Robust and Automated Sorting Algorithm Using General Spike Detection

Background and Aim : Neural activity monitoring is the basis for understanding the brain behavior. The recorded activity is a combination of multiple neurons activities corrupted by noise. A main step in analyzing of this data, is to differentiate among different neuron activities. Spike sorting is the process of assigning each detected spike to the corresponding neurons. That is to say, spike sorting is a clustering procedure where spikes are the samples that are clustered. Thus spikes in one cluster are assumed to belong to the specific neuron. There exist several challenges for spike sorting algorithms: complex noise, non-Gaussian, skewed cells and time overlapped spikes due to the simultaneous activity of several neurons. While different spike sorting algorithms are developed, yet there exists no universal algorithm that performs well in all situations.Methods : Mixture modeling is one of the successful clustering algorithms usually used in spike sorting methods. Primary works focused on Gaussian mixture models, however, it has been shown that mixture of t-distributions is more powerful than Gaussian mixtures in modeling neural data. By using mixture of Gaussian or t-distribution, we assume that the clusters are symmetrical, however, skewed clusters are reported in the literature. In this paper, we propose a new clustering method based on skewed t-distribution. Our proposed clustering method, could handle non-symmetrical clusters alongside preserving powerful features of symmetrical t-distribution like heavy tails. In addition, we introduce two preprocessing algorithms, which are a new alignment algorithm, and a statistical filtering algorithm for noise removal. In alignment phase, we propose to align detected spikes based on multiple target (aligning) points corresponding to the shapes of other spikes, instead of just aligning according to their extremum. For noise removal, the statistical characteristics of the spike shapes are used and false alarms (noise detected as spike) are removed accordingly.Results : The performance of the proposed method is investigated using artificially simulated and real datasets. Results show that the preprocessing algorithms improve the recall and precision of detection phase. Also, the alignment algorithm improves the compactness of clusters in terms of within cluster distance. The noise removal algorithm is also compared with a manual sorting results, which shows that, on average, 90% of removed spikes by the algorithm is also removed by human operator. In the clustering phase, results on simulated data shows an improvement over purity and accuracy of the sorting results. Also, the automatically determined number of clusters, is closer to the manual sorting results, than mixture of t-distributions method. Conclusion : The proposed method has three main contributions: 1) considering skewness for clusters, 2) new alignment algorithm, 3) statistical filtering for noise removal. We show that using the proposed method have some advantages in both detection and clustering phases. In detection, the correct detection rate is increased, while the false alarm is decreased. In clustering phase, the number of neurons could be determined more precisely alongside with pure and accurate clusters. However, estimating the skewness parameters, adds additional computational complexity to the proposed method.

کلیدواژه ها:

Spike Sorting ، Mixture of Skew t-distributions ، Spike Detection ، Spike Alignment