Deflation technique for neural spike sorting in multi-channel recordings

Zoran Tiganj, Mamadou Mboup

TL;DR

This paper introduces a novel "deflation" technique that iteratively refines Independent Component Analysis (ICA) for more accurate neural spike sorting in multi-channel recordings. By systematically isolating and removing the activity of individual neurons, the algorithm significantly improves the ability to distinguish between closely firing neurons, outperforming standard ICA and feature extraction methods.

What Problem Does This Paper Address?

Neuroscience research often relies on recording the electrical activity of neurons, specifically their "spikes" or action potentials. When recording from multiple sites in the brain, these spikes from different neurons can overlap and mix, making it challenging to isolate the activity of a single neuron. This is known as the "spike sorting problem." While techniques like Independent Component Analysis (ICA) can help separate mixed signals, their effectiveness is limited when the number of recording channels is much smaller than the number of neurons, a common scenario. This paper aims to overcome this limitation by proposing a more robust and accurate spike sorting method.

Methods

The core of the proposed algorithm is an iterative application of ICA combined with a "deflation" technique. This process occurs in two nested loops:

  • Outer Loop (Neuron Isolation): In each iteration, the algorithm attempts to isolate the activity of one neuron. It first applies ICA to the multi-channel recordings. Then, it identifies the Independent Component (IC) that most likely represents a single neuron's activity. This identified neuron's spikes are then "deflated" or removed from the original recordings. This process is repeated until no more distinct neurons can be reliably isolated.
  • Inner Loop (Noise and Non-Targeted Spike Removal): Within each outer loop iteration, an internal process refines the signal. This involves a sequence of ICA and spike detection to remove noise and any spikes that do not belong to the currently targeted neuron.

The algorithm also includes steps for:

  • Spike Detection: Identifying potential spike events in the recordings.
  • Feature Extraction: Representing detected spikes in a reduced-dimensional space (e.g., using PCA).
  • Clustering: Grouping similar spike features to identify distinct neuronal clusters. The paper mentions using a clustering algorithm from a previous work [15] that is specifically designed for neural recordings and relies on the non-uniform distribution of spike features.
  • Overlapped Spike Detection: A specific step is dedicated to identifying and separating spikes that fire simultaneously from different neurons. This is achieved by analyzing pairs of identified neurons and applying ICA to segments where their activity is isolated.

Datasets

The authors evaluated their algorithm on two types of datasets:

  1. Simulated Neural Mixtures: 10 simulated extracellular recordings, each 2 million samples long, generated from populations of 1000 neurons distributed around the electrodes.
  2. Real Extracellular-Intracellular Recordings: Five real simultaneous extracellular-intracellular recordings. The intracellular recordings were used to verify the sorting results for neurons that were simultaneously recorded.

Key Insights and Ideas

  • Deflation as a Powerful Refinement for ICA: The core insight is that iteratively removing the activity of identified neurons from the recording significantly "cleans up" the signal for subsequent ICA runs. This effectively reduces the "sources" problem that limits standard ICA in multi-channel recordings.
  • Nested Iterative Approach: The combination of an outer loop for neuron isolation and an inner loop for signal refinement provides a multi-stage approach to tackle the complexity of spike sorting.
  • Robustness through Iterative Refinement: By systematically isolating and removing neuronal activity, the algorithm becomes more robust to noise and the inherent difficulties of separating overlapping spikes.
  • Addressing the "Sources vs. Sensors" Problem: The paper directly tackles the limitation where the number of neurons (sources) exceeds the number of recording channels (sensors), a common bottleneck for ICA.

Key Results

The paper presents quantitative results in terms of Sorting Accuracy (SA) and Sorting Mistake (SM) for both real and simulated recordings.

  • Real Recordings (Table 1): The proposed algorithm, particularly with the "noise removal" and "cluster removal" steps, significantly improved SA and reduced SM compared to basic feature extraction or ICA alone. For example, one recording showed an improvement from 79% SA with basic ICA to 93% SA with the full proposed algorithm.
  • Simulated Recordings (Table 2): The full algorithm (including noise removal, cluster removal, sorted removal, and overlap detection) consistently achieved the highest SA and lowest SM across multiple sorted clusters. The "overlap detection" step, in particular, showed a substantial improvement in SA (e.g., from 76% to 97% for the first cluster).
  • Qualitative Findings: Visualizations (Figures 3, 4, 5, and 6) demonstrate how the algorithm progressively isolates neuronal activity and separates overlapping spikes, showing clearer cluster separation in feature spaces as the algorithm progresses.

Limitations and Future Work

  • Error Propagation: The authors acknowledge that a mistake made in an earlier iteration can influence subsequent iterations. However, they note that since only one neuron is targeted at a time, an error in one iteration does not necessarily affect the sorting results of all neurons.
  • No Explicit Error Compensation: The proposed algorithm does not have a dedicated error compensation step.
  • Future Work: The paper suggests that investigating different choices of the contrast function for ICA and exploring other versions of ICA could be beneficial. The authors also mention the potential for further research into improving the error compensation mechanisms.

Summary updated

Abstract

We propose an ICA based algorithm for spike sorting in multi-channel neural recordings. In such context, the performance of ICA is known to be limited since the number of recording sites is much lower than the number of the neurons around. The algorithm 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 spike detection for removing the noise and all the spikes that are not coming from the targeted neuron. We validate the performance of the algorithm on simulated data, but also on real simultaneous extracellular-intracellular recordings. The results show that the proposed algorithm performs significantly better than when only ICA is applied.
BibTeX
@inproceedings{tiganj2011deflation,
  author = {Zoran Tiganj and Mamadou Mboup},
  title = {{Deflation technique for neural spike sorting in multi-channel recordings}},
  booktitle = {MLSP},
  year = {2011},
  url = {https://openreview.net/forum?id=sZWKS_gakR},
}
Figure