Estimating muscle activation from EMG using deep learning-based dynamical systems models.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Estimating muscle activation from electromyographic (EMG) signals is crucial for understanding neural control of movement.
• Current methods for extracting neural command signals from EMG are limited, often requiring independent muscle analysis or manual hyperparameter tuning.
• The complex relationship between neural commands and recorded EMG signals presents a significant challenge in movement neuroscience.

Purpose of the Study:

• To adapt and apply AutoLFADS, a deep learning framework, for unsupervised estimation of multi-muscle activation from EMG recordings.
• To evaluate the performance of AutoLFADS in capturing dynamic muscle activation patterns during different behavioral tasks.
• To compare AutoLFADS-derived muscle activation estimates against traditional filtering methods and their correlation with neural activity.

Main Methods:

• Utilized AutoLFADS, an unsupervised deep learning approach employing recurrent neural networks, to model spatial and temporal patterns in multi-muscle EMG data.
• Applied the adapted AutoLFADS model to EMG recordings from rat hindlimb locomotion and monkey forearm isometric force tasks.
• Compared AutoLFADS performance against low-pass and Bayesian filtering for predicting joint kinematics and correlating with motor cortical activity.

Main Results:

• AutoLFADS dynamically adjusted its frequency response during rat locomotion, improving joint kinematics prediction compared to conventional filters.
• In monkey forearm tasks, AutoLFADS identified uncharacterized high-frequency EMG oscillations that correlated strongly with measured force.
• AutoLFADS-inferred muscle activation estimates showed higher correlation with simultaneously recorded motor cortical activity than other methods.

Conclusions:

• The adapted AutoLFADS method effectively estimates multi-muscle activation from EMG signals using dynamical systems and deep learning.
• This approach offers improved accuracy in muscle activation estimation, enhancing the prediction of movement kinematics and force.
• AutoLFADS provides a powerful tool for studying multi-muscle coordination, neural control, and advancing brain-machine interfaces.