Solving musculoskeletal biomechanics with machine learning.

Area of Science:

• Biomechanics
• Computational Science
• Machine Learning

Background:

• Musculoskeletal dynamics involve complex, high-dimensional spatial transformations.
• Machine learning (ML) offers a promising approach to approximate these relationships.
• Understanding posture-dependent muscle geometry is crucial for biomechanical modeling.

Purpose of the Study:

• To evaluate general ML algorithms for approximating posture-dependent moment arm and muscle length relationships.
• To compare the performance of artificial neural networks (ANN) and light gradient boosting machines (LGB) in modeling human arm and hand muscles.
• To assess the efficiency and accuracy of ML models for complex musculoskeletal transformations.

Main Methods:

• Utilized ANN and LGB algorithms to model wrapping kinematics for 33 arm and hand muscles across 18 degrees of freedom.
• Generated training and testing datasets using a phenomenological model with joint angles as input and muscle length/moment arms as output.
• Compared model accuracy, sample efficiency, and evaluation speed between ANN and LGB approaches.

Main Results:

• Both ANN and LGB models demonstrated high accuracy, with low percentage errors for muscle lengths and moment arms.
• LGB models required significantly fewer training samples (10^3) compared to ANN models (10^6).
• ANN models were substantially faster during evaluation (39x) than LGB models.

Conclusions:

• Developed ML models sufficiently approximate musculoskeletal transformations, validating their applicability.
• The study highlights the potential of ML in diverse applications, including advanced powered prosthetics.
• Both ANN and LGB present viable, albeit different, solutions for modeling complex biomechanical relationships.