Estimation of Joint Torque by EMG-Driven Neuromusculoskeletal Models and LSTM Networks

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Accurate joint torque prediction is vital for developing effective assist-as-needed exoskeleton controllers.
• Wearable sensors, electromyography (EMG), and kinematics are key inputs for estimating human movement dynamics.

Purpose of the Study:

• To estimate joint torques (ankle, knee, hip) during daily activities using neuromusculoskeletal (NMS) and long short-term memory (LSTM) models.
• To evaluate the impact of a clustering approach on the accuracy of torque estimation models.
• To compare the performance of pooled, individual, and clustered NMS and LSTM models.

Main Methods:

• Utilized electromyography (EMG) and kinematic data to estimate joint torques.
• Employed neuromusculoskeletal (NMS) models and long short-term memory (LSTM) networks for torque prediction.
• Implemented a clustering approach to group movements by similarity and assessed its effect on model accuracy.
• Obtained ground truth joint torques via inverse dynamics from motion capture data.

Main Results:

• Both NMS and LSTM models demonstrated high joint torque estimation accuracy.
• Clustered and individual NMS models performed similarly well, outperforming the generic pooled model.
• The clustering approach showed minimal impact on the already high accuracy of LSTM models.
• Individual, clustered, and pooled LSTM models all achieved relatively high estimation accuracy.

Conclusions:

• Neuromusculoskeletal and LSTM models are effective for estimating joint torques from wearable sensor data.
• Clustering can enhance NMS model accuracy for specific movement patterns, though its benefit for LSTM is limited.
• Findings provide guidelines for selecting appropriate models to improve exoskeleton controller design and functionality for rehabilitation.