Classification of epileptiform events in raw EEG signals using neural classifier

This study evaluated the capability of neural classifier to perform the separation between epileptiform and non-epileptiform events. To processing the EEG signals was used the Wavelet Transform through the use of the Coiflet1 function. The main elements present in the EEG signals were separated in five distinct event classes (spikes, sharp waves, blinks, background activity and noise). All the events were processed from first to the tenth decomposition level of the Wavelet Transform, where were generated graphics with the dispersion of each event class. Some experiments were made to try define a decision threshold to separate the groups of elements using the Coiflet1 function. The obtained results showed that only the amplitude of a decomposed signal don't show a distinction between the events classes. Thus, the raw epochs of EEG signals were applied directly in the neural network inputs. To evaluate the neural networks was used the method of cross-validation with early stopping. For the neural classifier was used ROC analysis and performance indexes applied to the diagnostic tests. The experiments have shown that the use of any epoch of training, indicated by the performance indexes (AUC and accuracy) showed the better results. The epochs indicated by the performance indexes were located close to the epoch indicated by the early stopping. The evaluation through of those indexes showed be an efficient method to verify the performance of the classifier, getting the following performance values: AUC index of 0,99910, sensitivity of 97,14%, specificity of 94,55% and an accuracy of 96,14%.