Deep learning and session-specific rapid recalibration for dynamic hand gesture recognition from EMG.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Wearable electromyographic (EMG) interfaces face challenges adapting to daily user variations across multiple sessions.
• Existing literature inadequately addresses session-specific differences for long-term EMG model learning.
• Developing robust EMG interfaces requires methods to handle inter-session variability.

Purpose of the Study:

• To introduce the novel MiSDIREKt dataset for multi-session EMG and kinematics recordings.
• To analyze the dataset using a non-linear encoder-decoder for gesture classification.
• To evaluate methods for adapting EMG models to session-specific variations.

Main Methods:

• Acquisition of the MiSDIREKt dataset: 43 sessions over 12 days from a single participant performing hand interaction tasks in virtual reality (totaling 814 minutes).
• Utilized a novel hardware design for recording EMG and kinematics.
• Employed a non-linear encoder-decoder architecture for dimensionality reduction and gesture classification.

Main Results:

• A recalibration approach using a small amount of single-session data achieved 79.5% accuracy for gesture classification within that session.
• Models trained solely on single-session data achieved 49.6% accuracy.
• Models trained only on the initial training data achieved 55.2% accuracy.

Conclusions:

• Recalibration with minimal new data is crucial for effective long-term EMG interface performance.
• The MiSDIREKt dataset provides a valuable resource for developing adaptive EMG systems.
• Non-linear encoder-decoder models show promise for robust gesture recognition in dynamic user environments.