Method for automatic detection and classification of N1 and P2 auditory evoked potentials in EEG recordings

The current averaging method for event related evoked potentials does not allow to study the potentials´ variability, thus a detection algorithm that allows to recognise the potentials elicited by each stimulus is desirable. This paper compares Neural Networks and Supported Vector Machines as classifiers to be integrated to an automated detection algorithm. EEG recordings of 5 subjects (3 female, 2 male) were made, while auditory stimulus was provided. The recordings were filtered with a 2^nd order Butterworth lowpass filter with cutoff frequency at 40 Hz. A free database provided by California Institute of Technology with several auditory evoked potential (EP) events was used to train the models used to identify de N100-P200 complex. Once a model was achieved, it was first validated with new data, and then incorporated to an algorithm that identified the complex in a non segmented EEG recording. An average sensitivity of 93.26% was achieved with only 136 false positives in over 25 minutes of 6 channel EEG recording. These results prove that individual detection of EPs is possible, thus enabling future studies in variability.