Spike Sorting

Spike sorting is the process of assigning each spike in an extracellular recording to the neuron that produced it, so that a channel of threshold crossings becomes a set of units with their own spike trains. It runs in four steps: detect threshold crossings in the spike band, extract each waveform's features, cluster similar waveforms, and validate each cluster against isolation criteria. Automated sorters — Kilosort, MountainSort, SpyKING CIRCUS and others — handle thousands of spikes across hundreds of channels; manual curation is still the norm for analyses that depend on single units.

The electrode decides how well sorting can work, because a sorter separates neurons on how differently they appear across neighboring sites. A neuron seen on one site is a waveform; a neuron seen on four sites at four amplitudes is a position, and two cells with similar waveforms are separable by position when they are not by shape. That is what a tetrode cluster, a polytrode's 20 to 25 µm spacing and a high-density shank provide, and why site pitch ↗ inside the detection radius is chosen for sorting rather than for coverage. Overlapping spikes, drift over a session and low signal-to-noise are the failure modes, and a stable, low-impedance site with a tight impedance distribution across the probe is the defense against the last of them. Quality is reported per unit as isolation distance, signal-to-noise and refractory violations. Sorting should not be the first place a bad session is discovered: the impedance distribution before the session and the live signal-quality view during it, which Allego provides, show the site and reference problems that sorting can only report afterwards.