Objectives: In rehabilitation, hand robotics, and kinesiology, a crucial challenge is establishing a connection between surface electromyogram (sEMG) signals and the kinematics of joints and upper limbs.
Methods: The present research introduces a deep recurrent neural network regressor that uses LSTM (Long Short Term Memory) and GRU (Gated Recurrent Unit) cells, leveraging sEMG signals to estimate finger joint angles. This investigation used the DB2 series of the Ninapro dataset.
Results: Remarkably, many joint angles can be estimated with up to 97% accuracy when based on the correlation coefficient criterion. This indicates a strong positive correlation between the estimated values and the actual values of these joints. However, it is significant to note that there is a trade-off between accuracy, learning time, and the number of parameters used, which varies based on the type of cells employed in the network.
Conclusions: Using deep LSTM neural networks, it is possible to estimate the interphalangeal joint angles of the fingers with an accuracy of 97%. LSTM cells offer more precise predictions but require more learning time and parameters than GRU cells. The study also explores the significance of efficient muscles in executing movements, potentially influencing the varying estimation accuracy among different joints.