Abstract
Identification of neuron cell type helps us connect neural circuitry and behavior; greater specificity in cell type and subtype classification provides a clearer picture of specific relationships between the brain and behavior. With the advent of high-density probes, large-scale neuron classification is needed, as typical extracellular recordings are identity-blind to the neurons they record. Current methods for identification of neurons include optogenetic tagging and intracellular recordings, but are limited in that they are expensive, time-consuming, and have a limited scope. Therefore, a more automated, real-time method is needed for large-scale neuron identification. Data from two recordings was incorporated into this research; the single-channel recording included data from three neuron types in the motor cortex: FS, IT, and PT neurons. The multi-channel recording contained data from two neuron subtypes also in the motor cortex: PT_L and PT_U neurons. This allowed for an examination of both general neuron classification and more specific subtype classification, which was done via artificial neural networks (ANNs) and machine learning (ML) algorithms. For the single-channel neuron classification, the ANNs achieved 91% accuracy, while the ML algorithms achieved 98% accuracy, using the raw electrical waveform. The multi-channel classification, which was significantly more difficult due to the similarity between the neuron types, yielded an ineffective ANN, reaching 68% accuracy, while the ML algorithms reached 81% using 8 calculated features from the waveform. Thus, to distinguish between different neuron cell types and subtypes in the motor cortex, both ANNs and specific ML algorithms can facilitate rapid and accurate near real-time large-scale classification.
Footnotes
The author has declared that no competing interests exist.
