Continuous Lower Limb Biomechanics Prediction via Prior-Informed Lightweight Marker-GMformer

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Traditional lower limb musculoskeletal dynamics simulation faces challenges like force plate dependency, poor generalization, and high computational cost.
• These limitations hinder real-time applications in robot control systems needing rapid feedback.

Purpose of the Study:

• To propose the Marker-GMformer model for efficient and accurate continuous prediction of lower limb kinematics and dynamics.
• To reduce computational complexity while maintaining high performance and strong generalization capabilities.

Main Methods:

• Developed a marker trajectories-driven deep learning model integrating prior knowledge with global-local and spatial-temporal features.
• Inputted marker coordinate time series data for continuous prediction.
• Validated predictions against musculoskeletal multibody dynamics simulations and force plates across 13 motion patterns.

Main Results:

• Marker-GMformer achieved excellent performance with average Pearson correlation coefficients (ρ ≥ 0.97).
• Low root mean square errors were observed: 1.95° for angles, 0.036 body weight for ground reaction forces (GRFs), and 0.099 N·m/kg for moments.
• The model demonstrated strong generalization across diverse motion patterns.

Conclusions:

• The proposed Marker-GMformer model offers a computationally efficient and accurate solution for predicting lower limb mechanics.
• It shows significant promise for real-time monitoring and optimizing assistive robot control through timely feedback.