Page 81 - Read Online

Wang et al. Art Int Surg. 2025;5:465-75 https://dx.doi.org/10.20517/ais.2025.03 Page 467

prosthetic devices, presenting an opportunity for machine learning algorithms to play a crucial role. These
advanced computational techniques can decode and interpret the complex nerve signals from a patient’s
muscles, enhancing control strategies for more precise prosthetic operation. They can be leveraged to
analyze the multifaceted signals generated by Regenerative Peripheral Nerve Interfaces (RPNIs), translating
them into specific commands for advanced multi-articulated prosthetic devices. By combining the
biological advantage of RPNI with the computational power of machine learning, researchers can
significantly improve the performance of modern-day prosthetics. The primary objective of this narrative
review is to demonstrate how machine learning algorithms have enhanced the use of prosthetics, specifically
in their application to RPNI surgery, ultimately improving the quality of life for individuals suffering from
upper extremity amputations. While numerous alternative control strategies exist, including targeted
muscle reinnervation (TMR) and conventional myoelectric systems, this narrative review is intentionally
focused on RPNIs due to their role in biologically amplifying neural signals for advanced prosthetic control.

APPLICATION OF MACHINE LEARNING IN RPNI SURGERY
Continuous control
RPNIs [Figure 1] produce high-amplitude electromyography (EMG) signals with large SNRs, which are
ideal for machine learning algorithms to process. To decode the EMG signals from RPNIs, researchers have
applied various machine learning algorithms focusing on continuous control and pose identification. For
continuous control, the Kalman filter, a recursive filter and algorithm that predicts the current state of a
system (e.g., finger position) based on a model of the system’s dynamics and noisy measurements (e.g.,
EMG signals), has been used to translate EMG signals directly into finger movements [14-16] . It predicts finger
positions using a model of hand movement dynamics, updates these predictions based on EMG signal
measurements, and filters out noise to provide smooth, accurate estimates of intended movements. The
[17]
Wiener filter, another method used for continuous reconstruction of movement, works differently . It uses
a statistical model of the system’s input and output to estimate the system’s state . It can be thought of as a
[18]
system that learns the patterns between input (EMG signals) and output (finger movements) and uses this
knowledge to make predictions. In prosthetic applications, it uses a model of the relationship between EMG
signals and finger movements to estimate the most likely finger position based on observed EMG data,
[16]
minimizing the difference between the estimated and actual finger positions . Both the Kalman and
Wiener filters enable continuous control of prosthetic devices by translating EMG signals into smooth,
accurate finger movements, helping to bridge the gap between a user’s neural signals and the physical
actions of their prosthetic limb.

Vu et al. implanted RPNIs and intramuscular bipolar electrodes (IM-MES) in two rhesus macaques to
[7]
enable closed-loop continuous hand control. The non-human primates (NHPs) were trained to perform
finger movement tasks, and a flex sensor was attached to the NHP’s index finger to measure finger position
[Figure 2]. They hit targets on the screen by moving their fingers, and the virtual hand could be controlled
by either the flex sensor or by decoding RPNI signals in real time. EMG signals from the RPNIs were
recorded and filtered. A Kalman filter was used to decode continuous finger position from the EMG signals
using two temporal features of the EMG waveform: mean absolute value (MAV) and waveform line length
(LL). A Wiener filter was also implemented for comparison. The decoding algorithms were evaluated in
offline and online closed-loop sessions. Both Kalman and Wiener filters were applied to decode continuous
finger movements using RPNI signals under identical experimental conditions in rhesus macaques. While
both filters enabled real-time control, the Kalman filter demonstrated superior responsiveness, requiring a
shorter duration of neural data collection to generate accurate predictions. In contrast, the Wiener filter
performed similarly but was more computationally intensive and less suited for real-time performance.
These direct comparisons suggest that Kalman filtering may be more optimal for low-latency continuous
control in RPNI-based applications.