A computational framework for effective isolation of single-unit activity from in-vivo electrophysiological recording

Modern spike sorting techniques are heavily reliant on unsupervised machine learning algorithms for isolation of single-unit activity from noisy channel recordings. One of the most common methods, k-means clustering, is highly sensitive to the number of clusters selected (k) prior to analysis. A robust automated method for determining k is required, in particular for the large datasets currently being analyzed by the Neuroscience community. Information criteria, often applied for this analysis, can yield over-fitted clustering recommendations and employ strong assumptions about cluster gaussianity which do not necessarily hold for real in-vivo neuro-electrophysiological recordings. An algorithmic approach to spike sorting is applied utilizing tandem multi-level wavelet decomposition and principal component analysis to construct a discriminant feature space. K-means clustering is applied to this feature space using a variety of distance metrics to determine which approach yields optimal cluster separation. Clustering outcomes are evaluated using the Entropic Product, an information entropy-based measure that makes no assumptions about the underlying distribution of spikes within a cluster. This measure is demonstrated to be more informative about clustering outcomes than other information criteria when sorting spike data collected using bundled microwire arrays implanted in the Primary Visual Cortex of marmosets conducting a visual-stimulation task.