AbstractThe shape of extracellularly recorded action potentials is a product of several variables, such as the biophysical and anatomical properties of the neuron and the relative position of the electrode. This allows for isolating spikes of different neurons recorded in the same channel into clusters based on waveform features. However, correctly classifying spike waveforms into their underlying neuronal sources remains a main challenge. This process, called spike sorting, typically consists of two steps: (1) extracting relevant waveform features (e.g., height, width), and (2) clustering them into non-overlapping groups believed to correspond to different neurons. In this study, we explored the performance of Gaussian mixture models (GMMs) in these two steps. We extracted relevant waveform features using a combination of common techniques (e.g., principal components and wavelets) and GMM fitting parameters (e.g., standard deviations and peak distances). Then, we developed an approach to perform unsupervised clustering using GMMs, which estimates cluster properties in a data-driven way. Our results show that the proposed GMM-based framework outperforms previously established methods when using realistic simulations of extracellular spikes and actual extracellular recordings to evaluate sorting performance. We also discuss potentially better techniques for feature extraction than the widely used principal components. Finally, we provide a friendly graphical user interface in MATLAB to run our algorithm, which allows for manual adjustment of the automatic results.
Spike sorting with Gaussian mixture models
