Selective peripheral nerve recordings from nerve cuff electrodes using convolutional neural networks

Abstract

Objective. Recording and stimulating from the peripheral nervous system are becoming important components in a new generation of bioelectronics systems. Although neurostimulation has seen a history of successful chronic applications in humans, peripheral nerve recording in humans chronically remains a challenge. Multi-contact nerve cuff electrode configurations have the potential to improve recording selectivity. We introduce the idea of using a convolutional neural network (CNN) to associate recordings of individual naturally evoked compound action potentials (CAPs) with neural pathways of interest, by exploiting the spatiotemporal patterns in multi-contact nerve cuff recordings. Approach. Nine Long-Evan rats were implanted with a 56-channel nerve cuff electrode on the sciatic nerve and afferent activity was selectively evoked in different fascicles (tibial, peroneal, sural) using mechanical stimuli. A recurrent neural network was then used to predict joint angles based on the predicted firing patterns from the CNN. Performance was measured based on the classification accuracy, F1-score and the ability to track the ankle joint angle. Main results. Classification accuracy and F1-score of the best CNN configuration were and 0.747 ± 0.114, respectively. The mean Pearson correlation coefficient between the manually measured ankle angle and the angle predicted from the estimated firing rate was Significance. The proposed method demonstrates that CAP-based classification can be achieved with high accuracy and can be used to track a physiological meaningful measure (e.g. joint angle). These results provide a promising direction for realizing more effective and intuitive neuroprosthetic systems.

Compensation Strategies for Bioelectric Signal Changes in Chronic Selective Nerve Cuff Recordings: A Simulation Study

“…The spatiotemporal signature used as input to ESCAPE-NET can exhibit different patterns depending on the ordering of the contacts. Alternatively, the network can be provided with more than one such representation, in order to exploit possibly complimentary information [42]. In the case of this project, ESCAPE-NET’s dual input version was used.…”

“…Convolutional Neural Network A previously described CNN, ESCAPE-NET [42], was used to classify the spatiotemporal signatures produced by the CAPs in the multi-contact nerve cuff. As such, the inputs to ESCAPE-NET were 56x100 spatiotemporal signatures created as described above.…”

“…Alternatively, a number of studies have attempted to achieve spatial selectivity of neural recordings using beamforming [15], [35]- [37] or source localization methods [38], [39]. Building on the results above, Koh et al used spatiotemporal signatures to characterize the recordings from a nerve cuff with an array of contacts [40]- [42]. By using these spatiotemporal signatures as input to a convolutional neural network (CNN), individual natural compound action potentials (nCAPs) could be classified according to the neural pathway that produced them [42].…”

“…Building on the results above, Koh et al used spatiotemporal signatures to characterize the recordings from a nerve cuff with an array of contacts [40]- [42]. By using these spatiotemporal signatures as input to a convolutional neural network (CNN), individual natural compound action potentials (nCAPs) could be classified according to the neural pathway that produced them [42]. This CNN is referred to as the extraneural spatiotemporal compound action potential extraction network, or ESCAPE-NET.…”

“…ESCAPE-NET architecture with Spatial and Temporal Emphasis representations of the multi-contact nerve cuff recordings (SNR = -5 dB) shown as input to the network[42]. Visual representation of the processes involved in self-learning approach.…”