Classification of Theoretical Extracellular Action Potentials Based on Unsupervised Machine-Learning

Abstract

Extracellular action potentials (EAP) are one of the most important features in biological study. Many researchers have studied the classification of EAP by their differences in voltage and magnitude. However, most research ignored the fundamental origin of the EAP variation around the neurons in their classification and treated waveforms of different shapes as signals recorded from different neurons. In our research, we theoretically investigated the shapes of EAP by clustering the spatially-varied EAP around the neuron. We use an unsupervised machine-learning algorithm to classify all EAPs measured around the same neuron. To eliminate the influence of the non-characteristic part of the EAP curve, we also compared the classification results by eliminating the unchanged part at the front and end of the curve in the second group of our study. Our results illustrate the previously overlooked relationship between different shaped EAP and the biological structure of the neuron. The results show that EAP measured is closer to classical theory prediction in the axon while more eccentric, even with a shape similar to an intracellular action potential in the dendrite. Our research has important implications for further device design to record accurate electric signals and extracting biological related information from extracellular recordings.