Neural spike sorting using iterative ICA and a deflation-based approach

Zoran Tiganj, Mamadou MboupJournal of Neural Engineering 2012 · DOI

TL;DR

This paper introduces a novel iterative spike sorting algorithm that leverages Independent Component Analysis (ICA) and a deflation technique to improve the separation of neural signals from multi-channel recordings. By iteratively isolating and removing individual neuron signals, the method effectively addresses the challenge of having fewer recording sites than neurons, leading to significantly enhanced spike sorting accuracy.

What Problem Does This Paper Address?

Accurately identifying and separating the electrical signals (spikes) from individual neurons is a fundamental challenge in neuroscience, especially when using multi-channel extracellular recordings. These recordings capture a mixture of signals from many neurons, and the number of recording channels is often far fewer than the number of neurons present. This "underdetermined source separation" problem limits the effectiveness of standard techniques like Independent Component Analysis (ICA), which typically perform better when the number of sources (neurons) is less than or equal to the number of observations (recording channels). Existing methods struggle to disentangle the signals of closely located neurons or neurons that fire simultaneously, leading to misclassifications or missed spikes.

Methods

The proposed method employs a two-nested-loop iterative approach combined with a deflation technique.

  • Iterative ICA and Deflation: The core idea is to repeatedly apply ICA to the multi-channel recordings, isolate the signal of one neuron, and then "deflate" (remove) that signal from the recordings. This process is repeated until no more distinct neuron signals can be reliably separated or a predefined number of neurons is identified.
  • Nested Loops:
    • The external loop focuses on identifying and isolating one neuron's activity at a time.
  • The internal loop refines this process by applying ICA and a sorting method to remove noise and spikes not belonging to the currently targeted neuron.
  • Handling Overlapped Spikes: A crucial final step addresses the problem of simultaneous firing. For pairs of identified neurons, ICA is applied to signal segments where at least one of the neurons in the pair was active, allowing for the separation of these overlapping spikes.
  • Spike Sorting Algorithm: A simple greedy method, based on the non-uniform distribution of spike features, is used for sorting. This method identifies clusters by finding the nearest neighbors in a feature space.

Datasets

  • Simulated Recordings: The authors generated artificial neural mixtures from a population of 1000 neurons recorded by a four-channel electrode. This dataset allowed for precise knowledge of which neuron fired which spike, enabling accurate validation.
  • Real Recordings: Simultaneous extracellular-intracellular recordings from a rat hippocampal area CA1 were used. The intracellular recording provided a ground truth for one neuron's activity, allowing validation of the extracellular recording's spike sorting performance for that specific neuron.

Key Insights and Ideas

  • Deflation as a Key Enabler: The iterative deflation of identified neuron signals is the critical innovation that allows ICA to work effectively in an underdetermined scenario (more neurons than channels).
  • Nested Iteration for Refinement: The nested loop structure allows for progressive refinement, first isolating a neuron and then cleaning up its signal and the remaining mixture.
  • Targeted ICA for Overlaps: Applying ICA specifically to segments where pairs of neurons are active is an elegant solution to the challenging problem of simultaneous firing.
  • Spike Dynamics as a Feature: Using spike dynamics (average peak-to-peak amplitude difference) as a criterion for selecting the most informative ICA component and for identifying spikes is an effective heuristic.

Key Results

  • The proposed algorithm significantly improved spike sorting accuracy (SA) and detection (SD) compared to basic spike sorting and ICA alone, as demonstrated in both simulated and real recordings.
  • Each step of the proposed algorithm contributed to improved results, with the final step of detecting overlapped spikes providing a notable boost, especially for more distant neurons.
  • While the algorithm generally increased the number of correctly sorted spikes, it also slightly increased the number of missed spikes, a trade-off considered acceptable in spike sorting.
  • The algorithm showed good robustness, with standard deviations generally decreasing as more sophisticated steps were incorporated.

Limitations and Future Work

  • Simultaneous Firing Limitation: The deflation process inherently struggles to detect simultaneous firing from multiple neurons, typically only identifying one spike per simultaneous event. The proposed final step partially addresses this but might not be exhaustive.
  • Time Delays in Recordings: The current FastICA implementation is not suitable for recordings with significant time delays in spike arrival across channels (e.g., peripheral nerve interfaces). The authors suggest that replacing FastICA with a convolutive version of ICA could address this limitation.
  • Feature Extraction: While the paper used simple peak amplitude features, they note that other feature extraction methods could potentially lead to even better performance.
  • Noise Sensitivity: The authors acknowledge that in very noisy recordings, the algorithm might fail to detect any clusters.

Summary updated

Abstract

We propose a spike sorting method for multi-channel recordings. When applied in neural recordings, the performance of the independent component analysis (ICA) algorithm is known to be limited, since the number of recording sites is much lower than the number of neurons. The proposed method uses an iterative application of ICA and a deflation technique in two nested loops. In each iteration of the external loop, the spiking activity of one neuron is singled out and then deflated from the recordings. The internal loop implements a sequence of ICA and sorting for removing the noise and all the spikes that are not fired by the targeted neuron. Then a final step is appended to the two nested loops in order to separate simultaneously fired spikes. We solve this problem by taking all possible pairs of the sorted neurons and apply ICA only on the segments of the signal during which at least one of the neurons in a given pair was active. We validate the performance of the proposed method on simulated recordings, but also on a specific type of real recordings: simultaneous extracellular-intracellular. We quantify the sorting results on the extracellular recordings for the spikes that come from the neurons recorded intracellularly. The results suggest that the proposed solution significantly improves the performance of ICA in spike sorting.
BibTeX
@article{tiganj2012neural,
  author = {Zoran Tiganj and Mamadou Mboup},
  title = {{Neural spike sorting using iterative ICA and a deflation-based approach}},
  journal = {Journal of Neural Engineering},
  year = {2012},
  doi = {10.1088/1741-2560/9/6/066002},
}
Figure