Delsys Prize 2007 Winning Proposal
Decoding a new neural-machine interface for control of artificial limbs
Dr.Ping Zhou, Rehabilitation Institute of Chicago, Northwestern University, Chicago, IL, USA.
A new neural-machine interface, termed targeted muscle reinnervation (TMR), has been developed to improve the function of upper limb prostheses. The control information contained in TMR was assessed. Our results demonstrate that TMR can provide a rich source of motor control information and this information in turn promises to dramatically improve artificial arm function for people with proximal arm amputations.
Improving the control and function of artificial arms remains a great challenge, especially for proximal amputations where the disability is greatest. A new neural-machine interface, termed targeted muscle reinnervation (TMR), has been developed to improve the function of upper limb prostheses, especially for those with high level amputations (Kuiken et al. 2004). TMR involves transferring the residual amputated nerves to non-functional muscles in amputees (Figure 1). The reinnervated muscles then act as a biological amplifier of motor commands in the amputated nerve. TMR thus provides physiologically appropriate surface electromyogram (EMG) control signals that are related to functions in the lost arm. While initial clinical success with TMR has been promising (using simple myoelectric control paradigms solely based on amplitude measurement of the EMG signal), the number of degrees of freedom of the robotic arm that can be controlled has been limited to date by the number of reinnervated muscle sites (Kuiken et al. 2004; Kuiken et al. 2007). An analysis of the motor control information content made available with TMR is then performed in this study using advanced EMG signal processing and classifying techniques.
The high density EMG experiments (Figure 2) were performed on 4 TMR patients (from 7 to 52 months after the surgery): a man with a shoulder disarticulation amputation (3 nerve transfers), a woman with a very short transhumeral amputation (4 nerve transfers) and two men with long transhumeral amputations (2 nerve transfers) (Zhou et al. 2007).
The subjects were asked to imagine and actuate 16 different movements involving the amputated limb (see Table 1 for a
list of these movements). The subjects were instructed to exert a comfortable level of contraction at a medium force which
was held for approximately 4-5 s. Ten repetitions were performed for each movement. A series of pattern recognition
analyses were performed on 256 ms data windows using a time domain (TD) feature set, and a feature set of
autoregressive (AR) coefficients in combination with the signal’s root mean square (RMS) amplitude. A linear
discriminant analysis (LDA) classifier was used for classification of different movements. The first half of the active data
was used to train the classifier; the second half of the active data was used as a test set to evaluate the classifier’s
accuracy. The classification analysis was conducted using a monopolar electrode configuration and three bipolar configurations in transversal, longitudinal and diagonal directions, respectively.
Classification accuracy for the 16 intended movements was high using all of the surface EMG recordings from the reinnervated muscles. Table 1 shows class-to-class results from a typical experiment; examination of the specific
movements revealed only a few movements that had accuracies below or equal 95%. Across all the subjects, with the
monopolar channels the average overall classification accuracy was 90.5±6.3% for TD feature sets and 90.0±7.3% for
AR+RMS feature sets. Spatial filtering consistently improved the accuracy of classification to an average of 96.0±3.9%
with TD and 95.0±5.2% with AR+RMS features, using bipolar electrode configurations. Analyses of the required number
of EMG channels and computational demands demonstrate clinical feasibility of these methods. The channel reduction
analysis indicated that only 5-10 optimally placed bipolar electrodes were required to maintain 90% of the maximum
accuracy. The pattern recognition methods employed in this investigation are computationally efficient; a speed benchmark was described previously in which a 1-GHz Pentium III based workstation required roughly 4 ms to process
each EMG channel. Improvements in coding efficiency and processor speed currently place this delay at less than 0.25
ms per channel, resulting in roughly 4 ms delay for 16 EMG channels, which is adequate for real-time control. This
indicates that TMR combined with pattern recognition techniques has the potential to further improve the function of
Kuiken T, Dumanian G, Lipschutz R, Miller L, and Stubblefield K (2004) The use of targeted muscle reinnervation for
improved myoelectric prosthesis control in a bilateral shoulder disarticulation amputee, Prosthet Orthot Int. 28: 245-253.
Kuiken T, Miller L, Lipschutz R, Lock B, Stubblefield K, Marasco P, Zhou P, and Dumanian G (2007) Targeted
reinnervation for enhanced prosthetic arm function in a woman with a proximal amputation. Lancet 369: 371-380.
Zhou P, Lowery M, Englehart K, Huang H, Li G, Hargrove L, Dewald J, and Kuiken T, Decoding a new neural-machine
interface for control of artificial limbs, J Neurophysiol. (Accepted for publication).
History of Dissemination: This study was accepted for publication by Journal of Neurophysiology on August 29, 2007.