Delsys Prize 2008 Winning Proposal
Toward neural control of artificial legs: a new strategy to identify locomotion modes using EMG
Dr. Helen (He) Huang of the University of Rhode Island, Kingston, RI, USA
Innovation
A phase-dependent EMG pattern classification strategy is proposed to promptly identify the user’s locomotion mode. This
method was tested on both able-bodied subjects and subjects with long transfemoral amputations. The EMG signals from gluteal muscles and muscles in the thigh or residual limb contained sufficient neuromuscular control information for
reliable classification. This proposed approach has a great potential for design of neural-controlled, powered artificial
legs, which can further enhance the locomotion function in individuals with lower limb amputations.
Description
Recent design of powered artificial legs uses locomotion mode-based control, which has allowed leg amputees to
efficiently perform various locomotion modes such as stair ascent/descent beyond level ground walking. However, the
selection of control mode can only be achieved by using a manual or body-powered switch; it is quite cumbersome for the
users and it does not allow smooth task transition. Hence, neural control of powered artificial legs is demanded for
seamless mode transition. EMG signals recorded from the residual limb are effective neural signals for intuitive
prostheses control (Basmajian and De Luca, 1985). EMG pattern classification method has been applied to decipher user
intent for multifunctional prosthetic arm control (Englehart and Hudgins, 2003); however, no EMG-controlled lower-limb
prostheses are commercially available, and published studies in this area are limited. Providing accurate and timely
updates of the user’s locomotion modes via EMG pattern recognition is challenging. This is because the EMG signals
applied to prosthetic arm control are stationary, while the EMG signals recorded from leg muscles during ambulation are
time-varying. To correctly identify user locomotion mode, we propose a new phase-dependent EMG pattern recognition
strategy, assuming that (1) EMG recorded from leg muscles are quasi-stationary within a short time window, and (2)
muscle activation patterns for the same locomotion mode are similar at the same gait phase. The idea is to build multiple
classifiers corresponding to several short gait phases (Fig. 1a)1. While the user is walking, the phase detection module
detects gait phase and switches on the associated classifier to decide the user’s current performing mode.
To test this design, eight able-bodied subjects and two subjects with long transfemoral amputations were recruited to
perform six locomotion modes (level ground walking, stepping over an obstacle, stair ascent and descent, and turning
right and left with 90 degrees) and a standing task. For able-bodies subjects, EMG from two gluteal muscles and eight
thigh muscles were monitored; for amputee subjects, EMG recorded from two gluteal muscles and muscles in the residual
limb were measured. Four gait phase windows with 200 ms in length were defined; they were aligned with heel-contact
(HC) and toe-off (TO) (Fig. 1b). In each phase, a linear discriminant analysis (LDA)-based classifier was built by EMG
training data and evaluated by the testing data set offline1. Fig. 2 demonstrates the confusion matrix averaged over eight
able-bodied subjects. The diagonal elements denote the classification accuracies for individual modes; other elements are
classification errors between two modes. Impressively, the high risk activities of stair ascent and descent were the most
reliably identified modes for all studied phases with 93.8-99.3% classification accuracy. The classifiers confused the
turning modes and the level walking mode. Fortunately, the turning modes may be unnecessary because amputee subjects
were able to perform the turning tasks easily and safely using the prostheses with level walking control mode. In addition,
high average classification accuracy over all modes and subjects were observed in the Pre-TO (94.0%) and Pre-HC
(94.8%) phases. This is desirable to assure the safety because the controller could decide the motion of prosthetic joint
before the execution of the swing and the next weight acceptance. Perhaps the most exciting result was the comparable
classification accuracy between using EMG from the able-bodied subjects and using EMG of the amputee subjects
(Figure 3). It implies that sufficient neural control information may be extracted from EMG signals recorded from
individuals with long transfemoral amputations. The outcome in this study suggests the concept of phase-dependent EMG
pattern recognition design is promising for realizing neural selection of prosthesis control mode, which will in turn permit
leg amputees to control their artificial legs easily and intuitively.
Innovation:
A phase-dependent EMG pattern classification strategy is proposed to promptly identify the user’s locomotion mode. This
method was tested on both able-bodied subjects and subjects with long transfemoral amputations. The EMG signals from
gluteal muscles and muscles in the thigh or residual limb contained sufficient neuromuscular control information for
reliable classification. This proposed approach has a great potential for design of neural-controlled, powered artificial
legs, which can further enhance the locomotion function in individuals with lower limb amputations.
Title of Entry: Toward neural control of artificial legs: a new strategy to identify locomotion modes using EMG
 |
| |
|
|