Performance comparison of extracellular spike sorting algorithms for single-channel recordings

Jiri Wild1, Zoltan Prekopcsak3, Tomas Sieger1,
Daniel Novak1, Robert Jech2

1Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University, Karlovo nam. 13, 121 35 Praha 2, CZ
2Department of Neurology, 1st Faculty of Medicine and General Teaching Hospital, Charles University in Prague, Katerinska 30, 128 21 Praha 2, CZ
2Department of Telecommunications and Media Informatics, Budapest University of Technology and Economics, Magyar tudosok korutja 2., H-1117 Budapest, HU

 

Abstract

Proper classification of action potentials from extra-cellular recordings is essential for making an accurate study of neuronal behavior. Many spike sorting algorithms have been presented in the technical literature. However, no comparative analysis has hitherto been performed. In our study, three widely-used publicly-available spike sorting algorithms (WaveClus, KlustaKwik, OSort) were compared with regard to their parameter settings. The algorithms were evaluated using 112 artificial signals (publicly available online) with 2-9 different neurons and varying noise levels between 0.00 and 0.60. An optimization technique based on Adjusted Mutual Information was employed to find near-optimal parameter settings for a given artificial signal and algorithm. All three algorithms performed significantly better (p<0.01) with optimized parameters than with the default ones. WaveClus was the most accurate spike sorting algorithm, receiving the best evaluation score for 60% of all signals. OSort operated at almost five times the speed of the other algorithms. In terms of accuracy, OSort performed significantly less well (p<0.01) than WaveClus for signals with a noise level in the range 0.15-0.30. KlustaKwik achieved similar scores to WaveClus for signals with low noise level 0.00-0.15 and was worse otherwise. In conclusion, none of the three compared algorithms was optimal in general. The accuracy of the algorithms depended on proper choice of the algorithm parameters and also on specific properties of the examined signal.

 

Acknowledgement

This work has been supported by Czech Science Foundation research program 309/09/1145 and by Grant Agency of the Czech Technical University in Prague; grant No. SGS10/279/OHK3/3T/13. The second author has been supported by Visegrad Fund scholarship No. 51000739.

 

Artificial data

JW Artificial data - 20s, 60s signals (mat-files ~ 664MB)
JW Artificial data - 960s signals (mat-files ~ 4.9GB)
QQ Artificial data (→ WaveClus website)

 

Result tables

Artificial
data
Signal
length
Parameters
used
JW short default html
JW short optimized html
JW long default html
JW long optimized html
QQ long default html
QQ long optimized html
Annotated list of parameters