Fig. 1.

(A) As a user transitions from one steady-state movement pattern to another, EMG classification models can exhibit erratic prediction behavior. (B) Frame-wise classification of EMG signals, wherein movement intentions are predicted from the most recent feature-extraction window, or frame, of EMG. Frame-wise models lack a mechanism for factoring longer-term temporal information into their predictions. (C) Sequential classification of EMG signals, wherein consecutive frames provide additional temporal information to a model. (D) With ED-TCN, the convolutional encoder-decoder framework learns latent temporal patterns which can improve prediction accuracy and stability.