Strategies for Efficient ENG Signal Classification Using Data Augmentation and Data Balancing Techniques

Abstract

Peripheral nerve injuries interrupt essential brain-body communications, leading to significant functional impairments. This study evaluates five Convolutional Neural Network (CNN)-based classification methods for Sciatic Nerve Electroneurographic (ENG) signals in rats, aiming to restore the lost connection. The examined networks include classic CNN, Convolutional Tiny Transformer (CTT), InceptionTime (IT), IT with Derivative (IT-D), and ENG Network (ENGNet). Light preprocessing ensures real-time application, essential for human sensory systems, with a maximum delay of 300ms. Different approaches of data augmentation and data balancing strategies are used to address dataset imbalances. These approaches prove to be optimal for rebalancing, particularly the overlapping techniques, which significantly enhance the classification of the previously underestimated classes.