Gaussian Process Autoregression for Joint Angle Prediction Based on sEMG Signals.

Area of Science:

• Biomechanics
• Neuromuscular modeling
• Machine learning for human movement analysis

Background:

• Deterministic models struggle with neuromusculoskeletal system uncertainty, impacting prediction accuracy.
• Surface electromyography (sEMG) signals are crucial for understanding muscle activity but can be unstable.
• Accurate joint angle prediction is vital for rehabilitation and understanding human movement.

Purpose of the Study:

• To propose a novel knee joint angle prediction model using sEMG signals.
• To address the challenges of sEMG signal instability and neuromusculoskeletal uncertainty.
• To develop a non-parametric probabilistic model for enhanced human movement prediction.

Main Methods:

• Developed a Gaussian process (GP) model incorporating muscle activation physiology.
• Combined the GP model with a non-linear autoregressive with exogenous inputs (NARX) model, creating a Gaussian process autoregression (GAR) model.
• Evaluated the GAR model against a standard GP model using normalized root mean square error (NRMSE) and correlation coefficient (CC) across healthy and hemiplegic datasets.

Main Results:

• The GAR model demonstrated superior performance compared to the GP model in predicting knee joint angles across various speed scenarios.
• Significant improvements (p < 0.05) in both NRMSE and CC were observed for the GAR model in both healthy and hemiplegic datasets.
• The probabilistic approach effectively captured and predicted human movement dynamics with higher accuracy.

Conclusions:

• The Gaussian process autoregression model provides an accurate and robust method for non-parametric probabilistic joint angle prediction.
• This approach successfully accounts for the inherent uncertainty in the neuromusculoskeletal system and sEMG signal variability.
• The findings suggest a promising tool for applications requiring precise human movement analysis and prediction.