Predicting xerostomia after IMRT treatments: a data mining approach

Abstract

Xerostomia is one of the most frequent long term side-effects experienced by head-and-neck cancer patients undergoing radiation therapy, reducing drastically the quality-of-life of patients. In the present study, a prediction model for xerostomia after radiotherapy is proposed. Model construction was based on a dataset of 138 patients with head-and-neck cancer treated at the Portuguese Institute of Oncology of Coimbra (IPOCFG) with Intensity Modulated Radiation Therapy, using different data mining predictors. The models considered dosimetric information and patient specific features known prior to treatment to estimate which patients will experience xerostomia (G0 vs G1/G2 according to RTOG/EORTC). The quality of the classifiers was assessed by applying cross-validation procedures and was validated by different datasets. ROC/AUC, precision and recall were the measures used to evaluate the models' performance. Age, gender, severity of xerostomia prior to radiation therapy and planned mean (physical) dose in both parotids revealed to be relevant predictors of xerostomia. The best model was the one based on random forests. The method produced an AUC equal to 0.73, a precision of 72% and a recall of 83% considering the threshold 0.5. The ability to discriminate patients according to their features helps to achieve personalized radiation therapy treatments. Random forests revealed to be a good classification method for predicting the binary response "risk for xerostomia induced by radiation therapy at 12 months", showing a high discriminative ability.

http://ift.tt/2sK7JZ1