AI Model for Extracting Patterns from EMG Data

Application

Aritificial neural network-based dynamical systems modeling on electromyography (EMG) data for simultaneously estimating de-noised, high-resolution muscle activation signals across multiple muscles with millisecond-timescale precision.

Key Benefits

• Generates models that produce estimates that avoid trivial output solutions (e.g., replicating EMG data).
• Captures underlying spatial and temporal structure across multi-muscle EMG signals to eliminate noise.
• Improved performance for models trained with these methods by providing superior “de-noising” of the data, resulting in better prediction of movement-related variables associated with the EMG data.
• Leverages shared information across simultaneous multi-muscle EMG recordings to inform the de-noising algorithm so as not to treat them as independent signals.

Market Summary

Electromyography (EMG) is a bio-sensing technique used for measuring electrical activity from muscles over time. Accurately capturing these muscle activation signals can provide critical information to diagnose neuromuscular disorders or provide useful control signals for assistive devices like prosthetics. However, EMG data can often be temporally complex and often corrupted with noise, making it difficult to extract and interpret meaningful signals from the raw data. Standard approaches (e.g., filtering) often treat individual muscles independently, failing to leverage key information accessible only when studying co-activation of multiple muscles. Further, they arbitrarily set boundaries on what information is considered “signal” versus “noise”. The rise in more complex machine learning models trained on EMG data have the potential to apply more complex de-noising operations than standard approaches, however, these may suffer from the susceptibility to finding trivial solutions, in other words, preserving noise and other irrelevant information that is corrupting the raw EMG data.

Technical Summary

To address the need for effective signal extraction for valuable muscle activation estimation, researchers have developed an AI model to provide estimates of muscle activation patterns across multiple muscles simultaneously. This model utilized AutoLFADS (Automatic Latent Factor Analysis via Dynamical Systems), a machine learning tool typically used to de-noise and analyze neural activity data, to extract and model underlying latent, or hidden, patterns in EMG data. Testing on a rat hindlimb during locomotion showed superior estimation of muscle activations, resulting in improved prediction of multiple hindlimb joint angular kinematics in comparison to traditional filtering techniques such as low-pass or Bayesian filtering. Testing on a monkey forearm showed that AutoLFADS-inferred estimates of muscle activation preserved high-frequency, behaviorally relevant features better than other tested approaches. This technology has the potential to improve brain-machine interfaces, neuromotor interfaces, and muscle-controlled prostheses that rely on EMG signals to provide accurate estimates of the “neural drive” of motor intent.

Development Stage

• Successfully estimated muscle activations from EMG signals in rat hindlimbs, monkey forearms, and cat hindlimbs using the model.
• Currently developing methods to apply this approach to high-density surface EMG signals in humans.
• Human application is still in progress and under development.