Prediction of Late Xerostomia after Radiotherapy by Integrating Dosiomics and Conventional Predictive Factors in Patients With Oropharyngeal Cancer

Abstract

Xerostomia (dry mouth) is one of the main toxicities after radiotherapy in patients with oropharyngeal cancer. Although the mean dose to the parotid glands is well known to a predictive factor for xerostomia, it cannot show where a dose is received within the structures due to no spatial information of dose-volume-histogram (DVH). The purposes of this study were to develop the dosiomics-based predictive model for xerostomia after radiotherapy and to clarify the efficacy of the dosiomics approach for patients with oropharyngeal cancer.A total of 133 patients with oropharyngeal cancer treated with intensity modulated radiation therapy between 2010 and 2018 were included. Xerostomia was scored for each patient by radiation oncologists according to common terminology criteria for adverse effects (CTCAE) v4.0. Then, the patients were assigned as grade 2 or lower to predict moderate-to-severe xerostomia. Over 825 dosiomic features were extracted from each contour of ipsi- and contra-lateral parotid gland, and bi-lateral parotid glands, and the redundant features were eliminated using the Spearman's correlation coefficient. A total of 21 conventional predictive factors such as DVH parameters and an estimated salivary output from the parotid glands were also considered for a model building. All patients were divided into the training cohort of 93 patients and the independent validation cohort of 40 patients, and informative features were selected by Bruta algorithm in the training cohort. Xerostomia at 12 months was predicted using random forest in the validation cohort, and area under the curve (AUC) was used for evaluating the predictive accuracy. DeLong test was used to compare AUC values between the different models and the calculated P-values were adjusted using the Bonferroni correction for multiple comparisons.Xerostomia grade < 2 and 2 at 12 months were 75 (56%) and 58 (43%) patients, respectively. There were no patients who developed grade 3 xerostomia or more at 12 months. Although the predictive accuracy improved in the combined model, no statistically significant difference was shown between the models. AUCs of the dosiomics, conventional, and combined models were 0.67, 0.69, and 0.77, respectively. Nine features (8 from dosiomics and 1 from conventional factors) were selected as informative features for the combined model.The predictive accuracy of dosiomics model was comparable to that of the model with conventional predictive factors. Some dosiomic features could improve the predictive accuracy for predicting the moderate-to-severe xerostomia in patients with oropharyngeal cancer.