Semi-supervised self-training approaches in small and unbalanced datasets: Application to xerostomia radiation side-effect

Abstract

Supervised learning algorithms have been widely used as predictors and applied in a myriad of studies. The accuracy of the classification algorithms is strongly dependent on the existence of large and balanced training sets. The existence of a reduced number of labeled data can deeply affect the use of supervised approaches. In these cases, semi-supervised learning algorithms can be a way to circumvent the problem. In the present study, we apply several semi-supervised learning methodologies to a small clinical dataset with 222 (138 labeled and 84 unlabeled) head-and-neck cancer patients treated at the Portuguese Institute of Oncology of Coimbra (IPOCFG) with Intensity Modulated Radiation Therapy (IMRT). In order to predict the aptness for xerostomia induced by radiation treatments, we considered random forest classifiers. Xerostomia is one of the most frequent long term side-effects experienced by head-and-neck cancer patients undergoing radiation therapy, reducing drastically their quality- of-life. Therefore, being able to predict xerostomia at early stages of the treatment would make it possible to adjust the treatment plan in order to minimize or avoid this complication The quality of the semi-supervised classification rule was validated by using different subsets of patients. Our experiments evidenced an improved performance of the classifier as the size of the training labeled dataset increased.