K-means clustering method for auditory evoked potentials selection

Abstract

Surface auditory evoked potentials are generally recorded using a headset of 32, 64 or 128 electrodes, but the quality of the responses is quite heterogeneous on the scalp surface. In some contexts, such as the analysis of auditory evoked potentials recorded in radio-frequency fields, the signal quality is essential, and it appears pertinent to consider only a limited number of electrodes. Therefore, before analysing signals influenced by radio-frequency fields, it is necessary to consider the preliminary step of selecting the channels where auditory activity is strong. This step is often realised by human visual selection and can take a considerable time. In this paper, a simple k-means clustering method is proposed, to select automatically the important channels, and the results are compared with traditional methods of selection. The method detected channels that showed a concordance rate of 86.5% with the visual selection (performed by five individuals) and gave the same final selection (only two extra electrodes in the automatic case). Moreover, the time needed for this automatic selection was 100 times less than that for the visual selection, and also human variability was avoided.