This package contains matlab code for sorting/estimating spikes of neurons recorded with one or more extracellular electrodes. Unlike commonly used clustering methods, this method can recover temporally overlapping spikes through use of a sparse inverse algorithm known as Continuous Basis Pursuit (CBP). The method is described in this publication:

A unified framework and method for automatic neural spike identification. C Ekanadham, D Tranchina, and E P Simoncelli. J. Neuroscience Methods, vol. 222, pp. 47--55, Jan 2014. DOI: 10.1016/j.jneumeth.2013.10.001 http://www.cns.nyu.edu/~lcv/pubs/makeAbs.php?loc=Ekanadham13

For optimization, the code uses the Embedded Conic Solver (ECOS) package, an interior-point algorithm for second-order cone programming, written by A. Domahidi, E. Chu, and S. Boyd.
See http://web.stanford.edu/~boyd/papers/ecos.html For convenience, this package is included in the ecos subDirectory.

We also include two example datasets with ground truth (correct spike times), as obtained from these sites: *** INCLUDE URLS here ***

TO GET STARTED: we suggest you work through the code in the demonstration script, spikesort_demo/cbp_spikesort_demo_script.m executing one section at a time.

We also include a ChangeLog file documenting changes made to the code, as well our current ToDo file.

• Conceptualization and design by Chaitanya Ekanadham and Eero Simoncelli.
• Original matlab code written by Chaitanya Ekanadham, 12/26/2012.
• Rewritten to use ECOS, and interface updated by Peter H. Li, Fall/Winter 2013.
• Changes/adjustments made by Eero, Spring/Fall 2014.
• Further changes made by Mike Battaglia in conjunction with Eero, Summer 2016.
• Current version available at http://www.cns.nyu.edu/~lcv/software.php

===================================================================================== OUTLINE OF METHOD:

(0) Load raw data, stored in an array containing electrode voltages (each row is a separate electrode). Sampling rate should be at least 5kHz.

PRE-PROCESSING:

(1) Filter. Purpose is to eliminate low and high frequency noise (increasing the signal-to-noise ratio of the data) and to allow isolation of noise regions (for whitening, step 2) by thresholding.

(2) Whiten noise. The CBP objective function is most easily and efficiently computed when the background noise is uncorrelated (white), over both time and electrodes. Covariance over time and electrodes is computed on low-amplitude portions of data. The entire data array is then (separably) whitened by linearly transforming across electrodes, and filtering over time.

(3) Select number of cells, and initialize spike waveforms. Initial waveforms are obtained from the centroids of k-means clustering in a principal components space. Note: this NOT used to identify/estimate spikes - it is only used to determine the number of cells, and to obtain initial estimates of their waveforms.

CBP SPIKE SORTING:

(4) Partition data (optional). To improve efficiency (especially when using multiple cores or machines), the data array is separated into smaller segments, separated by spike-free intervals.

(5) Use CBP (see article listed above) to estimate amplitudes/times of spikes associated with each waveform.

(6) Decide on which spikes to keep/discard, by thresholding the recovered amplitudes.

(7) Re-estimate waveforms (given the spike times, this is a simple regression problem). If changes are substantial, repeat from step (5) until convergence.

POST-PROCESSING:

(8) Compare CBP results to Clustering, and ground truth (true spikes, if available).

=====================================================================================