Assessing Ambiguity Of Spike Detection By Thresholding

Spike sorting is the art of extracting reliable neuronal activity information from noisy electrophysiologic recordings. For intracranially implanted microelectrodes and neuro-prosthetics, it is paramount to have reliable spike sorters for identifying different neuronal events to have a better understanding of neuronal networks and even brain functions. Unfortunately, up till now, neuroscientists could not agree on one trustworthy Gold Standard algorithm performing robust identification under all conditions. Even worse, already the very first step in the sorting pipeline is plagued by ambiguities and very subjective decisions. Here we compare, based on examples from one huge data set, some event detecting methods such as wavelet denoising, non-linear energy operator, shift and multiply techniques as pre-processors for thresholding, along with different pre-emphasis such as standard deviation and median absolute deviation for each pre-processors. We evaluated the performance of different spike sorting algorithms such as waveclus, kilosort, combinato, hdsort, tridesclous on our dataset. Although for the same threshold coefficient, different spike sorters had the same number of spikes, there was a slight difference in the time of the detected spike. In this scenario, the question arises, which spike sorter to trust? We suggest not relying on one method only but using several curated spikes based on a consensus decision.