Inter-Subject Domain Adaptation for CNN-Based Wrist Kinematics Estimation Using sEMG

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Convolutional Neural Networks (CNNs) are used for decoding surface Electromyography (sEMG) signals.
• CNN models struggle with inter-subject variability due to domain shift in sEMG data.
• Existing methods like fine-tuning face limitations in maintaining performance across diverse subjects.

Purpose of the Study:

• To propose a novel regression scheme for supervised domain adaptation (SDA) to enhance CNN performance for wrist kinematics estimation.
• To effectively reduce domain shift issues in sEMG signal analysis.
• To improve the reusability and inter-subject performance of CNN models.

Main Methods:

• A two-stream CNN with shared weights was developed to process source and target sEMG data concurrently.
• Domain-invariant features were extracted by simultaneously exploiting both datasets.
• CNN weights were tuned using regression losses for supervised learning and a domain discrepancy loss to minimize distribution divergence.

Main Results:

• The proposed regression SDA method demonstrated superior performance compared to fine-tuning in wrist kinematics estimation.
• The method showed effectiveness in both single-subject and multiple-subject scenarios.
• Regression SDA maintained better performance in original domains, unlike fine-tuning which suffers from catastrophic forgetting.

Conclusions:

• The novel regression SDA scheme effectively mitigates domain shift in sEMG data for CNN-based kinematics estimation.
• This approach significantly improves inter-subject performance and model reusability.
• The proposed method offers a robust solution for real-world applications requiring adaptable sEMG decoding.