Abstract
Chronic extracellular recordings are a powerful tool for systems neuroscience, but spike sorting remains a challenge. A common approach is to fit a generative model, such as a mixture of Gaussians, to the observed spike data. Even if non-parametric methods are used for spike sorting, such generative models provide a quantitative measure of unit isolation quality, which is crucial for subsequent interpretation of the sorted spike trains. We present a spike sorting strategy that models the data as a mixture of drifting t-distributions. This model captures two important features of chronic extracellular recordings—cluster drift over time and heavy tails in the distribution of spikes—and offers improved robustness to outliers. We evaluate this model on several thousand hours of chronic tetrode recordings and show that it fits the empirical data substantially better than a mixture of Gaussians. We also provide a software implementation that can re-fit long datasets (several hours, millions of spikes) in a few seconds, enabling interactive clustering of chronic recordings. Using experimental data, we identify three common failure modes of spike sorting methods that assume stationarity. We also characterize the limitations of several popular unit isolation metrics in the presence of empirically-observed variations in cluster size and scale. We find that the mixture of drifting t-distributions model enables efficient spike sorting of long datasets and provides an accurate measure of unit isolation quality over a wide range of conditions.
