Objective: To evaluate the variability in spike sorting results using different clustering algorithms

Background: Extracellular currents recorded from the Subthalamic Nucleus (STN) of PD patients, obtained during Deep Brain Stimulation surgery, can provide important information about pathological spiking activity of STN neurons. However, a critical step in analyzing neural spike trains is the isolation of single-unit (SU) activity from the recorded multi-unit signal. To this end, a group of algorithms commonly referred to as “spike sorting” is used, which assign the observed spike waveforms in the filtered signal to putative neurons based on their similarity. However, the differences in the sorting results generated by various algorithms pose a challenge for subsequent spike train analysis.

Methods: To evaluate 3 widely used unsupervised spike sorting algorithms (K-Means, Valley-Seeking, and Expectation-Maximization provided by the “Plexon Offline Sorter”) we tested their sorting results on both artificial spike data (AD) with known ground truth and experimental data (ED) from human STN recordings. AD includes two SUs and background noise with statistical features comparable to those of ED. We took a multiple validation approach by comparing the spike times of the detected SUs to the ground truth for AD, and to a manually controlled sorting based on the template matching for ED, to determine how the sorting method influences characteristics of the SU activity (e.g. firing regularity).

Results: The spike sorting performance of each algorithm is similar when applied to AD and ED. However, the results show a wide variability between different algorithms. In AD the variability in sorting results increased if spikes of the 2 SUs had more similar shapes. Depending on the algorithm the mean number of detected SUs varied between 1 and 4.3 (AD) and between 1.3 and 3.8 (ED).  Measures such as the spiking regularity, firing rate, and refractory period violations within SUs also exhibited significant differences between the algorithms. Overall, Valley-Seeking produced the most accurate results compared to ground truth.

Conclusions: Our results strongly argue for the need of a standardized validation procedure for spike sorting algorithms based on ground truth data. Moreover, to ensure reproducibility of results a detailed description of spike sorting procedure becomes a necessity.

« Back to 2017 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/evaluation-of-spike-sorting-algorithms-applied-to-artificial-spikes-and-to-spikes-recorded-from-the-subthalamic-nucleus-of-pd-patients/