InverseMuscleNET: Alternative Machine Learning Solution to Static Optimization and Inverse Muscle Modeling.

Area of Science:

• Biomechanics
• Machine Learning
• Neuroscience

Background:

• Inverse muscle models traditionally rely on static optimization to resolve redundancy.
• This can be computationally intensive and may not capture dynamic temporal relationships.

Purpose of the Study:

• To propose InverseMuscleNET, a recurrent neural network (RNN) model, as an efficient alternative for estimating muscle activation signals.
• To evaluate the model's accuracy and identify key biomechanical inputs for muscle activation prediction.

Main Methods:

• A recurrent neural network (RNN) was configured, trained, and tested using biomechanical variables (joint angle, velocity, acceleration, torque, activation torque) as inputs.
• Surface electromyography (EMG) signals from shoulder flexion/extension experiments were used for training and validation.
• A sequential backward selection algorithm identified the most influential input variables.

Main Results:

• The RNN model achieved a normalized regression of 88-91% between experimental data and estimated muscle activation.
• Key inputs, in order of importance, were joint angle, activation torque, joint torque, joint velocity, and joint acceleration.
• The model effectively captured dynamic temporal relationships using previous biomechanical and EMG feedback signals.

Conclusions:

• InverseMuscleNET provides a fast, direct estimation of muscle activation, bypassing iterative solutions.
• The model enables potential real-time applications in functional rehabilitation and sports evaluation.
• This non-invasive method allows clinicians to estimate EMG activity without surface electrodes.