Blinking Artifact Recognition in EEG Signal by Neural Network

Recognition of blinking artifacts was the subject of my Master Project. I have created the computer program, which finds blinking artifacts in EEG signal. The main screen of the program is shown below. On this page you can also find abstract of my Master thesis.

This is the main screen of the Artefact system which finds blinking artifacts in EEG signal:

Click here to download the executable version of the Artefact system

Click here to download the postscript version of my Master Thesis (in Polish)

Abstract of my Master Thesis

Artifacts are noises introduced to an EEG signal by patient's movements and sources of electric field outside the patient's body. The artifacts impede a doctor?s expertise and an automatic analysis of the signal. The most common and characteristic kind of artifacts are blinking artifacts. The aim of this work was to create a computer program finding blinking artifacts in the EEG signal. There is no a set of rules determining if the fragment of the signal does contain a blinking artifact, or does not. Hence, neural networks were chosen as a classification tool. The input to the network is not a raw sampled signal, but different coefficients computed for a window of one second of the signal. The window moves by 0.25 second of signal. To create training set for the network, one analysed 500 MB of EEG signal, from which one chose over 27000 the most representative windows, containing different kind of artifacts, pathological and proper waves. For each window, one computed 41 coefficients expressing some characteristic properties of blinking artifacts. Some of the coefficients were designed by the author and were based on his knowledge about the artifact recognition. Then the network was trained. Other coefficients, which were introduced, were based on an analysis of the signal's fragment, which were incorrectly classified by the network. The sensitivity analysis was used to choose the coefficient bringing the larger amount of information about the presence of an artifact. Then, correlated coefficients were eliminated and the number of neural network's inputs was limited to 14. Three classification algorithms were tested: k-neighbours, RBF networks and back propagation networks. The lowest classification error for the test set was obtained for the back propagation network with 5 hidden units. The program was tested on the EEG signal and was highly evaluated by a domain expert.

Back to Rafal Bogacz Homepage