Classification of naturally evoked compound action potentials in peripheral nerve recordings via convolutional neural networks

Ryan Koh (1), Adrian Nachman (1), Jose Zariffa (1)

(1) Institute of Biomaterials and Biomedical Engineering, University of Toronto

Objective: Recording and stimulation from the peripheral nervous system are becoming important components in a new generation of bioelectronics systems. Neurostimulation in humans using implanted peripheral neural interfaces has seen a long history of success, including in applications such as reducing phantom pain in amputees, treatment for overactive bladder, and implanted functional electrical stimulation for movement. Unfortunately, recording applications using implanted peripheral neural interfaces has not been as successful and remains a challenge. Improvements to recording devices and signal processing techniques to extract useful information from those devices are needed. With the objective of recording selectively and reliably from different neural pathways in a peripheral nerve, we propose to use a convolutional neural network to exploit the spatiotemporal structure of compound action potentials (CAPs) recorded from a 56-channel nerve cuff.

Approach: 9 Long-Evan rats were implanted with a 56-channel spiral nerve cuff electrode on the sciatic nerve. Afferent activity was selectively evoked in three fascicles of the sciatic nerve (tibial, peroneal, sural) using mechanical stimuli. Spatiotemporal signatures of recorded CAPs were used to train the CNN. A recurrent neural network was then trained to predict the joint angle based on predicted firing patterns from the CNN. Performance was measured based on the CNN’s classification accuracy and F1-score, and correlation between the ground truth and predicted joint angle of the rat’s ankle.
Main Results: Our novel technique using CNNs yielded a mean classification accuracy of 0.808 ±0.104 with corresponding mean F1-score of 0.747 ±0.114. In contrast, the mean classification accuracy and F1-score for the previous state-of-the-art were 0.686 ±0.126 and 0.605 ±0.212, respectively. Using the CNN classification results, the mean Pearson correlation coefficient was 0.826 ± 0.176 for the ankle angle predicted using the estimated firing rate vs the manually labelled ankle angle.
Significance: The proposed method demonstrates that CAP-based classification can be used to track a physiological meaningful measure (e.g. joint angle) and will allow for more precise control signals in neuroprosthetic systems.