EEG signal classification using Neural Networks

The aim of electroencephalography is to record and measure samples of electromagnetic fields during certain states and sequences of behavior, in order to explain some of the mechanisms by which behavior is generated. The electroencephalographic (EEG) signals represent electrical changes of the brain during its function. The digital signal processing of EEG, especially the spectra analysis, yields more meaningful information than the visual inspection of the EEG curve. However, there are still problems encountered like the artifacts during the registration, the individual differences among normal and among diseased brains, and the uncontrolled conditions of brain functions. Additional problems arise due to the ambiguities of the algorithms and the methods used for analysis. It is believed that Neural Network methods can be utilized to classify pathological and normal EEG signals because of their ability to discriminate among EEGs with minor differences. To test this hypothesis the EEG signals of 12 positions over the scalp for 12 car-painters (exposed to neurotoxic organic solvents) and 12 normal volunteers were registered and analyzed. The existing results of the neurophysiological examination of persons exposed to organic solvents are controversial because of the slight differences existing in their EEG signals. Comparing the changes of the power spectrum and the coherence between rest and task conditions, statistical important differences, using the Wilcoxon test,were found in two measurements between car-painters and normal volunteers out of twelve utilized in the power spectrum. Important statistical differences were not found at all in the ten coherence values used.

In this study, a modified Back-propagation learning rule, which results in faster learning rates, is implemented. Three different ANN experiments are performed for EEG classification. The first experiment is based on twelve power spectrum EEG measurements. Ten EEG coherence values are used as inputs to the second ANN experiment. Finally, both power spectrum and coherence EEG indications (twenty-two values) are combined to constitute the inputs for the third ANN. Preliminary results indicate that on the average 90% of the cases are correctly classified by the Neural Networks. Detailed results of the ANN architectures and parameters used along with statistical method comparisons are presented.

George Bebis
November 22, 1995 at 5:22 PM