Joint waveform and firing rate spike-sorting for continuous extracellular traces

This paper discusses recent work in automatic spike-sorting for continuous extra-cellular cortical traces. Our spike-sorting framework jointly models neuronal firing times and corresponding action potential waveforms as a discrete-state latent variable process. We model the likelihood of the observed firing occurrence times as the aggregation of multiple hidden point processes based on inter-arrival probability distributions. We evaluate our method on two real, continuous, partially labeled recordings of extracellular traces from rat hippocampus obtaining total error rates (false positives + false negatives) of 5.60% and 1.86% in clean conditions, outperforming both a Gaussian mixture model (GMM) baseline and the state-of-the-art WaveClus method. Our method continues to outperform in the presence of added noise on the same data. We then perform an empirical study of two free parameters for our method on a semi-artificial dataset. We find that our method is more sensitive to parameter tuning in more difficult data and noise conditions.