Biomedical Engineering Reference
In-Depth Information
and error; no other way was known to select the appropriate MLP model.
Further investigation revealed that the same pathological cases were always
misclassified although different MLP models were used. These were subjects
that suffered from a typical motor neuron disease with bulbar involvement
and no serious involvement in the biceps brachii. Another commonly misclas-
sified subject suffered from polymyositis myopathy. This suggested that the
disorders possess finer EMG characteristics, which could not be easily differ-
entiated from normal subjects. Similar results were obtained when applying
Kohonen's SOM with 91% accuracy on the training set but again only 80-85%
on the test set. The results on genetics-based machine learning (GML) were
reported separately (Pattichis and Schizas, 1996) where the same data was
classified using GML models derived by genetic learning. Better classification
was achieved with less evolution indicating that frequent changes in the clas-
sifier state should be avoided. It was argued that the GML method yielded
simple implementation models, were easier to train, and provided quicker clas-
sification due to parallel-rule implementation. This is, however, questionable
as the number of classifiers from the evolution algorithm numbered greater
than 500 on average. Comparisons with unsupervised learning methods was
undertaken in 1999 by Pattichis and Elia (1999) using SOM, SOM with learn-
ing vector quantization (LVQ), and statistical methods. In this experiment,
1213 MUAPs were analyzed and classified using a proposed EMG decomposi-
tion and classification system. The EMG signal was signal thresholded using
average-signal and peak-signal values and located using a 120 bit window cor-
responding to a sampling frequency of 6 ms or 20 kHz. The hybrid SOM-LVQ
method was found to produce the highest classification accuracy of all the
methods used by Schizas and Pattichis to classify between myopathies and
motor neuron diseases.
NNs have also been used as decision-making tools for online classifica-
tion of various neuromuscular diseases (Bodruzzaman et al., 1992). The main
challenge was to design an automatic classification system based on signal-
processing techniques such as AR modeling, short-time Fourier transform,
the Wigner-Ville distribution, and chaos analysis to define abnormalities in
diseases such as neuropathy and myopathy. It was often found that features
derived from these different methods were nonseparable and, hence, dicult
to classify using a single method. This was demonstrated by the probability
density function of the features, which often overlapped. Ten features were
extracted from the processing methods described, and feature selection was
used with NN to provide at least 95% classification accuracy.
5.6.2 Support Vector Machines
The SVM classifier has also been applied to EMG signals for diagnosis of neu-
romuscular disorders. Xie et al. (2003) trained a multiclass SVM classifier to
distinguish between healthy subjects and patients suffering from motor neu-
ron disease or myopathies. They extracted the physical features such as EMG
spike amplitude, number of turns, and spike duration. The SVM classifier was
Search WWH ::
Custom Search