Enhanced Joint Angle Estimation Using Support Vector Machine-Long Short-Term Memory Fusion with Electromyography Signals

Due to its non-invasive nature, the Electromyography (EMG) signal has become crucial in predicting human motor intention and facilitating human-robot collaboration across various fields, including rehabilitation, assistive technologies, ergonomics, clinical diagnosis, and sport science. This study aims to enhance joint angle prediction using a two-step algorithm, leveraging a substantial dataset of twenty human subjects. The EMG data were acquired from four upper-limb muscles during an elbow flexion-extension and shoulder abduction-adduction task using Kinect V2 and Noraxon's TeleMyo Mini DTS system. In stage one, a Support Vector Machine (SVM) is used to predict joint angles directly, followed by refining predictions with a Long Short-Term Memory (LSTM) network. The impact of different sliding window sizes on algorithm performance is analyzed, and a statistical comparison between direct angle prediction and the two-stage approach is conducted to assess the proposed method's effectiveness. The proposed SVM-LSTM algorithm consistently yields enhanced Root Mean Square Error (RMSE) values (ranging from 8.45° to
$9.36^{\circ}$
) for both elbow and shoulder angles, compared to RMSE values reported in the literature (ranging from
$8.2^{\circ}$
to
$19.04^{\circ}$
). Statistical analysis through Friedman Analysis of Variance (ANOVA) underscores significant differences between SVM-direct and SVM-LSTM for both elbow and shoulder angles
$(\mathrm{p} < 0.001)$
. In the dynamic field of EMG-based joint angle prediction, these results hold profound significance for applications like prosthetics and rehabilitation robotics, where precise joint angle estimations are crucial.