Computational Intelligence in Analysis of Electromyography Signals - Computational Intelligence in Biomedical Engineering - page 255

TABLE 5.3
Classification Accuracy (Acc.) for Limb Motion Used in Prosthesis Control

Classifier Year Acc. (%) Remark

Nonlinear discriminant 1982 99 Long training times

(Graupe et al., 1982) limited by computer

technology then.
AR parameters used.

Bayes classifier 1984 91 Data collected from

(Lee and Grimson, immobilized limb

2002) measurements. Zero crossing

and variance features used.

Neural network 1995 93.7 Neural network model

(Englehart et al., with AR features.

1995)
Fuzzy systems

2000 91.2 Fuzzy inputs based on

(Chan et al., 2000) basic isodata.

Fuzzy clustering 2003 98.3 Requires two step processing,

with neural network finding fuzzy clusters, and

(Karlik et al., 2003) training neural network.

Recently, there has been some work toward developing a prosthesis for
spinal cord-injured patients (Au and Kirsch, 2000; Hincapie et al., 2004).
An adaptive NN (Hincapie et al., 2004) was used to control a neuroprosthesis
for patients with C5/C6 spinal cord injury. The initial experiment employed
healthy subjects to obtain kinematic data for several upper-limb muscles such
as the upper trapezius, anterior deltoid, infra spinatus, and biceps. The EMG
information was then fed into a finite element neuromuscular model, which
simulated the actions of spinal cord-injured patients. The model performed an
inverse dynamic simulation to transform the healthy data into data that mim-
icked spinal cord-injured patients. This was then used to train a time-delayed
neural network (TDNN) to predict the muscle activations and subsequently
produced a low RMS error of 0.0264. Hussein and Granat (2002) proposed
a neurofuzzy EMG classifier to detect the intention of paraplegics to sit or
stand. Their work applied the Gabor matching pursuit (GMP) algorithm to
model the EMG waveform. A genetic algorithm was applied to increase the
GMP algorithm speed whereas a neurofuzzy model was used to classify the
intention of the patient to sit or stand. Results from 30 tests yielded high-
classification accuracies with only one or two false-negative classifications and
no false positives. False-positive classifications were when the subject per-
formed an action, which they did not intend, whereas false negatives were not
performing an intended action. In this situation, false positives would be more
hazardous and highly undesirable in the design of myoelectric prostheses.
Huang et al. (2003) proposed an intelligent system to automatically evalu-
ate the EMG features for control of a prosthetic hand. The system was built
on a supervised feature-mining method, which used genetic algorithms, fuzzy