Computational Intelligence in Analysis of Electromyography Signals - Computational Intelligence in Biomedical Engineering - page 248

TABLE 5.2
Comparison of Classification Accuracies of Myopathy, Motor Neuron
Disease, and Normal Subjects

Method Acc. (%) Feature Set

BP-NN (Pattichis et al., 1995) 80.0 MUAP time domain

NN-majority vote (Pattichis 80.0 AR, cepstrum,

et al., 1995) time domain

SOM 80.0 MUAP time domain

SVM (Xie et al., 2003) 82.40 MUAP time domain

Neurofuzzy hybrid (Christodoulou 88.58 AR, cepstrum,

and Pattichis, 1999) time domain

Hybrid SOM (Christodoulou 94.81 MUAP time domain

and Pattichis, 1999)
Statistical (Christodoulou 95.30 MUAP time domain

and Pattichis, 1999)
Hybrid SOM-LVQ (Christodoulou

97.61 MUAP time domain

and Pattichis, 1999)
Genetics-based machine learning ≤ 95 MUAP time domain

(Pattichis and Schizas, 1996)

Note : The data sets used are not the same, hence the results are just a general indication
of the potential of these techniques.

In 1990, Schizas et al. (1990) looked at the application of NN to classify
action potentials of a large group of muscles. The early results yielded accu-
racies of approximately 85% and simple physiological features of the EMG
waveforms were extracted. The supervised data were obtained from the deci-
sions of two expert neurologists using the Bonsett protocol that consisted of
a series of questions used to arrive at the final diagnosis. Their work was later
extended (Schizas et al., 1992) to comparing classification algorithms using
K-means, MLP-NN, SOMs, and genetic-based classifiers. It was discovered
that the simple K-means algorithm was not suitable as it gave the lowest clas-
sification accuracy, but both NN and genetic-based models produced promis-
ing results (Table 5.2).
Pattichis et al. (1995) extended the work by Schizas et al. by applying NN
to 880 MUAP signals collected from the biceps brachii muscle. The data con-
sisted of 14 normal subjects, 16 patients suffering from motor neuron disease,
and another 14 patients suffering from myopathy of which 24 were randomly
selected to form the training set and the remainder formed the test set. They
compared the MLP network against the K-means clustering and Kohonen's
SOM. The best performance using a K-means clustering algorithm using 3
cluster means was 86% accuracy on the training set and 80% on the test set.
The MLP network fared better giving 100% accuracy on the training data
but required a fair amount of tuning for the network gain η and momentum
α coecients. The best model had a hidden layer with 40 neurons providing
a test accuracy of only 80-85%. The problem then was that short of trial