Estimating IMU signals from surface EMG using physics-informed and domain-adaptive neural networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Biomechanics

Background:

• Estimating human body kinematics from surface electromyography (EMG) signals is challenging due to the complex, non-linear relationship between muscle activity and motion.
• Existing methods often require domain-specific tuning or retraining for different tasks, limiting their generalizability.

Purpose of the Study:

• To develop a physics-informed, domain-adaptive neural network architecture for robust body kinematics estimation from EMG signals across heterogeneous human task domains.
• To address the biomechanical inverse mapping problem by directly predicting inertial measurement unit (IMU) outputs from EMG signals without task-specific adaptation.

Main Methods:

• A shared convolutional feature extractor followed by two branches: a regression head for IMU signal prediction and an adversarial domain-classification head using a gradient reversal layer.
• Incorporation of a physics-informed loss term penalizing the derivative of predicted acceleration (jerk) to ensure biomechanical plausibility.
• Evaluation using EMG-IMU data across five distinct task categories with varied sensor configurations.

Main Results:

• Achieved a root mean squared error of 0.42 ± 0.11 m/s² for acceleration and 6.38 ± 1.1°/s for gyroscope outputs across 57 IMU channels.
• The model demonstrated domain prediction accuracy of nearly 75%, indicating effective learning of task-invariant representations.
• Performance metrics varied based on domain complexity and overlap in EMG/motion patterns.

Conclusions:

• The proposed physics-informed, domain-adaptive neural network effectively estimates body kinematics from EMG signals across diverse tasks without retraining.
• The architecture learns generalizable representations while enforcing biomechanical constraints, offering a promising approach for motion analysis.
• This method advances the potential for non-invasive, accurate motion tracking in various applications.