Quantitative measures of cluster quality for use in extracellular recordings

Neil Schmitzer-Torbert, Jadin C. Jackson, Darrell A. Henze, Kenneth D. Harris, A. David RedishNeuroscience 2005 · DOI

TL;DR

This paper introduces two novel quantitative measures, Lratio and Isolation Distance, to objectively assess the quality of single neuron recordings from multi-channel electrodes. These measures are crucial for reliably distinguishing individual neuron signals from noise and other neurons, improving the reproducibility and rigor of neuroscience research.

What Problem Does This Paper Address?

Neuroscience research increasingly relies on extracellular recordings with multi-channel electrodes (like tetrodes) to simultaneously monitor large ensembles of neurons. A critical step in analyzing this data is "spike sorting," where individual action potentials (spikes) are assigned to specific neurons. While traditional single-electrode recordings used Signal-to-Noise Ratio (SNR) as a quality metric, this becomes insufficient with multi-channel data. Here, spikes from multiple neurons and noise events are mixed, and their waveforms are analyzed in a high-dimensional feature space. Existing methods for assessing the quality of these sorted clusters (i.e., how well a cluster represents a single neuron) are often subjective and unreliable, especially for clusters of intermediate quality. This paper aims to fill this gap by providing objective, quantitative measures for cluster quality.

Methods

The authors propose two new quantitative measures: Lratio and Isolation Distance. Both measures operate in a high-dimensional feature space derived from extracellular spike waveforms.

  1. Feature Extraction: For each spike recorded on a tetrode, they extract features. The primary features used are the energy of the waveform and the first principal component coefficient of the energy-normalized waveform, calculated for each of the four tetrode channels. This results in an 8-dimensional feature space for each spike.
  2. Mahalanobis Distance: Both measures utilize the Mahalanobis distance, which accounts for correlations between dimensions in the feature space. This is crucial because spike features across different channels are often correlated.
  3. Lratio: This measure quantifies the amount of "noise" (spikes not belonging to the cluster) found in the immediate vicinity of the cluster. It's calculated based on the cumulative distribution function of the chi-squared distribution, essentially measuring how many non-cluster spikes fall within a "moat" around the cluster. A low Lratio indicates a well-separated cluster.
  4. Isolation Distance: This measure defines the radius of the smallest hypersphere that encloses all the cluster's spikes and separates them from the nearest non-cluster spikes. It's determined by the Mahalanobis distance to the Nth nearest noise spike, where N is the number of spikes in the cluster. A high Isolation Distance indicates good separation.

Datasets

The authors evaluated their measures using:

  • Simulated Data: They generated synthetic spike data with known cluster and noise distributions (Gaussian and non-Gaussian) to test the measures under controlled conditions.
  • Actual Extracellular Recordings:
    • Data from rodent hippocampus and striatum using tetrodes.
  • Paired intracellular and extracellular recordings from rodent hippocampus. This dataset is particularly valuable as it allows for direct comparison of extracellularly sorted clusters with intracellularly identified single neurons, enabling the evaluation of Type I (false positive) and Type II (false negative) errors.

Key Insights and Ideas

  • Objective Quality Assessment is Crucial: Subjective assessment of spike sorting quality is a bottleneck for reproducible neuroscience. Quantitative measures are essential.
  • Mahalanobis Distance is Key: Accounting for correlations between spike waveform features across channels is vital for accurate distance calculations in high-dimensional spaces.
  • Lratio and Isolation Distance Complement Each Other: Lratio focuses on the "noise moat" around a cluster, while Isolation Distance measures how far the cluster is from its nearest neighbors. Together, they provide a robust assessment.
  • Feature Space Choice Matters (but not too much): While the specific values of Lratio and Isolation Distance depend on the chosen feature space (e.g., energy, principal components), the measures consistently differentiate well-separated from poorly-separated clusters across various feature combinations.
  • Robustness to Non-Gaussian Distributions: The measures perform well even when spike clusters deviate from ideal Gaussian shapes, which is common in real neural data.

Key Results

  • Simulation Success: Both Lratio and Isolation Distance successfully discriminated between well-separated and poorly-separated clusters in simulations, even with increasing dimensionality of the feature space.
  • Real Data Validation: The measures aligned with subjective assessments of cluster quality from actual tetrode recordings in the hippocampus and striatum.
  • Correlation with Error Rates: In paired intracellular/extracellular data, both Lratio and Isolation Distance showed significant negative correlations with Type I (false positive) and Type II (false negative) spike classification errors.
  • Lratio was more strongly correlated with Type II errors (missed spikes).
  • Isolation Distance was more strongly correlated with Type I errors (spikes incorrectly included).
  • Feature Space Recommendation: Using energy and the first principal component coefficient of the energy-normalized waveform for an 8-dimensional feature space provided strong correlations with error rates and is recommended for consistency across studies.

Limitations and Future Work

  • Threshold Selection: The paper acknowledges that these measures provide quantitative scores but do not inherently define a universal threshold for what constitutes a "good" cluster. The appropriate threshold will depend on the specific experimental question.
  • Feature Space Comparability: While the measures are robust to different feature choices, direct comparison of cluster quality values across studies is only justified if the same feature space (dimensionality and features) is used. The authors recommend a standardized feature space (energy + first principal component coefficient) for reporting.
  • Future Work: The authors suggest that reporting these quantitative measures would greatly improve the reproducibility and evaluation of neuroscience research. They also imply that further exploration of optimal feature spaces and thresholding strategies could be beneficial.

Summary updated

BibTeX
@inproceedings{schmitzertorbert2005quantitative,
  author = {Neil Schmitzer-Torbert and Jadin C. Jackson and Darrell A. Henze and Kenneth D. Harris and A. David Redish},
  title = {{Quantitative measures of cluster quality for use in extracellular recordings}},
  booktitle = {Neuroscience},
  year = {2005},
  doi = {10.1016/j.neuroscience.2004.09.066},
}
Figure