Forearm Motion and Hand Grasp Prediction Based on Target Muscle Bioimpedance for Human-Machine Interaction

This paper introduces a novel methodology for simultaneously predicting hand grasp and forearm motion using target muscle bioimpedance measurements and regression models. A total of six channels, formed by nine electrodes, are employed for this multi-degree of freedom (DoF) prediction. Given the time-dependent nature of bioimpedance variation, the long short-term memory (LSTM) regression model is more competent in multi-DoF prediction, compared to linear regression (LR), support vector regression (SVR) and multilayer perceptron (MLP). In intra-subject cross-validation, MLP yields an average coefficient of determination (R2) of 0.9256 for predicting hand grasping angle, while LSTM achieves an average R2 of 0.9483 for predicting random simultaneous forearm two-DoF motion. Operation by amputees without the need to train the regression models is possible by mapping muscle bioimpedance variation directly to the prediction angle, allowing for the approximate estimation of single-DoF motion. The efficacy of these prediction approaches is demonstrated in a real-time object grasping task.