Page 216 - Handbook of Biomechatronics

Upper-Limb Prosthetic Devices 213

2.3 Sensing Many-DoFs
Nowadays, the sensing of many EMGs is done via superficial EMG elec-
trodes or intramuscular EMG electrodes as described later in Section 2.6.
The following can apply for both.

2.3.1 Artificial Intelligence, Neural Networks, and Pattern Recognition
Pattern recognition for myoelectric control was first introduced (Herberts
et al., 1973).
The University of New Brunswick has reinvented and championed since
1990s the pattern recognition paradigm for upper-limb multifunction pros-
thesis control (Hudgins et al., 1993) by showing that the myoelectric signal is
deterministic at the onset of the contraction and that an artificial neural
network can be used to classify patterns of the myoelectric signals
(Hudgins et al., 1993).
If the physiological musculature exists, then its EMG can be used to
control the prosthetic device of an upper-limb amputee, for example,
biceps and triceps pair to control a prosthetic elbow (Scheme and
Englehart, 2011), leading to intuitive control and low mental loading
which are both desirable attributes for prosthesis control (Childress,
1992). In the case that the musculature does not exist, then logical substi-
tutions should be used, for example, using wrist flexor and extensor pair to
control a prosthetic hand. The above one muscle pair to one DoF schema
becomes impractical when applied for the control of a multi-DoF upper-
limb prosthesis, due to co-activation patterns of muscles, deep common
muscle activation, and EMG crosstalk (Scheme and Englehart, 2011).
One of the proposed solutions of the above problem is the tool of pattern
recognition. The idea is simple: a set of myoelectric signals is recorded, for
which different features are extracted (amplitude, zero crossings, etc.).
During training a set of basic hand postures is performed and their associ-
ated features pattern is associated and a classifier is trained. During a real-
time hand movement, a classifier algorithm will try to classify the real-time
features set to one of the basic classes which is associated to a particular
hand posture (from training).
A complete review of the pattern recognition techniques for upper-
limp prostheses shows that current pattern recognition success rates are
90%–95%, in a controlled laboratory environment (Scheme and
Englehart, 2011). That is, open the hand 9 out of 10 times but one time
statistically do not open it. This is not clinically acceptable. On top of this