Validity and Impact of Methods for Collecting Training Data for Myoelectric Prosthetic Control Algorithms

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Robotics

Background:

• Intuitive control of prosthetic devices requires accurate correlation between biological signals and motor intent.
• High-quality training datasets are crucial for effective regression algorithms in prosthetics.
• Labeling intended hand kinematics for prosthetic training is challenging, especially for amputees.

Purpose of the Study:

• To compare the accuracy and precision of two training paradigms: mimic training and mirror training, for labeling hand kinematics.
• To evaluate the impact of these training methods on the performance of prosthetic control in both healthy individuals and amputees.

Main Methods:

• Quantified kinematic accuracy and precision using motion capture in healthy participants under mimic and mirror training conditions.
• Assessed offline regression performance using non-linear neural networks with large datasets.
• Evaluated online prosthetic control performance in unilateral transradial amputees performing a functional task.

Main Results:

• Mimic training failed to account for biomechanical coupling and temporal hand posture changes.
• Mirror training demonstrated significantly higher accuracy and precision in kinematic labeling.
• Mirror training resulted in superior offline regression performance and faster online task completion speeds for amputees.

Conclusions:

• Mirror training generates more faithful datasets, leading to improved prosthetic control.
• This approach offers a valuable strategy for developing next-generation myoelectric and neuroprostheses with enhanced dexterity and intuitive operation.