Adaptive Spike Sorting with a Gaussian Mixture Model

Abstract

Spike sorting is an important step in the processing of action potentials recorded from electrodes implanted into the brain. When such signals are acquired over long time spans, the shape of action potential waveforms often changes, requiring corresponding changes in spike sorting algorithm parameters. To avoid manual update of these parameters, an adaptive spike sorting method is needed. Here we present an adaptive spike sorting method based on the Gaussian mixture model and variational Bayesian inference. This approach treats classification of new spikes as clustering under a Gaussian mixture model with a strong Bayesian prior comprising of the previous clustering parameters. The result of this clustering gives the label of the new waveform as well as updates the clustering parameters. We also include a model for how clustering parameters change over time. We demonstrate this method on synthetic data as well as real neural data.