SLAy-ing oversplitting errors in high-density electrophysiology spike sorting

The growing channel count of silicon probes has substantially increased the number of neurons recorded in electrophysiology (ephys) experiments, rendering traditional manual spike sorting impractical. Instead, modern ephys recordings are processed with automated methods that use waveform template matching to isolate putative single neurons. While scalable, automated methods are subject to assumptions that often fail to account for biophysical changes in action potential waveforms, leading to systematic errors. Consequently, manual curation of these errors, which is both time-consuming and lacks reproducibility, remains necessary. To improve efficiency and reproducibility in the spike-sorting pipeline, we introduce here the Spike-sorting Lapse Amelioration System (SLAy), an algorithm that automatically merges oversplit spike clusters. SLAy employs two novel metrics: (1) a waveform similarity metric that uses a neural network to obtain spatially informed, time-shift invariant low-dimensional waveform representations, and (2) a cross-correlogram significance metric based on the earth-movers distance between the observed and null cross-correlograms. On a diverse set of datasets with realistic simulated oversplitting, SLAy achieves high recall and near-perfect precision in identifying ground truth merges. We also demonstrate that SLAy achieves [~] 85% with human curators across a diverse set of animal models, brain regions, and probe geometries. To illustrate the impact of spike sorting errors on downstream analyses, we develop a new burst-detection algorithm and show that SLAy fixes spike sorting errors that preclude the accurate detection of bursts in neural data. SLAy leverages GPU parallelization and multithreading for computational efficiency, and is compatible with Phy and NeuroData Without Borders, making it a practical and flexible solution for large-scale ephys data analysis.