Evaluation of feature projection techniques in object grasp classification using electromyogram signals from...

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Signal Processing

Background:

• Myoelectric prostheses utilize electromyogram (EMG) signals for control, but high-dimensional feature vectors can lead to computational inefficiency and data redundancy.
• Effective feature projection is crucial for improving the performance and reducing the complexity of EMG-based control systems.

Purpose of the Study:

• To evaluate and compare the effectiveness of four feature projection techniques (PCA, LDA, t-SNE, SRELM) for object grasp classification in myoelectric prostheses.
• To assess the impact of combining different feature projection types (linear/nonlinear, supervised/unsupervised) with various classifiers.

Main Methods:

• Applied Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Spectral Regression Extreme Learning Machine (SRELM) to EMG and IMU sensor data.
• Evaluated all pairwise combinations of the four feature projection techniques with seven different classifiers for nine upper limb positions.
• Utilized data from six EMG channels and IMU sensors for object grasp classification tasks.

Main Results:

• SRELM (nonlinear supervised) outperformed LDA (linear supervised), and t-SNE (nonlinear unsupervised) outperformed PCA (linear unsupervised).
• Classification errors ranged from 1.50% to 2.65% for SRELM and 1.27% to 17.15% for t-SNE across classifiers.
• SRELM showed no significant difference in performance based on classifier type (p=0.334), achieving a lowest error of 1.50% with a neural network.
• t-SNE achieved the lowest overall classification error of 1.27% when paired with a k-nearest neighbors classifier, but performance varied significantly with classifier choice.

Conclusions:

• Nonlinear supervised (SRELM) and unsupervised (t-SNE) feature projection methods offer significant advantages over linear methods for EMG-based grasp classification.
• SRELM provides robust performance irrespective of the chosen classifier, simplifying system design for myoelectric prostheses.
• t-SNE demonstrates potential for achieving very low classification errors but necessitates careful selection of the accompanying classifier for optimal prosthetic control.