpoint, multivariable NTCP models were created. Univariable logistic regression analyses and correlation statistics were performed to select candidate predictors for each endpoint that were significantly asso-ciated with the endpoints in univariable analysis (P < .05), but not mutually correlated (r < 0.80). Next, a stepwise backward multivariable logistic regression procedure was used to exclude the variables with P >.157 from the model. The resulting model was then manually explored further in the following 2 ways: (1) by testing whether the models would significantly deteriorate when one or more variables would be removed and (2) by exchanging the selected dose volume variables by other potentially relevant dose vari-ables that were highly correlated to the selected dose var-iable and therefore discarded in an earlier stage. The final best model was chosen primarily by the applying the likelihood-ratio test and by evaluating the general model performance measuresdthat is, received operative curve area under the curve, discrimination slope, explained vari-ance, and calibration. For each endpoint, the final model was subjected to internal validation with a bootstrapping procedure to correct (shrink) the models (slope and inter-cept) for optimism. This was done to obtain realistic regression coefficients for the model variables that are representative for populations like the development sample.
Potential model variables: Univariable logistic Correlation matrix:
keep variables with
with lowest P-value
One-by-one removal of
variable with highets P-value Conditional stepwise
Establish final model Keep variable in the case of backward variable selection
for multivariable model:
(likelihood ratio test)
Fig. 1. Variable selection and logistic regression modeling for each endpoint.
556 Rwigema et al. International Journal of Radiation Oncology Biology Physics
A figure summarizing the steps in model generation is shown in Figure 1.
Candidate variables that were initially entered in the model were sex (male versus female), age (as contin-uous variable), concomitant chemotherapy (no vs yes), weight loss at baseline (0-10 vs >10%), accelerated radiation therapy (no vs yes), T stage (stage 1-2 vs 3-4), N stage (negative vs positive), target volume (local or unilateral vs bilateral neck irradiation), surgery (no vs yes), and baseline toxicities (grade 0 vs grade 1). Paired t tests and Wilcoxon rank tests were used to compare mean NTCP results for endpoints between PBT and IMRT. Data were analyzed using SPSS Statistics for Windows, version 23.