A NMF-Based Non-Euclidean Adaptive Feature Extraction Scheme for Limb Motion Pattern Decoding in Pattern Recognition...

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Signal Processing

Background:

• Electromyogram (EMG) feature extraction is vital for motor intent decoding in prosthetic and assistive devices.
• Existing feature extraction methods often suffer from low decoding performance and are susceptible to dataset drift.
• The impact of training-test data drift on feature extraction performance is often overlooked in current evaluations.

Purpose of the Study:

• To propose a novel unsupervised feature extraction scheme for EMG-based pattern recognition systems.
• To address limitations of existing methods by reducing dataset drift and improving decoding accuracy.
• To enhance the reliability and robustness of EMG control systems for clinical and commercial applications.

Main Methods:

• Developed an unsupervised feature extraction scheme utilizing Non-negative Matrix Factorization (NMF) and Riemann operations for feature adaptation.
• Implemented data distribution alignment to minimize drift between training and testing datasets.
• Evaluated the proposed technique on 13 hand and finger movements for both amputee and able-bodied participants.

Main Results:

• Achieved significantly higher motor intent decoding performance (p < 0.05) compared to existing techniques.
• Attained high average accuracies: 99.91 ± 0.35% for amputees and 99.99 ± 0.02% for able-bodied individuals.
• Demonstrated superior decoding performance across varied signal-to-noise ratios (SNRs), highlighting robustness.

Conclusions:

• The proposed unsupervised feature extraction technique effectively reduces dataset drift and enhances EMG-based motor intent decoding.
• The method offers significant improvements in accuracy and robustness, outperforming conventional approaches.
• This technique holds promise for advancing the reliability of EMG control systems in real-world applications.