Evaluation of Random Forest and LSTM for sEMG-Based Estimation of Two DOF Wrist Kinematics

Abstract

Robotic wrist rehabilitation after stroke requires control signals that reflect the user’s volitional intent, while clinically deployable systems are constrained by the number of surface electromyography (sEMG) channels that can be practically placed on a hemiparetic forearm. This study evaluates how far a single sEMG channel can support discrete motion-intent classification and continuous two-degree-of-freedom (2-DOF) wrist kinematic estimation for exoskeleton control. Two anonymised datasets, recorded at 200 Hz using one forearm sEMG channel and one inertial measurement unit (IMU), were processed through detrending, 50 Hz notch filtering, fourth-order zero-phase 20–90 Hz band-pass filtering, and 250 ms windowing with 50% overlap. Seven time- and frequency-domain features were extracted and provided to Random Forest (RF), long short-term memory (LSTM), and hybrid RF–LSTM models. Under a strict held-out-trial protocol, LSTM achieved higher five-class classification accuracy than RF (95.56% versus 87.28%). Regression performance was weaker: LSTM produced the best flexion–extension result (R² = 0.290, RMSE = 12.49°), while hybrid RF–LSTM performed best for radial–ulnar estimation (R² = 0.204, RMSE = 12.35°). Closed-loop Dynamixel MX-28AT tracking showed zero overshoot. Overall, single-channel sEMG supports reliable intent classification, but proportional angle estimation remains limited, indicating that further sensing or regression-specific feature optimization is required for precise control.