Heuristic Spike Sorting Tuner for the Determination of an Optimal Parameter Set for a Generic Spike Sorting Algorithm

Extracellular microelectrodes frequently record neural activity from multiple sources in the vicinity of the electrode. Spike sorting generally describes the process of labeling each recorded spike waveform with the identity of its source neuron, which is required to conduct any further analysis of the neuronal spiking patterns. This process for spike sorting or isolating neural activity is often approached from an abstracted mathematical perspective such as calculating the Euclidean distance between spike waveform features in some lower dimensional space or using probability distributions to describe the isolation of neural activity or recorded spikes. However, these approaches ignore neurophysiological realities and result in the loss of important information that could improve the accuracy of these methods. Furthermore, standard algorithms typically require manual selection of at least one free parameter, which can have significant effects on the ultimate quality of the spike sorting and all resulting neurophysiological inferences.\ud \ud We describe a Heuristic Spike Sorting Tuner (HSST) which determines the optimal choice of the set of free parameters for a given spike sorting algorithm. A set of heuristic metrics computes a neurophysiologically-based qualification score of an algorithm’s output across a range of parameters. This qualification score measures unit isolation and signal discrimination, allowing HSST to select the best set of parameters for a sorting algorithm, resulting in high sort quality. \ud \ud We demonstrate the power of this spike sorting framework, by comparing its performance across many existing spike sorting methods while using HSST to set their free parameters. The algorithm is robust over varied data (signal-to-noise ratio, number of units, relative size of units to each other, etc), and importantly; this approach requires no human supervision. \ud \ud With simulated datasets, HSST reliably selects the optimal set of free parameters for many different sorting algorithms, allowing simple clustering techniques (such as K-Means) whose performance is highly dependent on correct parameter settings to outperform more complex algorithms. HSST outperforms expert manual sorters and is more robust at parameter estimation than other unsupervised algorithms, achieving this without sacrificing speed or stability. \ud \ud Rather than being a spike sorting algorithm in its own right, HSST is a general framework that can incorporate any existing spike sorting algorithm, has an extendable set of heuristics and can be integrated in any existing neural signal processing stream. HSST makes use of known neurophysiological priors while simultaneously taking advantage of the power of abstract mathematical tools. We believe that this approach enables unsupervised spike sorting that exceeds the performance of previous methods, thereby enabling principled processing of large data sets without the significant confound of human intervention.\u