Scalability of random forest in myoelectric control

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Robot Interaction

Background:

• Myoelectric control systems translate electromyographic (EMG) signals into commands for human-robot interaction.
• EMG signal variability degrades system performance, often necessitating large, complex deep neural networks.
• A need exists for simpler, explainable, and scalable models for practical myoelectric control implementation.

Purpose of the Study:

• To investigate the scalability of Random Forest (RF) models for myoelectric control.
• To explore methods for scaling RF models up and down for pre-training, fine-tuning, and self-calibration.
• To identify factors influencing RF model size and accuracy in EMG-based control.

Main Methods:

• Systematic study of RF scalability using EMG data from 106 participants with varying electrode densities.
• Analysis of factors including bootstrapping, decision tree editing (pruning, grafting, appending), and training data size.
• Evaluation of RF model size reduction and accuracy changes under different configurations.

Main Results:

• Optimized RF models achieved a ≈500× reduction in size with only a 1.5% decrease in accuracy.
• Increased EMG electrode input dimensions led to a reduction in RF model size.
• The study systematically analyzed the impact of various factors on RF model performance and size.

Conclusions:

• Random Forests provide a highly efficient and scalable solution for myoelectric control.
• RF models can be flexibly scaled, facilitating real-time deployment in practical applications.
• Findings support the use of RFs as a viable alternative to large deep neural networks for EMG-based control.