Myoelectric and Inertial Data Fusion Through a Novel Attention-Based Spatiotemporal Feature Extraction for...

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Human-Computer Interaction

Background:

• Individuals with transhumeral amputation require intuitive control for full-arm prostheses.
• Existing control strategies often lack reliability and intuitive activation.
• Underinvestigated need for advanced feature extraction in prosthetic control.

Purpose of the Study:

• To propose and evaluate a novel feature extraction scheme fusing accelerometric (ACC) and electromyographic (EMG) data.
• To improve shoulder movement identification for individuals with transhumeral amputation.
• To assess the impact of ACC data fusion on pattern recognition models for prosthetic control.

Main Methods:

• A novel spatiotemporal warping architecture was used for feature-level fusion of EMG and ACC data.
• Data collected from participants with intact limbs and transhumeral amputations at 1000 Hz.
• Leave-one-trial-out (LOTO) cross-validation used for training and testing pattern recognition models (LDA, ELM, kNN).

Main Results:

• Fusion of ACC data positively impacted identification accuracy for window lengths below 150 ms.
• Linear Discriminant Analysis (LDA) and Extreme Learning Machine (ELM) classifiers achieved >90% accuracy for both groups.
• k-Nearest Neighbors (kNN) and autonomous learning multi-model classifiers showed <87% mean accuracy, indicating model-specific performance.

Conclusions:

• The proposed feature extraction scheme effectively fuses ACC and EMG data for improved shoulder movement identification.
• The findings support the applicability of shallow pattern recognition models in real-world prosthetic control scenarios.
• This study provides a foundation for future real-time validation and larger population studies.