Long-term upper-extremity prosthetic control using regenerative peripheral nerve interfaces and implanted EMG electrodes

Abstract

Objective.Extracting signals directly from the motor system poses challenges in obtaining both high amplitude and sustainable signals for upper-limb neuroprosthetic control. To translate neural interfaces into the clinical space, these interfaces must provide consistent signals and prosthetic performance.Approach.Previously, we have demonstrated that the Regenerative Peripheral Nerve Interface (RPNI) is a biologically stable, bioamplifier of efferent motor action potentials. Here, we assessed the signal reliability from electrodes surgically implanted in RPNIs and residual innervated muscles in humans for long-term prosthetic control.Main results.RPNI signal quality, measured as signal-to-noise ratio, remained greater than 15 for up to 276 and 1054 d in participant 1 (P1), and participant 2 (P2), respectively. Electromyography from both RPNIs and residual muscles was used to decode finger and grasp movements. Though signal amplitude varied between sessions, P2 maintained real-time prosthetic performance above 94% accuracy for 604 d without recalibration. Additionally, P2 completed a real-world multi-sequence coffee task with 99% accuracy for 611 d without recalibration.Significance.This study demonstrates the potential of RPNIs and implanted EMG electrodes as a long-term interface for enhanced prosthetic control.

Keywords: myoelectric control; neuroprosthetics; pattern recognition; peripheral nerve regeneration.

Figure 1.

Experiment set-up and task description. (a) Participants mimicked a virtual hand with their phantom limb to collect intramuscular EMG for offline analyses and decoder calibration (solid red arrow). P2 controlled a second virtual hand to match cued postures during online decoding sessions (dashed red arrow). (b) To quantify decoding performance over time, P2 completed the virtual task in four static arm positions, with her inactive prosthesis donned to simulate weight effects. (c) P2 then used her prosthesis to complete a multi-grasp coffee making task.

Figure 2.

Measured signal-to-noise ratios (SNRs) over time. (a) Measured SNRs for P1 and P2 during volitional phantom movements of thumb and small finger flexion. SNRs remained high for both RPNIs and residual muscle channels with no decreasing linear trend (p > 0.05, F-test). However, SNRs did vary from session to session. The dashed line represents the mean SNR of surface EMG estimated from literature [67, 68]. (b) Boxplots representing the median root mean squared (RMS) of EMG for P1 and P2’s intramuscular electrode channels. Orange lines show the median, blue box shows the interquartile range (IQR), black dashed lines show the most extreme non-outlier values and red crosses show outliers more than 1.5 times the IQR.

Figure 3.

Online four-grip classifier decoding performance across time without recalibration. (a) Decoding accuracy was measured every 50 ms, accounting for transition errors that occurred within each trial. Performance was quantified over four different arm positions: arm at side (blue), arm raised (navy blue), arm front (gold), and arm across (red). (b) Cumulative decoding performance within each session showed no significant linear decreasing performance over time (p = 0.11, F-test). Gold dashed line represents the linear fit across data points. (c) Timeline of when P2 received RPNIs, electrode implantation, and when the four-grip classifier was trained. Last quantified decoding session occurred 1424 d after RPNI creation.

Figure 4.

Breakdown of online decoding performance for each arm position across all sessions. (a) Online confusion matrices representing the overall accuracy for each grip (rest, fist, pinch, point) at each arm position (arm at side, arm raised, arm in front, arm across). Confusion matrix captures transition errors to cued grips while P2 controlled the virtual hand in real-time. (b) EMG mean absolute value (MAV) features from both RPNIs, and residual muscles were decomposed into a two-dimensional space for cluster visualization. Cluster separation was seen across all grips: rest (light blue), fist (navy blue), pinch (gold), point (red). Magnitude of the black solid lines represent the contribution each channel provided to each grip. (c) Decoder latency was measured as the time difference between the onset of new EMG activity and a successful posture. The median (dashed line) and middle 50% (shading) is overlaid on histograms binned in 50 ms increments (n = 1179 trials). Trials with latency greater than a second (>1) are aggregated in the orange rectangle.

Figure 5.

Offline decoding analysis of nine movements over time. (a) Decoding accuracy of Hidden Markov Model (HMM-NB), Naïve Bayes (NB), and Linear Discriminant Analysis (LDA) classifiers over time, without recalibration. Each classifier was trained on nine finger and wrist movements: rest, thumb flexion (T), index finger flexion (I), middle finger flexion (M), ring finger flexion (R), small finger flexion (S), wrist flexion (WF), finger abduction (Ab), and finger adduction (Ad). (b) Offline confusion matrices representing performance of LDA averaged across all sessions (n = 5 and 7 sessions for P1 and P2).

Figure 6.

P1 offline decoding performance with RPNIs and residual muscles vs. only residual muscles. (a), (b) Confusion matrices show a Linear Discriminant Analysis (LDA) classifier predicting four movements for grasp control: fist (F), pinch (Pi), point (Po), and rest (Re) with and without input from P1’s RPNIs. (c), (d) Prediction performance of the individual finger movements: rest, thumb flexion (T), index finger flexion (I), middle finger flexion (M), ring finger flexion (R), small finger flexion (S), wrist flexion (WF), finger abduction (Ab), and finger adduction (Ad), with and without RPNIs. (e), (f) Distinguishing rest and index finger flexion from four thumb and intrinsic finger movements: thumb flexion, finger abduction, finger adduction, and thumb opposition (TO), with and without RPNIs.

Figure 7.

P2 offline decoding performance with RPNIs and residual muscles vs. only residual muscles. (a), (b) Confusion matrices show a Linear Discriminant Analysis (LDA) classifier predicting the four movements for grasp control with and without RPNIs. (c), (d) Prediction performance of the nine individual finger movements analyzed with and without RPNIs. (e), (f) Distinguishing rest and index finger flexion from the four thumb and intrinsic finger movements with and without P2’s RPNIs.