Intelligent software for spike separation in multiunit recordings.

Abstract

An algorithm allowing unsupervised spike separation on the basis of three parameters is described. It rests on the assumption that the first incoming spike is part of the first family of spikes. The second spike is compared to the first and, if similar, is included in the first family; otherwise, it constitutes the first member of the second family, and so on. From the first to the 30th spikes of a given family, similarity is established on a percentage basis (10 per 100 around the family's centre of gravity), and when the family includes over 30 spikes, similarity is measured by the standard deviation (1·96 SD around the family's centre of gravity). As new families occur, the families are sorted according to the number of spikes they comprise. The algorithm, its implementation and related software are fully described. Results were tested with both artificial and natural material. Using artificial spikes as input, it was demonstrated that on average 68 per cent of spikes were correctly classified, 30 per cent were rejected, and only 2 per cent were wrongly classified. For natural spike inputs, 65 per cent of recorded spikes were classified, and their separation into several families were confirmed on a physiological basis.