An unsupervised and refractoriness-supported algorithm design for real-time spike sorting

Neural spike sorting algorithms have been used to group the action potentials in electrophysiological signals according to their characteristics to use them in neuroscience studies. In this study, we developed an unsupervised and online spike sorting algorithm that performs better than existing online and unsupervised sorting algorithms, calculates thresholds variably, and can be used for real-time studies when applied on FPGA boards. A template matching algorithm OSort, based on Euclidean distance, functioned as the base for this algorithm. We used the NEO method to detect the spikes and the Windowed Sinc Interpolation method to upsample the detected spikes four times. We developed a refractoriness control mechanism that works according to the peak points of the spikes to prevent assigning a spike to the wrong cluster. We designed a second block that considers the probability of a spike belonging to the second closest cluster to itself. It controls this issue to prevent assigning a spike to the wrong cluster if the spike waveforms of the different neurons are similar. We avoided using more complex mathematical operations like calculating standard deviation with the aim of future real-time studies on cheaper FPGA boards. When we compared the performance of our algorithm with the well-known algorithms in the literature, we saw that ours performed significantly better than the online ones and insignificantly worse than offline ones. The presented thesis shows that more accurate online and unsupervised spike sorting is possible without complex algorithms. It is expected that this kind of algorithms will support the increment in the number of neuroscience studies that use spike sorting.