Estimating finger joint angles on surface EMG using Manifold Learning and Long Short-Term Memory with Attention mechanism

Abstract The success criteria of electromyography (EMG) rely on recognizing the pattern and correlating it to its target, especially in the regression task, which has become a popular topic in this decade because of its flexibility in accessing targets compared to classification tasks. Therefore, it will be beneficial when the methods used in stroke therapy devices require wide access to estimate flexion/extensions in the degree of freedom of fingers. Previous researchers have made many efforts to achieve accuracy and performance results in this regression task. Therefore, many variations of the regression method were found. However, the existed method combination brings drawbacks because the signals containing sparse samples and noise could make the regressor overfitting. For encountering this problem, this work deploys a Manifold Learning to handle sparse data and Long Short-Term Memory (LSTM) based attention mechanism as a regressor, distinguishing between time series data with different importance by allocating weights of Manifold Learning results. Besides applying those methods, this work compares with various combination techniques such as different preprocessing signals, feature extraction, dimensional reduction, and regressor that are widely used to determine the proposed investigated contribution various parameters applied according to a different combination of methods. The results show that the proposed methods outperform performance by getting 0.957 in the R-Square score. In addition, Manifold Learning as dimensional reduction contributes to the improvement of the performance in various parameters such as different feature domains and regressors.