A Robust Method for Spike Sorting with Overlap Decomposition

Abstract

It is difficult to identify the spikes to different classifications especially when the neurons have many similar spike waveforms or lots of overlapped spikes. Our previous study proposed the window-slope representation (WSR), and it improved the classification accuracy of high similar spike waveforms. The classification accuracy of the method, however, will be affected by lots of overlapped spikes or low signal-to-noise ratio. In this paper, the secondorder difference method is introduced to solve those problems. First the second-order difference of every spike is calculated to describe the convexity-concavity of the waveform. Then we use the window-slope representation to describe the tendency of waveform at each time. The method is tested at various signal-to-noise ratio levels based on simulation data coming from the Wave_clus. The experiment results show that the accuracy of classification can be improved by using the method together with K-means Cluster. In our experiments, the classification accuracy is above 95% on all datasets in Wave_clus, even under the low signal-to-noise ratio. It shows that this method has a strong robustness.

ELVISort: encoding latent variables for instant sorting, an artificial intelligence-based end-to-end solution

“…This trend creates a need for fast, reliable and highly automated data processing algorithms. Using extracellularly implanted MEAs, raw data is typically recorded at a high sampling rate (20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30) in order to capture temporal features of individual action potentials of nearby neurons (single-unit activities, spikes). To replace the slow and time-inefficient human evaluation of these data, spike detecting and sorting algorithms have emerged to automate the process.…”

“…To properly cluster spikes even in these cases, a robust feature extractor is needed. For this, several methods have been proposed, such as feature extraction based on principal component analysis (PCA) [19][20][21], wavelet coefficient [22], wavelet packet coefficient [16], and wavelet packet decomposition used with support vector machine [23].…”

“…the segmentation of extracted features, has a wide range of approaches as well. The most widely applied semiautomatic method is kmeans clustering, which is featured in many research papers [11,20,24], combined with other methods. Other approaches such as Bayesian classification [10], superparamagnetic clustering [16], support vector machine classification [25], and independent component analysis [24] are also used in the segmentation phase of spike sorting.…”