Subject-Independent sEMG-Based Prosthetic Control Using MAMBA2 with Domain Adaptation

Abstract

Integrating functional wrist articulation in prosthetic robot arms is crucial for enhancing natural movement and reducing compensatory upper limb motions. However, two significant challenges remain in electromyography (sEMG)-based prosthetic control: (1) real-time processing via efficient model design and (2) cross-subject generalization to address the individual variability in muscle signals. This study employs the MAMBA2 architecture to address the first challenge, leveraging Structured State Space Models (SSM) for efficient long-sequence inference. This enables real-time control with minimal computational overhead, making it well-suited for prosthetic robot arm applications. To tackle the second challenge, we implement a Representation Subspace Distance (RSD)-based Unsupervised Domain Adaptation (UDA), which preserves feature scale while aligning inter-subject variations, mitigating domain shift effects, and improving subject-independent wrist movement estimation. The model is trained on the Ninapro DB2 dataset, utilizing multi-channel sEMG signals and corresponding wrist kinematics. Evaluation results demonstrate that the MAMBA architecture outperforms conventional recurrent neural networks, achieving lower Mean Squared Error (MSE) and higher R2 values, with the Attention variant exhibiting the best prediction performance. Furthermore, this study highlights that the proposed UDA approach, combined with RSD-based alignment, significantly enhances cross-subject performance, reducing the need for extensive calibration. By enabling real-time processing through a computationally efficient model structure and effectively addressing cross-subject variability, this study contributes to developing a more reliable and generalizable sEMG-based robotic prosthesis controller, ultimately improving its applicability across diverse individuals.