Standardized assessment of extracellular single unit isolation quality

Single unit ensemble studies utilize multiple extracellular single unit recordings from electrodes in an array or tetrode. A neuron can be identified if its spikes have a distinct combination of features across different electrodes: features clustering in feature space. Clustering is typically done manually but has also been automated. Ideally, spikes of individual neurons cluster clearly in scatter plots in multiple 2D cuts of feature space. However, the quality of cluster separation may vary with operator expertise, choice of 2D slices, and visual assessment of cluster boundaries. Thus, it is valuable to quantify cluster isolation quality. Isolation distance and Lratio are a pair of such measures proposed as a standard (Schmitzer-Torbert et al., 2005, Neurosci., 131:1). Both measures are defined in a space where features are restricted to a spike's energy and first principal component of the energy-normalized spike. The standard was chosen to make comparison across data sets straightforward. These two features rarely provide optimal waveform isolation and we find that a variable number of features are needed for good isolation. We therefore sought to provide a measure with a straightforward interpretation when data sets are described in feature spaces differing in component features. Our method quantifies the information about the waveform feature distribution a cluster's identification provides. We compute a metrical distance between a cluster's waveform feature distribution and the waveform feature distribution of spikes not in the cluster using a symmetrized Kullback Leibler divergence (KLD). This quantifies the cluster's isolation quality in units of information, bits. KLD is proportional to the number of component features that are used, but not to the specific features. To facilitate comparison of KLD values across different feature spaces of arbitrary dimensionality, we perform a reduction as follows: for each cluster, the 8 features giving the highest pairwise 2D KLDs are used to compute the final 8D KLD. These automatically chosen features typically match the choices of expert users and maximize cluster isolation and compactness. There are advantages to using information theoretic measures over measures like isolation distance and Lratio. Measuring the information a single unit classification provides about waveform features has intuitive appeal and a standard interpretation when different features are used for the isolation. The measure may facilitate single unit isolation by giving researchers freedom to choose the most discriminative waveform features and allowing standardized verification of data quality.