Program Information

Using Machine Learning to Predict Radiation Induced Toxicities for Oropharyngeal Cancer

T Cui

T Cui*, M Ward , E Murray , J Potter , J Dorfmeyer , N Joshi , J Greskovich , S Koyfman , P Xia , The Cleveland Clinic Foundation, Cleveland, OH

Presentations

TH-AB-FS4-10 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: Four Seasons 4

Purpose: To establish a machine learning methodology to predict the incidence of toxic effects after definitive radiotherapy for oropharyngeal cancer.

Methods: 210 patients treated with IMRT for oropharyngeal carcinoma at our institution from 2009 to 2016 were retrospectively identified and randomly divided into two datasets for training and validation, respectively. For each patient, more than 40 features were retrieved, including patient’s characteristics and dosimetric endpoints of various organs at risk (OARs). A feature selection technique was developed to filter out clinically relevant features for the training dataset. Four different machine learning models, multinomial logistic regression (MLB), random forest (RF), naïve Bayes (NB), and support vector machine (SVM), were trained on the training dataset to predict grade ≥2 acute/late xerostomia and dysphagia, respectively. The prediction accuracy of each model was evaluated on the validation dataset using receiver operating characteristics (ROC) analysis. Each model was trained multiple times to determine a subset of features with the best prediction accuracy.

Results: 102, 72, and 39 of the patients had experienced grade ≥2 acute dysphagia, acute xerostomia, and late xerostomia, respectively, whereas late dysphagia was excluded from modeling due to the low incidence. Evaluated by the area under the ROC curve (AUC), MLB model with treatment technique, HPV status, disease subsite, sex, and pharyngeal constrictors Dmean as the subset of features resulted in the most accurate prediction for acute dysphagia (AUC = 0.77), RF with age, grouped stage, sex, and submandibular Dmean for acute xerostomia (AUC = 0.77), and SVM with disease subsite, parotid V30, and submandibular Dmean for late xerostomia (AUC = 0.80), respectively.

Conclusion: The proposed machine learning-based methodology can help clinicians and planners to evaluate the risk of grade ≥2 acute dysphagia and acute/later xerostomia, and therefore to improve the quality of treatment planning and reduce the radiation induced toxic effects.


Contact Email: