Lower Limb Torque Prediction for Sit-To-Walk Strategies Using Long Short-Term Memory Neural Networks.

Area of Science:

• Biomechanics
• Assistive Technology
• Machine Learning

Background:

• Accurate joint torque prediction is vital for biomechanics research, treatment evaluation, and designing powered assistive devices.
• Sit-to-walk (STW) movements require personalized torque assistance, adapting to individual strategies and anthropometry.

Purpose of the Study:

• To develop and evaluate long short-term memory (LSTM) neural networks for predicting hip and knee joint torques during STW.
• To generate strategy-specific and user-oriented torque trajectories for personalized assistive controllers.

Main Methods:

• Trained three LSTM neural networks on STW data from 65 subjects across different age groups.
• Utilized hip and knee angles and center of mass velocity as model inputs, with windowing for real-time adaptation.
• Compared LSTM performance based on STW strategies, focusing on temporal and spatial relationship recognition.

Main Results:

• The encoder-decoder LSTM demonstrated optimal performance, accurately recognizing temporal features.
• Achieved low root mean square error (0.24 ± 0.07 Nm/kg for hip, 0.15 ± 0.02 Nm/kg for knee) and high Spearman's correlation (93.43 ± 2.86% and 84.83 ± 2.96%).
• Exhibited good intraclass correlation coefficients (>0.75), confirming model reliability.

Conclusions:

• The developed encoder-decoder LSTM reliably predicts user- and strategy-specific reference torques for assistive devices.
• Enables more natural and personalized assistance during sit-to-walk transitions.
• Highlights the potential of deep learning for advancing personalized rehabilitation and assistive technologies.