tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

BiLSTM-Transformer: A Novel Hybrid Model for Accurate Prediction of Hand Joint Angles from sEMG Signals

Authors :

Anita Sadat Sadati Rostami¹ Alireza Nazari² Mohammadreza Nayeri³

1- دانشگاه آزاد اسلامی واحد کرج 2- دانشگاه علم و صنعت ایران 3- دانشگاه تهران

Keywords :

surface electromyography،sEMG،hand kinematics،time series،Joint estimation،BiLSTM-Transformer،NinaPro DB8،Human-machine interfaces،Regression

Abstract :

Surface electromyography (sEMG) signals offer a non-invasive pathway for advancing human-machine interfaces, yet predicting continuous hand joint angles is challenging due to signal variability and complexity. This study proposes a novel hybrid model integrating bidirectional long short-term memory (BiLSTM) and Transformer architectures to accurately predict 10 degrees of freedom (DOF) in hand joints—metacarpophalangeal (MCP) and proximal interphalangeal (PIP)—using the NinaPro DB8 dataset. The model captures temporal muscle dynamics through BiLSTM and models long-range dependencies with a Transformer encoder, enhanced by residual connections for training stability. To enhance efficiency, a comprehensive preprocessing pipeline extracts time-domain features, applies dimensionality reduction, and employs standardization to optimize performance. A custom regularized loss function mitigates overfitting, ensuring robust predictions. The model achieves a mean absolute error of 8.2° across all 10 joints, with 6.8° for metacarpophalangeal joints and 9.7° for proximal interphalangeal joints, outperforming prior benchmarks focused on fewer aggregated joint angles. This approach enables precise, naturalistic hand pose reconstruction for prosthetic control, offering scalability and personalization for rehabilitation applications.