Spiking neural network based classification of task-evoked EEG signals

This paper presents an improved technique to detect evoked potentials in continuous EEG recordings using a spiking neural network. Human EEG signals recorded during spell checking, downloaded from the BC1 Competition website, were pre-processed using a Wavelet Transform to remove the noise and to extract the low frequency content of the signal. Analysis of the signals was performed on the ensemble EEG and the task of the neural network was to identify positive and negative peaks of different shapes. The network has a time-warp invariance property, which means that an input linearly compressed or elongated in time is still recognisable by the network. This enabled the network to train on one peak shape and generalize it to recognise similarly shaped peaks. The neural network presented was trained on one epoch of filtered EEG and was tested on the remaining samples. A post hoc examination of the averaged evoked EEG signal pre-designated as target and non-target show a nadir in the non-target, but not in the target signals. A new supplementary template containing a nadir was therefore created and the effectiveness of this was tested on the ability of the network to correctly identify evoked EEG. After final testing 94.7% of the signals assigned as containing P300 by the paradigm used for the data on the website were correctly classified as P300s, and 83.7% of the non-P300s were also classified as non-P300s. The sensitivity of the technique, utilising the data from this paradigm was 94.7%, specificity 83.68%, and positive predictive value was 53.71%.