Deep Learning for Musculoskeletal Force Prediction.

Area of Science:

• Biomechanics
• Computational modeling
• Machine learning

Background:

• Traditional musculoskeletal models are crucial for understanding internal forces during dynamic movement but are often slow and data-intensive.
• Supervised learning offers a promising alternative for creating faster, more flexible computational models.

Purpose of the Study:

• To develop and validate a deep neural network for rapid and accurate estimation of internal musculoskeletal forces.
• To compare the performance of the deep neural network against traditional musculoskeletal modeling and benchmark competition results.

Main Methods:

• A deep neural network was trained using kinematic, kinetic, and electromyographic data from 156 subjects during gait.
• The network learned the mapping from movement (kinematic/kinetic) to muscle (force) space.
• Model performance was evaluated against traditional musculoskeletal modeling and benchmark data from the Grand Challenge competitions.

Main Results:

• The deep neural network demonstrated good concordance with internal force predictions from traditional musculoskeletal models.
• The trained network outperformed winning submissions in four out of six international Grand Challenge modeling competitions.
• Computational speedup allows for real-time force estimation in a lab-based system.

Conclusions:

• Deep neural networks provide a computationally efficient and accurate method for estimating internal musculoskeletal forces.
• This approach facilitates real-time clinical applications and offers new insights into movement biomechanics.
• The trained networks reveal population-level relationships between kinematic and kinetic factors in human movement.