Context Informed Incremental Learning Improves Myoelectric Control Performance in Virtual Reality Object Manipulation...

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Electromyography (EMG)-based gesture recognition systems excel in labs but falter in real-world applications due to performance degradation during real-time control.
• This performance decline stems from uncaptured goal-directed behaviors in static, offline training scenarios.

Purpose of the Study:

• To investigate the efficacy of Context Informed Incremental Learning (CIIL) for real-time adaptation of EMG classifiers in an object-manipulation task.
• To compare the performance of CIIL with a traditional open-loop approach in a virtual reality (VR) environment.

Main Methods:

• Implemented CIIL for continuous classifier adaptation using contextual cues in a VR object-manipulation task.
• Recruited nine participants without upper limb differences to perform functional tasks involving object transport with life-like grips.
• Compared a CIIL-based real-time adaptation scenario against a traditional open-loop system without adaptation.

Main Results:

• The CIIL-based approach significantly enhanced task success rates and efficiency.
• Perceived workload was reduced by 7.1% using the CIIL method.
• A minor 5.8% reduction in offline classification accuracy was observed with CIIL, while real-time performance improved.

Conclusions:

• Real-time contextualized adaptation using CIIL shows significant potential for improving user experience and usability of EMG-based systems.
• This adaptive strategy is crucial for practical, goal-oriented applications, paving the way for wider adoption of EMG interfaces.
• The study demonstrates the value of dynamic adaptation in overcoming limitations of static classifiers in dynamic human-computer interaction scenarios.