An intuitive muscle-computer interface using ultrasound sensing and Markovian state transitions

In recent work regarding gesture recognition and muscle computer interfaces, ultrasound-based sensing strategies have been demonstrated as a viable alternative to the pervasive surface electromyography (sEMG) modality. However, in order to facilitate switching between available gestures, both sEMG and ultrasound-based strategies have traditionally relied on unintuitive control mechanisms. The most common among these are: requiring the users to return to rest as an intermediary state between motions; mode switching through co-contraction or other ad-hoc user input; and switching based on muscle activations that are functionally unrelated to the desired motion. The unintuitive nature of such control has historically led to increased user frustration, and is often cited a major reason for device abandonment in the prosthetic control setting. In this work, we propose using an approach inspired by Hidden Markov Models (HMMs) with a novel continuous gesture recognition mechanism, for ultrasound-based sensing. We empirically calculate the average classification accuracy of our novel method during non-transitionary periods to be 99%. We then demonstrate that including predictions made during transition periods reduces this value to 69% Finally, by encoding the temporal dependency of the system within a Hidden Markov Model framework, we show that we can reduce the error caused by the instability of predictions during transitions, measured as the normalized Levenshtein distance from the true ordering, by approximately 98.8%.