Automatic spike sorting correction and burst detection for high-density electrophysiological recordings

High-density electrophysiological probes have substantially increased the spatial resolution
and scale of neural recordings. Popular spike sorting algorithms, e.g. Kilosort (Pachitariu et al. 2016),
SpyKING CIRCUS, etc., use waveform template matching to isolate single units from the large-scale
recordings, but these algorithms make numerous errors which often require hours of manual curation
to correct. Bursting neurons pose an additional challenge in spike-sorting: the shape and amplitude
of a neuron’s waveform can change significantly over the course of a burst. Thus, waveform shape-
based template matching may fail to group together spikes from the same burst. Analyzing bursting
patterns can shed light on important aspects of brain function such as memory formation, synaptic
plasticity, and inter-region communication that would be difficult to study otherwise. Regardless, no
reliable method exists to automatically detect bursts from large-scale electrophysiological recordings. We
develop here a two-stage algorithm that corrects spike sorting errors and identifies single-unit bursts. To
automatically determine which spike clusters to merge, we devised novel metrics that quantify waveform
similarity and cross-correlation significance. To extract hierarchical bursting structure from single-unit
spike trains, we adapted a Hidden Markov model used to detect bursts in text data. On a real dataset
from mouse hippocampus, our spike sorting correction method performs comparably to a human curator.
Furthermore, we validated our burst detection model on both real and simulated data. The two stages
of our algorithm can be used independently and the algorithm is post hoc, i.e., it can be integrated into
existing spike sorting pipelines.