Handling class label noise in medical pattern classification systems

Pattern classification systems play an important role in medical decision support. They allow to automatize and speed-up the data analysis process, while being able to handle complex and massive amounts of information and discover new knowledge. However, their quality is based on the classification models built, which require a training set. In supervised classification we must supply class labels to each training sample, which is usually done by domain experts or some automatic systems. As both of these approaches cannot be deemed as flawless, there is a chance that the dataset is corrupted by class noise. In such a situation, class labels are wrongly assigned to objects, which may negatively affect the classifier training process and impair the classification performance. In this contribution, we analyze the usefulness of existing tools to deal with class noise, known as noise filtering methods, in the context of medical pattern classification. The experiments carried out on several real-world medical datasets prove the importance of noise filtering as a pre-processing step and its beneficial influence on the obtained classification accuracy.