Adaptive EMG decomposition in dynamic conditions based on online learning metrics with tunable hyperparameters

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Neural decoders require robustness to non-stationary conditions for sustained accuracy.
• Dynamic muscle contractions present significant challenges for traditional stationary decoders.

Purpose of the Study:

• To develop a novel adaptive electromyography (EMG) decomposition algorithm for real-time decoding of motor neuron (MN) discharges.
• To address challenges posed by non-stationary conditions, particularly during dynamic muscle contractions.

Main Methods:

• Utilized blind source separation methods enhanced with Kullback-Leibler divergence and kurtosis for online learning.
• Developed a theoretical framework for tuning adaptation hyperparameters and compensating for mixing matrix non-stationarities.
• Implemented real-time adaptation with a computational time of approximately 22 ms per 100 ms batch.

Main Results:

• Hyperparameters captured anatomical variations and generalized across subjects.
• The adaptive algorithm significantly improved decomposition performance metrics compared to non-adaptive methods across an 80° wrist range of motion.
• Achieved >=90% agreement rate, sensitivity, and precision in >=80% of cases for both simulated and experimental data.

Conclusions:

• The proposed online learning metrics and hyperparameter optimization are suitable for decoding MN discharges in dynamic conditions.
• The study validates an experimental method for EMG decomposition during dynamic tasks.