A non-parametric Bayesian approach for clustering and tracking non-stationarities of neural spikes

Neural spikes from multiple neurons recorded in a multi-unit signal are usually separated by clustering. Drifts in the position of the recording electrode relative to the neurons over time cause gradual changes in the position and shapes of the clusters, challenging the clustering task. By dividing the data into short time intervals, Bayesian tracking of the clusters based on Gaussian cluster model has been previously proposed. However, the Gaussian cluster model is often not verified for neural spikes.

We present a Bayesian clustering approach that makes no assumptions on the distribution of the clusters and use kernel-based density estimation of the clusters in every time interval as a prior for Bayesian classification of the data in the subsequent time interval.

The proposed method was tested and compared to Gaussian model-based approach for cluster tracking by using both simulated and experimental datasets.

The results showed that the proposed non-parametric kernel-based density estimation of the clusters outperformed the sequential Gaussian model fitting in both simulated and experimental data tests.

Using non-parametric kernel density-based clustering that makes no assumptions on the distribution of the clusters enhances the ability of tracking cluster non-stationarity over time with respect to the Gaussian cluster modeling approach.