Spike Sorting Using Time-Varying Dirichlet Process Mixture Models

Spike sorting is the task of grouping action potentials observed in extracellular electrophysiological recordings by source neuron. In this thesis a new incremental spike sorting model is proposed that accounts for action potential waveform drift over time, automatically eliminates refractory period violations, and can handle “appearance” and “disappearance” of neurons during the course of the recording. The approach is to augment a known time-varying Dirichlet process that ties together a sequence of infinite Gaussian mixture models, one per action potential waveform observation, with an interspike-interval-dependent term that prohibits refractory period violations. The relevant literature on spike sorting as well as (time-varying) Dirchlet process mixture models is reviewed and the new spike sorting model is described in detail, including Monte Carlo methods for performing inference in the model. The performance of the model is compared to two recent spike sorting methods on synthetic data sets as well as on neural data recordings for which a partial ground truth labeling is known. It is shown that the model performs no worse on stationary data and compares favorably if the data contains waveform change over time. Additionally, the behaviour of the model under different parameter settings and under difficult conditions is assessed and possible extensions of the model are discussed.