Automating synaptic plasticity analysis: A deep learning approach to segmenting hippocampal field potential signal

Abstract

Hippocampal field potentials are widely used in research on neurodegenerative diseases, epilepsy, neuropharmacology, and particularly long- and short-term synaptic plasticity. To conduct these studies, it is necessary to identify specific components within hippocampal field potential signals. However, manually marking the relevant signal points for analysis is a time-consuming, error-prone, and subjective process. Currently, there is no specialized software dedicated to automating this task. In this study, three different recurrent neural network-based deep learning architectures were examined for the automatic segmentation of hippocampal field potential signals in two separate experimental studies. In the first experimental study, 10,836 epochs of field potential signals recorded from 54 rats were used, and in the second experimental study, field potential signals with noise added to the above data at different rates were used. The best model achieved an average f-score of 98.1% on noise-free data and 97.15% on data with noise, highlighting its robustness in real-world scenarios. Furthermore, we assessed system stability using the repeated holdout method, which randomly split the data into training and testing sets 100 times, and each time trained a new version of the system. As a result, the proposed system was proven to be reliable and generalizable by showing similar average scores and low variability across all 100 iterations of the test.