High channel live spike sorting with GemSORT (Graph nEtwork Multichannel neural spike sorting)

The brain processes information by sending small electric events called action potentials or spikes between neurons. Thus, one obvious way to study how the brain works involves recording and analyzing these spikes over time, which is most commonly done with a technique called extracellular in-vivo recordings. While such recordings have been performed for decades, there have been several break-through innovations during the last few years which have greatly enhanced the power and the appeal of this technique. One of these is the advent of multi-channel and high-channel recording equipment that enables an investigator to record neural activity through dozens or even hundreds of channels and thus acquire a much more accurate and complete picture of how populations of neurons process neural information. 

A major challenge which is growing with the increase in the number of recording channels is the ability to process these large amounts of incoming data, particularly in real-time. Spike sorting, a common first step in analyzing neural activity measured from a multichannel probe, is one such bottleneck. Due to the close proximity of neurons to each other, whenever an electrode is placed into the brain, it will often pick up electrical signals from more than one neuron, resulting in the spike waveforms of several neurons being electrically combined into so-called “multi-unit” activity. Spike sorting algorithms are then used to separate this multi-unit activity into several sets of “single-unit” activity, each of which represents the action potential firing pattern of a single neuron. Although a number of sophisticated spike sorting algorithms have been developed, many of these are relatively slow and designed to sort data post-hoc after the conclusion of a recording session. This is especially true for recordings consisting of many or even hundreds of channels. 

Additionally, due to the mathematical approaches many algorithms were based upon, they have to examine the entire set of neural spikes only once it is completely recorded in order to come up with the best sorting accuracy, and thus can only provide sorting outcomes at the very end of the analysis. Some algorithms do allow for near-live sorting. However, due to computational demands, the investigator is asked to select a very small subset of the incoming data stream such as one channel out of many for sorting. Therefore, novel mathematical approaches to sort individual spikes with a short processing time are needed. 

The GEMsort algorithm is a real-time neural spike sorting algorithm using a novel mathematical approach of graph clustering. Using this new mathematical approach, the algorithm can learn the dissimilarities between neural spikes fired from different neurons as the neural spikes are being measured in real-time. With this unique graph clustering approach, a neural spike can be analyzed and sorted as soon as the spike is measured with very minimal processing time (< 1 ms). Additionally, the GEMsort algorithm is designed to allow parallel computations. When combined with the latest state-of-the-art digital electronics, the algorithm can provide real-time neural spike sorting even when neural spikes are simultaneously measured from hundreds of recording channels. This unique real-time sorting capability can open the door to new approaches to studying brain circuitry, including real-time neural dynamic decoding and closed-loop neural feedback stimulations based on the real-time analyzed neural dynamics.

Another unique capability of the GEMsort algorithm is its ability to self-adapt to changes in the electrophysiology environment, such as probe drifts during long recording sessions which cause profile changes in the electrical signatures of neural spikes. Due to the learning on-the-fly adaptability of the algorithm, the GEMsort algorithm can “follow” changes in the neural spikes for continuous and correct sorting. This self-adapting sorting feature allows extremely long recording sessions to track neural activities without wasting any “down times” for neuroscience studies.

PopNeuron is currently developing an electronic add-on to allow investigators using their existing recording systems and additionally have the capability to sort neural spikes in real-time, even with hundreds of simultaneously recorded channels streaming a large amount of data. This electronic add-on device can simply be plugged into the recording computer, without the needs of investing on expensive GPU or GPUs. The sorting will be entirely processed within the electronic add-on, and the computational resources of the main computer will be preserved for data analysis or other experimentally required computations.

Interested in being a beta tester? Contact us!