Comparing Spike Sorting Algorithms on Simulated Extracellular Multi-Electrode Array Recordings

Spike sorting plays a crucial role in extracting neural information from high-density multi-electrode array extracellular electrophysiological recordings. Despite the recent proposals of numerous spike sorting algorithms, the presence of inherent biases within these algorithms can result in significant variations in sorting performance, even when applied to the same recording data. Thus, a comprehensive and detailed comparative analysis of spike sorting algorithms remains unexplored. In this study, we address this gap by utilizing an extracellular electrophysiological simulator to generate synthetic recordings, which we use as ground truth for evaluating eight different spike sorting algorithms. We leveraged single-cell information encompassing electrophysiology and morphology from the Allen Brain Atlas in our simulation. With synthetic recordings, we described spike sorting performance for each algorithm by calculating temporal agreement and template similarity matrices against simulation ground truth. We performed extensive inter-algorithm comparisons and ground truth validation, such as precision-recall analyses and drift studies, to rigorously assess each algorithm. Our results reveal the precision-recall trade-off in spike sorting and highlight the two categories of intrinsic biases among different spike sorting algorithms. Our findings shed light on important considerations for selecting spike sorting algorithms and developing next-generation spike sorting algorithms.