A Study of Deep Clustering in Spike Sorting

TL;DR

This paper compares deep clustering algorithms with traditional methods for spike sorting, finding that deep clustering performs better, especially with complex datasets.

Contribution

The study introduces a large-scale benchmark of deep clustering algorithms for spike sorting, showing their superiority over traditional methods.

Findings

Deep clustering algorithms like ACeDeC, DDC, DEC, IDEC, and VaDE outperform traditional methods in spike sorting.

These algorithms excel in capturing complex spike data structures through non-linear representations.

Deep clustering combines feature extraction and clustering, improving accuracy in neuronal activity identification.

Abstract

Spike sorting is the process of identifying the source neurons for neuronal activity recorded from extracellular electrodes. Traditional spike sorting pipelines separate the process into distinct feature extraction and clustering steps, which may not optimally capture the complex structure of spike data. This study provides a large-scale benchmark of 12 deep clustering algorithms against traditional feature extraction methods combined with K-means clustering for spike sorting. We analyze performance across 95 synthetic datasets with varying cluster counts (2-20) and complexity from the perspective of six performance metrics. Our results demonstrate that a subset of deep clustering algorithms—particularly ACeDeC, DDC, DEC, IDEC and VaDE—significantly outperform traditional methods, especially as dataset complexity increases. These deep clustering approaches effectively learn non-linear representations that better capture the structure of spike data while simultaneously optimizing clustering objectives. This dual optimization produces feature spaces tailored for clustering, combining the two traditionally separate steps of spike sorting. Our findings indicate that deep clustering approaches are most suitable for accurately identifying individual neuronal activity in extracellular recordings, providing guidance for method selection for the increasingly complex modern multi-electrode recordings.

The online version contains supplementary material available at 10.1007/s12021-025-09751-4.

Figures