Evaluation of spike sorting algorithms: Simulations and application to human Subthalamic Nucleus recordings

An important prerequisite for the analysis of spike synchrony in extracellular recordings is the extraction of single unit activity from the recorded multi unit signal. To identify single units (SUs), potential spikes are detected and separated with respect to their potential neuronal origins ('spike sorting'). However, different sorting algorithms yield inconsistent unit assignments which seriously influences the subsequent analyses of the spiking activity. To evaluate the quality of spike sortings performed with different prevalent algorithms offered by the 'Plexon Offline Sorter' we first apply these algorithms to experimental data (ED), namely recordings in the Subthalamic Nucleus of patients with Parkinson's disease, obtained during Deep Brain Stimulation surgery. Since this procedure leaves us unsure about the best sorting result we then apply all methods again to artificial data (AD) with known ground truth (GT). AD consists of pairs of SUs with different shape similarity embedded in the background noise of the ED. The sorting evaluation is based on the influence of the respective methods on the SU assignments and its effect on the resulting firing characteristics. We find a high variability in the sorting results obtained by different algorithms that increases with SU shape similarity. We also find significant differences in the resulting firing characteristics of the ED. We conclude that Valley-Seeking produces the most accurate results if the exclusion of artifacts as unsorted events is important. If the latter is less important ('clean' data) K-Means is a better option. Our results strongly argue for the need of standardized validation procedures for spike sorting based on GT data. The recipe suggested here is simple enough to become a standard procedure.