Biomedical Engineering Reference
In-Depth Information
ECG
waveform
preprocessing
Feature
extraction
SVM
classification
FIGURE 4.31
Classification of ECG waveforms using an SVM-based system.
This method of obtaining features was shown to exceed features derived from
heart rate, QRS width, or coecients of the correlation waveform analysis.
The SVM classifier was first tuned using a bootstrap resampling procedure
to select proper SVM parameters and subsequently managed to achieve 96%
specificity and 87% sensitivity, but no comparisons were made against NNs.
Incremental learning was then used to establish a trade-off between popula-
tion data and patient-specific data. The authors argued that many machine-
learning algorithms build a black box statistical model for the differential
diagnosis without knowledge of the underlying problem. The SVM, however,
can be used to extract information from this black box model via support vec-
tors representing the critical samples for the classification task. Rojo-Alvarez
et al. (2002b) proposed a SVM-oriented analyses for the design of two new
differential diagnosis algorithms based on the ventricular EGM onset criterion.
Millet-Roig et al. (2002) applied SVMs on wavelet features extracted from
a mixture of MIT-BIH, AHA, and data from the Hospital Clinico de Valencia
in Spain. A holdout method was used where 70% of the dataset was applied to
training the SVM, whereas 30% was reserved for testing. The SVM classifier
was compared against an NN and logistic regression that statistically fitted
an equation based on the probability of a sample belonging to a class Y.
It was found that the SVM performance was comparable to the NN and
outperformed it in one case, whereas it was significantly superior than the
logistic regression method.
SVMs have also been used in combination with genetic algorithms for clas-
sification of unstable angina. Sepulveda-Sanchis et al. (2002) employed genetic
algorithms to select proper features from ECG waveforms collected by four
hospitals from 729 patients. These patients fulfilled several preselection cri-
teria such as having no history of MI but suffering from unstable angina. It
was found that the Gaussian kernel worked best for their data achieving a
successful prediction rate of 79.12% with a low false-positive rate.
SVMs have also been used in the design of rate-based arrhythmia-recognition
algorithms in implantable cardioverter-defibrillators. In the past, these algo-
rithms were of limited reliability in clinical situations and achieved improved
performance only with the inclusion of the morphological features of endo-
cardial ECG. Strauss et al. (2001) applied a coupled signal-adapted wavelet
together with an SVM arrhythmia detection scheme. During electrophysio-
logical examination data, segments were first recorded during normal sinus
rhythm (NSR) and ventricular tachycardia (VT), following which consecutive
Search WWH ::
Custom Search