Review Paper
Recent progress in multi-electrode spike sorting methods

https://doi.org/10.1016/j.jphysparis.2017.02.005Get rights and content

Highlights

  • Large and dense arrays of electrodes require new spike sorting methods.

  • Novel methods have been designed that rely on template matching.

  • Most of them follow the same general strategy.

  • We review their common points and differences.

Abstract

In recent years, arrays of extracellular electrodes have been developed and manufactured to record simultaneously from hundreds of electrodes packed with a high density. These recordings should allow neuroscientists to reconstruct the individual activity of the neurons spiking in the vicinity of these electrodes, with the help of signal processing algorithms. Algorithms need to solve a source separation problem, also known as spike sorting. However, these new devices challenge the classical way to do spike sorting. Here we review different methods that have been developed to sort spikes from these large-scale recordings. We describe the common properties of these algorithms, as well as their main differences. Finally, we outline the issues that remain to be solved by future spike sorting algorithms.

Introduction

Progress in neuroscience relies to a large extent on the ability to record simultaneously from large populations of cells, in order to understand how information is represented among neurons. One of the most popular techniques to measure such an activity is the use of arrays of extracellular electrodes. With these devices, each electrode records the extracellular field in its vicinity and can detect the action potentials emitted by the neighboring neurons. In contrast to intracellular recording, those extracellular recordings do not give a direct access to the neuronal activity: one needs to process the recorded signals to extract the spikes emitted by the different cells around the electrode. This process is termed spike sorting, and many algorithms have been suggested to do it efficiently (see Lewicki (1998) or Rey et al. (2015) for a review).

The first extracellular recordings were performed with a single electrode, and could only give access to 3-5 neurons (Gerstein and Clark, 1964). A recent study (Pedreira et al., 2012) highlighted that the maximal number of accessible neurons should lie between 8 and 10 in that case. Over the last decades, there has been a strong effort to increase the number of electrodes, and therefore the number of recorded neurons. Spike sorting algorithms had to be adapted to process this increasingly large amount of data. At first, electrodes were spaced by hundreds of microns such that the spike of one cell could only be detected on a single electrode (Jones et al., 1992, Shoham et al., 2003). In that case, spike sorting on a large amount of electrodes could simply be done by processing each electrode independently. The parallelization of the problem for large amount of independent electrodes was relatively easy to address.

However, devices where electrodes are packed with a high density have also been developed. The spacing between electrodes is much smaller (tens of microns). As a consequence, a spike from a single cell can be detected on several electrodes. Conversely, each electrode will detect the activity of many cells, a property already encountered in the case of single electrode. This increased density helps a lot to resolve single cells (Gray et al., 1995, Franke et al., 2015a), but electrode signals could not be processed independently. Spike sorting algorithms had to be adapted to this new type of data. While for small numbers of electrodes (e.g. tetrodes), methods that could be seen as adaptations of single electrode sorting worked very well (McNaughton et al., 1983, Harris et al., 2000, Gao et al., 2012), this is not the case with new devices designed with hundreds of electrodes all densely packed. CMOS-based devices with thousands of electrodes have been tested and are now frequently used (Berdondini et al., 2005, Fiscella et al., 2012, Müller et al., 2015, Hilgen et al., 2016), calling for new algorithmic methods, largely different from the usual sorting methods.

Here we review the different spike sorting algorithms that have been proposed to process recordings from these novel high-density devices. We will first explain the limitations of classical spike sorting approaches to process these large-scale, dense recordings. Then, we will outline the main changes introduced by these new algorithms compared to classical spike sorting approaches. We will emphasize that most of these new methods follow the same global strategy, although they have been developed independently by different groups. Therefore, we will outline the common properties shared by these algorithms, before explaining and discussing their main differences. Finally, we will discuss the issues that still need to be resolved by future spike sorting algorithms.

Section snippets

The challenge posed by large-scale multi-electrode recordings to classical approaches

Most of the classical approaches to spike sorting can be decomposed in two main steps. First, some specific features of the spike waveforms are extracted from the raw data. This allows each spike to be characterized by a small set of numbers/features. Using these features, each spike can now be seen as a point in a low dimension space, and the second step consists in clustering all the points in this reduced space.

For the first step, earliest methods only extracted the spike amplitude (Hubel,

Improvements of the clustering

In order to be able to scale up and perform spike sorting for large number of channels with the classical algorithms mentioned above, several refinements of the clustering have been proposed by various groups.

Template matching approaches

Several template matching approaches have been developed for spike sorting (Pillow et al., 2013, Pachitariu et al., 2016, Marre et al., 2012, Yger et al., 2016, Prentice et al., 2011). Note that historically the use of template matching (Gerstein and Clark, 1964) predates the use of clustering (Simon, 1965) and then experienced renewed interest. All these methods usually assume that the extracellular signal can be decomposed as a sum of so-called “templates” (one template is the average

Conclusion: challenges ahead

The methods described here have enabled to sort spikes from a large number of cells and electrodes (Yger et al., 2016, Pachitariu et al., 2016). However, there are still several challenges that need to be overcome. First, most of the algorithms described here have been tested on in vitro data, in the retina (but see Ekanadham et al., 2014, Franke et al., 2015b or Yger et al. (2016) for in vivo tests). In vivo tests on silicon probes with a large number of recording sites close apart will be

Acknowledgments

This work was supported by ANR OPTIMA and TRAJECTORY, the French State program Investissements d’Avenir managed by the Agence Nationale de la Recherche [LIFESENSES: ANR-10-LABX-65], a grant from the European Union Seventh Framework Programme (FP7/2007–2013, grant agreement No. 604102, Human Brain Project), and NIH grant U01NS09050 to OM.

