Classification of Sensory Neural Signals through Deep Learning Methods

Abstract

The recording and analysis of peripheral neural signals can be beneficial to provide feedback to prosthetic limbs and recover the sensory functionality in people with nerve injuries. Nevertheless, the interpretation of sensory recordings extracted from the nerve is not trivial, and only few studies have applied classifiers on sequences of neural signals without previous feature extraction. This paper evaluates the classification performance of two deep learning (DL) models (CNN and ConvLSTM) applied to the electroneurographic (ENG) activity recorded from the sciatic nerve of rats. The ENG signals, available from two public datasets, were recorded using multi-channel cuff electrodes in response to four sensory inputs (plantarflexion, dorsiflexion, nociception, and touch) elicited in response to mechanical stimulation applied to the hind paw of the rats. Different temporal lengths of the signals were considered (2.5 s, 1 s, 500 ms, 200 ms, and 100 ms). Both the two DL models proved to correctly discriminate sensory stimuli without the need of hand-engineering feature extraction. Moreover, ConvLSTM outperformed state-of-the-art results in classifying sensory ENG activity (more than 90% F1-score for sequences greater than 500 ms), and it showed promising results for real-time application scenarios.