LSTM-AE for Domain Shift Quantification in Cross-Day Upper-Limb Motion Estimation Using Surface Electromyography.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Deep learning (DL) models for upper-limb myoelectric control face challenges in cross-day robustness due to unstable surface electromyography (sEMG) signals.
• This signal variability leads to domain shift, negatively impacting DL model performance in real-world applications.

Purpose of the Study:

• To propose and validate a reconstruction-based method for quantifying domain shift in myoelectric control.
• To assess the impact of domain shift on the performance of hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models.
• To establish a correlation between quantified domain shift and model performance degradation.

Main Methods:

• A hybrid CNN-LSTM framework was utilized as the primary model for myoelectric control tasks (hand gesture classification and wrist kinematics regression).
• A Long Short-Term Memory Auto-Encoder (LSTM-AE) was developed to reconstruct features extracted by the CNN component.
• Domain shift was quantified by measuring the reconstruction errors (RErrors) generated by the LSTM-AE.

Main Results:

• Reconstruction errors (RErrors) from LSTM-AE increased significantly when performance degraded in between-day testing, distinguishing them from within-day errors.
• A strong negative correlation was observed between LSTM-AE errors and the performance of the CNN-LSTM model.
• Average Pearson correlation coefficients reached -0.986 ± 0.014 for classification and -0.992 ± 0.011 for regression.

Conclusions:

• The proposed LSTM-AE method effectively quantifies domain shift in sEMG signals for myoelectric control.
• Quantified domain shift is a reliable indicator of performance degradation in CNN-LSTM models.
• This approach offers a pathway to enhance the robustness and reliability of myoelectric control systems in daily use.