References (57)

  • E. Hulata et al.

    A method for spike sorting and detection based on wavelet packets and shannon’s mutual information

    J. Neurosci. Methods

    (2002)
  • C. Leibig et al.

    Unsupervised neural spike sorting for high-density microelectrode arrays with convolutive independent component analysis

    J. Neurosci. Methods

    (2016)
  • J.C. Letelier et al.

    Spike sorting based on discrete wavelet transform coefficients

    J. Neurosci. Methods

    (2000)
  • B.L. McNaughton et al.

    The stereotrode: a new technique for simultaneous isolation of several single units in the central nervous system from multiple unit records

    J. Neurosci. Methods

    (1983)
  • M. Meister et al.

    Multi-neuronal signals from the retina: acquisition and analysis

    J. Neurosci. Methods

    (1994)
  • C. Pedreira et al.

    How many neurons can we see with current spike sorting algorithms?

    J. Neurosci. Methods

    (2012)
  • C. Pouzat et al.

    Using noise signature to optimize spike-sorting and to assess neuronal classification quality

    J. Neurosci. Methods

    (2002)
  • V. Prochazka et al.

    A neuroelectric signal recognition system

    Electroencephalogr. Clin. Neurophysiol.

    (1972)
  • H.G. Rey et al.

    Past, present and future of spike sorting techniques

    Brain Res. Bull.

    (2015)
  • W.M. Roberts et al.

    Separation of multi-unit nerve impulse trains by a multi-channel linear filter algorithm

    Brain Res.

    (1975)
  • S. Shoham et al.

    Robust, automatic spike sorting using mixtures of multivariate t-distributions

    J. Neurosci. Methods

    (2003)
  • W. Simon

    The real-time sorting of neuro-electric action potentials in multiple unit studies

    Electroencephalogr. Clin. Neurophysiol.

    (1965)
  • A.F. Atiya

    Recognition of multiunit neural signals

    IEEE Trans. Biomed. Eng.

    (1992)
  • E. Chah et al.

    Automated spike sorting algorithm based on laplacian eigenmaps and k-means clustering

    J. Neural Eng.

    (2011)
  • C. Ekanadham et al.

    Recovery of sparse translation-invariant signals with continuous basis pursuit

    IEEE Trans. Signal Process.

    (2011)
  • J. Fournier et al.

    Consensus-based sorting of neuronal spike waveforms

    PloS One

    (2016)
  • F. Franke et al.

    An online spike detection and spike classification algorithm capable of instantaneous resolution of overlapping spikes

    J. Comput. Neurosci.

    (2010)
  • F. Franke et al.

    Spike sorting of synchronous spikes from local neuron ensembles

    J. Neurophysiol.

    (2015)
  • Cited by (46)

    • Online spike sorting via deep contractive autoencoder

      2022, Neural Networks
      Citation Excerpt :

      Several approaches have been proposed for the clustering step of spike sorting including clustering with a mixture of Gaussians (Rossant et al., 2016; Wood, Fellows, Donoghue, & Black, 2004), K-means clustering (Chah, Hok, Della-Chiesa, Miller, O’Mara, & Reilly, 2011), an ensemble of K-means solutions (Fournier, Mueller, Shein-Idelson, Hemberger, & Laurent, 2016), paramagnetic clustering (Quiroga, Nadasdy, & Ben-Shaul, 2004), and mean-shift clustering (Swindale & Spacek, 2014). Nature of spike waveforms imposes different challenges to the clustering step of spike sorting; two of the main problems are: (1) lack of mathematical definition for clustering, and (2) ending up in a sub-optimal solution in most cases (Lefebvre, Yger, & Marre, 2016). Such issues are originated from modality of waveform distribution.

    • Classification of overlapping spikes using convolutional neural networks and long short term memory

      2022, Computers in Biology and Medicine
      Citation Excerpt :

      Recording the neuronal electrophysiological activities is helpful to understand the manner of information transmission between neurons, where extracellular recording is one of the most popular and effective acquisition techniques. It can pick up the action potentials, i.e. spikes, emitted by several neurons adjacent to the electrode site [2–5]. Therefore, it is necessary to classify the spikes based on the similarity of their waveforms followed by identifying which neuron fired at what time.

    • Building population models for large-scale neural recordings: Opportunities and pitfalls

      2021, Current Opinion in Neurobiology
      Citation Excerpt :

      From extracellular recordings, spike trains of individual neurons are extracted through a process called spike sorting. There are several algorithmic approaches for spike sorting of data from dense probes, which are reviewed elsewhere [67,29,12,40]. Accurately extracting spike trains is essential because mistakes, including incorrectly merged or split units, unresolved overlapping spikes, or incorrect clustering due to probe drift, have been shown to bias subsequent analysis [4,90,91].

    View all citing articles on Scopus
    #

    These authors contributed equally to this work.

    View full text