Consensus-Based Sorting of Neuronal Spike Waveforms.

Area of Science:

• Computational Neuroscience
• Signal Processing
• Machine Learning

Background:

• Spike-sorting algorithm optimization is challenging due to the lack of ground truth data for validation.
• Current methods rely on assumptions about spike shape distributions and subjective visual inspection.
• Overlapping spike waveforms from different neurons introduce ambiguity in spike assignment.

Purpose of the Study:

• To develop a new spike-sorting approach that identifies consistent spike clusters from an ensemble of clustering solutions.
• To provide a method for estimating the proportion of misclassified spikes per cluster.
• To overcome the limitations of model-based assumptions and subjective validation in spike sorting.

Main Methods:

• Utilized a consensus partition across multiple clustering solutions to identify spike clusters.
• Employed variability in clustering solutions from iterative template matching (K-means based) to estimate spike clustering probabilities.
• Identified statistically indistinguishable groups of spikes, indicating consistent clusters.

Main Results:

• The proposed method identifies consistent spike clusters without relying on spike shape models.
• It provides accurate estimates of misclassification rates for each cluster.
• Performance on ground truth datasets approached the optimum achieved by a support vector machine.

Conclusions:

• This consensus-based spike-sorting approach offers a robust alternative to traditional methods.
• The method enhances the reliability of spike cluster identification and misclassification estimation.
• It demonstrates high performance, comparable to supervised machine learning methods, even without ground truth during training.