Visualized Evidences for Detecting Novelty in Myoelectric Pattern Recognition using 3D Convolutional Neural Networks^.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Myoelectric pattern recognition (MPR) enables dexterous control of multi-degree-of-freedom prosthetic devices.
• Outlier data interference poses a significant challenge to the stability and application of conventional MPR systems.
• Existing methods struggle to effectively distinguish between intended movements and unexpected signals.

Purpose of the Study:

• To propose a novel method using 3D convolutional neural networks (CNNs) for robust myoelectric pattern recognition.
• To effectively extract spatial-temporal features from high-density surface electromyogram (EMG) data processed as a video stream.
• To discriminate and reject outlier data interference for improved MPR control stability.

Main Methods:

• High-density surface EMG recordings were processed as a video stream to capture time-varying muscular activity.
• A 3D CNN was employed to extract spatial-temporal features for pattern characterization.
• t-Distributed Stochastic Neighbor Embedding (t-SNE) was used for visualization and confirmation of feature discrimination.
• Mahalanobis Distance (MD) was applied for novelty detection and classification of targeted tasks.

Main Results:

• The 3D CNN effectively extracted spatial-temporal features, clearly separating targeted task patterns from outlier patterns.
• t-SNE visualization confirmed distinct clustering of targeted patterns and scattered distribution of outlier patterns.
• The proposed method achieved an average error rate of 10.98% for targeted tasks and less than 5% for outlier data.
• Performance significantly outperformed a common baseline method in rejecting outlier interferences.

Conclusions:

• The proposed 3D CNN method offers an effective way to characterize myoelectric patterns by leveraging spatial-temporal features.
• This approach demonstrates significant potential for rejecting outlier data interference, enhancing the stability of MPR control.
• The findings support the advancement of dexterous prosthetic control systems through improved signal processing and pattern recognition techniques.