Fig. 1

Outline of the experiments. HD-sEMG recordings were processed (root mean square, data windowing with overlap) and used as inputs for classification/regression to estimate the level of activation of individual fingers during flexion (F) and extension (E) movements. Two machine-learning approaches for myoelectric control, a standard benchmark (LDA) and a recently presented novel method (CSP-PE), as well as direct control via simple thresholding (THR) were assessed in the context of selective finger control. Both offline and online tests were performed. In offline tests, isometric forces of individual fingers were measured and predicted by applying the above-mentioned methods. During the online tests, the task for the subjects was to track the reference trajectories specifying the desired individual finger activation levels assessed using EMG normalized to maximum voluntary contraction. To this aim, the subjects controlled a visual marker, which was moving according to the finger activation levels predicted online using the selected estimation method