Accurate Prediction of Knee Joint Angles Using a Hybrid CNN-LSTM-Attention Network from Surface Electromyography

In various fields such as robotics, rehabilitation, biomechanics, human-machine interfaces, and clinical research, human motion prediction has extensive applications. A hybrid networks model based on Convolutional Neural Network-Long Short-Term Memory-Attention (CNN-LSTM-Attention) was employed in this paper to extract features from the high-dimensional space and time series of surface electromyography (sEMG) data for continuous prediction of knee angles. Five able-bodied subjects participated in the study, walking at a self-selected pace while six surface electromyography (sEMG) signals from muscles involved in knee flexion and extension, along with knee angles, were synchronously recorded. The performance was assessed using the root mean square error (RMSE), Pearson Correlation Coefficient <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(r)$</tex>, and coefficient of determination <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(R^{2})$</tex> between estimated and measured knee joint angles. Compared to the CNN, LSTM, and CNN-LSTM algorithms, the adopted method estimated knee angles more accurately, achieving an average RMSE, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\rho$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.09\pm 0.90,\ 0.9911 \pm 0.0052$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.9817\pm 0.0108$</tex>, respectively. As the advance time increased, the accuracy of knee joint angle estimation decreased, particularly when the advance time reached 100 ms, at which point the RMSE is 4.83 degress. In addition, the computing time of the prediction model was less than 1 ms, which was shorter than the advance time 100 ms it could predict, thereby facilitating real-time computation in practical environments. Compared to previous studies, the proposed method excelled at extracting features directly from raw signals, avoiding the need for handcrafted feature extraction. Consequently, the model effectively achieved continuous and accurate predictions of joint motion angles and demonstrated potential to reduce latency in the human-machine interface.