Editorial overview: Large-scale recording technology: Scaling up neuroscience

Comparison of combined spike detection and clustering using mutual information

BACKGROUND

Spike sorting techniques involve both detection of spike waveform events and classification of those events into clusters of similar waveform shape. The one existing method of evaluating the combined effects of both detection and classification depends on assignment of cluster correspondence. Other methods of evaluation have focused on either clustering or detection, but not both, although these two steps may interact.

NEW METHOD

This paper develops an information theoretic measure of agreement between the output of two spike sorting techniques, AMI_all, which can be used even when the number of waveform events detected by the two techniques differs.

RESULTS

AMI_all is shown to be a useful measure for studying variations of parameters of spike sorting techniques in two examples: comparing outputs for simulated noisy spike sorting and spike sorting of human single neuron recordings. Comparison with existing methods Computing AMI_all does not require an explicit assignment of cluster correspondence, thereby eliminating a potential source of variation. By providing a single measure of performance, computing AMI_all is very useful when comparing large numbers of algorithmic or parametric variations of spike sorting techniques; prior comparison techniques have often required multiple measures of performance which complicates large scale comparisons.

CONCLUSIONS

The use of AMI_all to measure agreement between spike sorting techniques facilitates the comparison of the outputs of those techniques, including variations in both spike detection and waveform clustering. This measure should be useful for broad based and large scale comparisons between spike sorting techniques.