Lower-Limb Joint Torque Prediction Using LSTM Neural Networks and Transfer Learning.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Estimating joint torques is crucial for assessing athletic training, medical interventions, and assistive device efficacy.
• Wearable sensors offer a promising avenue for real-time joint torque estimation during daily activities.

Purpose of the Study:

• To predict lower limb joint torques during ten daily activities using long short-term memory (LSTM) neural networks and transfer learning.
• To evaluate the accuracy and generalizability of LSTM models in both intra-subject and inter-subject scenarios.

Main Methods:

• LSTM neural networks were trained using muscle electromyography (EMG) signals and lower limb joint angles.
• Models predicted hip, knee, and ankle joint torques.
• Transfer learning was employed to enhance model generalizability across different subjects and tasks.

Main Results:

• LSTM models achieved accurate predictions of lower limb joint torques with low error (RMSE ≤ 0.14 Nm/kg, NRMSE ≤ 8.7%) in intra-subject tests.
• Transfer learning significantly improved torque prediction accuracy in inter-subject tests, enabling accurate predictions from limited data from new subjects.

Conclusions:

• LSTM models, combined with wearable sensor data, provide accurate estimations of lower limb joint torques during daily activities.
• Transfer learning enhances the generalizability of these models, making them applicable to new subjects and tasks with minimal training data.