[Paper Review] Gait Cycle-Inspired Learning Strategy for Continuous Prediction of Knee Joint Trajectory from sEMG

Xueming Fu, Hao Zheng+10 more|arXiv (Cornell University)|Jul 25, 2023

Muscle activation and electromyography studiesEngineering3 citations

TL;DR

This paper proposes a gait cycle-inspired learning strategy to improve continuous prediction of knee joint trajectory from surface electromyography (sEMG). By decoupling joint angles into motion patterns and amplitudes and using muscle principal activation masks from prolonged walking, the model achieves a mean RMSE of 3.03° with 50ms prediction lead time—representing a 9.5% reduction over prior work.

ABSTRACT

Predicting lower limb motion intent is vital for controlling exoskeleton robots and prosthetic limbs. Surface electromyography (sEMG) attracts increasing attention in recent years as it enables ahead-of-time prediction of motion intentions before actual movement. However, the estimation performance of human joint trajectory remains a challenging problem due to the inter- and intra-subject variations. The former is related to physiological differences (such as height and weight) and preferred walking patterns of individuals, while the latter is mainly caused by irregular and gait-irrelevant muscle activity. This paper proposes a model integrating two gait cycle-inspired learning strategies to mitigate the challenge for predicting human knee joint trajectory. The first strategy is to decouple knee joint angles into motion patterns and amplitudes former exhibit low variability while latter show high variability among individuals. By learning through separate network entities, the model manages to capture both the common and personalized gait features. In the second, muscle principal activation masks are extracted from gait cycles in a prolonged walk. These masks are used to filter out components unrelated to walking from raw sEMG and provide auxiliary guidance to capture more gait-related features. Experimental results indicate that our model could predict knee angles with the average root mean square error (RMSE) of 3.03(0.49) degrees and 50ms ahead of time. To our knowledge this is the best performance in relevant literatures that has been reported, with reduced RMSE by at least 9.5%.

Motivation & Objective

• To address inter- and intra-subject variability in sEMG-based knee joint trajectory prediction.
• To improve continuous, real-time prediction of lower limb motion intent for exoskeleton and prosthetic control.
• To reduce prediction error by leveraging gait cycle structure and muscle activation patterns.
• To decouple joint angle features into low-variability motion patterns and high-variability amplitudes for better generalization.

Proposed method

• Decompose knee joint angles into motion patterns (low inter-subject variability) and amplitudes (high variability) for separate learning.
• Train dedicated neural networks for motion patterns and amplitudes to capture both shared and personalized gait features.
• Extract muscle principal activation masks from extended gait cycles to filter non-gait-related sEMG components.
• Use the masks as auxiliary guidance to enhance gait-related feature learning in the sEMG encoder.
• Integrate the decoupled predictions to reconstruct full knee joint angle trajectories.
• Optimize the model end-to-end with a loss function minimizing RMSE between predicted and actual joint angles.

Experimental results

Research questions

• RQ1 Can decoupling knee joint angles into motion patterns and amplitudes improve sEMG-based trajectory prediction accuracy?
• RQ2 How effective is using gait cycle-derived muscle activation masks in filtering non-gait-related sEMG noise?
• RQ3 To what extent does the proposed learning strategy reduce inter- and intra-subject variability in sEMG prediction?
• RQ4 Can the model achieve real-time, continuous prediction with a significant lead time over actual movement?
• RQ5 How does the model compare to state-of-the-art methods in terms of RMSE and prediction latency?

Key findings

• The model achieved a mean root mean square error (RMSE) of 3.03° (±0.49°) in predicting knee joint trajectories.
• The model predicted knee angles 50ms ahead of time, enabling early intervention in assistive devices.
• The proposed method reduced RMSE by at least 9.5% compared to the best-performing prior methods.
• The decoupling of motion patterns and amplitudes significantly improved generalization across subjects with diverse physiological traits.
• The use of gait cycle-inspired muscle activation masks enhanced feature learning by filtering out non-gait-related muscle activity.
• The model demonstrated robust performance on prolonged walking data, confirming the stability of the gait cycle-based feature extraction.