Classification of naturally evoked compound action potentials in peripheral nerve recordings using spatiotemporal signatures
Ryan G. L. Koh 1, 2 ; Adrian I. Nachman 1, 3, 4 ; José Zariffa 1, 2, 3
1. Institute of Biomaterials and Biomedical Engineering, University of Toronto; 2. Toronto Rehabilitation Institute – University Health Network, Toronto; 3. Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto; 4. Department of Mathematics, University of Toronto
Objective: Neural signals from peripheral nerves could potentially provide motor, sensory or autonomic information for robust control in many neuroprosthetic and neuromodulation applications. However, developing algorithms that can correctly recover the information encoded in these signals is a significant challenge. Current techniques do not fully exploit the distinct spatiotemporal signatures that compound action potentials (CAPs) produce as they propagate through the nerve. We integrated spatial and temporal approaches and introduce the idea of associating individual CAPs with neural sources of interest using spatiotemporal signatures extracted from a multi-contact nerve cuff electrode.
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 three different classifiers (matched filter, random forest, and neural net). Performance was measured based on the classification accuracy, F1-score, and the ability to reconstruct the original firing rates of neural pathways.
Main Results: The mean classification accuracy, for a 3-class problem, for the 3 classifiers were 0.510 ±0.108, 0.658 ±0.115, 0.686 ±0.126 and corresponding mean F1-score were 0.446 ±0.157, 0.578 ±0.210, 0.605 ±0.212, for the matched filter, random forest and neural net respectively. The mean Pearson correlation coefficient between the original firing rates and estimated firing rates found for the best classifier was 0.728 ±0.276.
Significance: The proposed method demonstrates the possibility of CAP classification in peripheral neural signals, allowing for more precise control signals in a wide range of applications.