Discovering optimal features for neuron-type identification from extracellular recordings.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Multichannel recordings of single-unit activity (SUA) in vivo offer new insights into neuron-type identification.
• Traditional methods for classifying neuron-types using extracellularly-recorded action potentials (EAPs) have limitations inherited from single-channel recording conventions.
• Spatiotemporal EAP waveforms result from the mixing of underlying current sources in the extracellular space.

Purpose of the Study:

• To introduce a novel machine learning approach for demixing underlying sources of spatiotemporal EAP waveforms.
• To utilize these demixed sources as features for identifying neuron-types from SUA recordings.
• To establish a hierarchical classification scheme of latent morpho-electrophysiological types.

Main Methods:

• Simulated EAP waveforms using biophysically realistic computational models.
• Characterized simulated EAPs by the relative prevalence of demixed underlying sources.
• Employed classification and clustering methods on simulated EAPs from detailed morphological models for validation.

Main Results:

• Identified distinct spatial and multi-resolution temporal patterns for EAP sources, robust to sampling biases.
• Demonstrated that EAP sources are shared across neuron-types, predictive of morphological features, and reveal underlying morphological domains.
• Organized known neuron-types into a hierarchy based on source prevalences, creating a multi-level classification scheme.

Conclusions:

• The proposed machine learning approach effectively demixes EAP sources for neuron-type identification.
• This simulation-based strategy provides a robust and accurate method for classifying neuron-types.
• The identified EAP sources offer insights into the morpho-electrophysiological properties of neurons.