An Incremental Approach to Clinical EEG Data Classification

Abstract

Experimental verification of two approaches of incremental learning to clinical EEG data classification is presented. Three datasets were used to evaluate their performance. In particular sleep, artifact and newborn EEG signals. Most accurate classifiers that are able to detect unwanted elements in EEG signal are found using artifact data set and tested on sleep and newborn datasets. Classification is performed using two incremental classifiers: Support Vector Machine and 1-Nearest Neighbors. The classifiers are able to classify sleep (10%) and newborn (0.5%) datasets by learning the shortest part of the data set with sufficient accuracy at least 70%. Better and quicker results were obtained by random rather than sequential selection of data. In conclusion it means that the classification is made fast. The incremental approach is supposed to save time, which neurologists spend during manual EEG signal scoring.