Computational Intelligence in Cardiology and Heart Disease Diagnosis - Computational Intelligence in Biomedical Engineering - page 196

the diseased, which is generally not desirable. Later, Zong and Jiang (1998)
showed that a better selection of fuzzy linguistic variables and decision rules
could be used to classify five different ECG beat types over 20 arrhythmias
obtained from the popular MIT-BIH database. The fuzzy variables included
premature beats, QRS widths, and beat times with the classification deci-
sion achieved by selecting a maximum likelihood method with a threshold
level of 0.5. Good prediction accuracies of up to 99.77% were achieved for
beat types such as supraventricular premature beats but results for VEBs
(63.01%) were low.
Silipo et al. (1999) investigated the relevance of ECG features in their design
of an arrhythmia classifier and argued that if the number of features is high,
the interpretability or significance of the corresponding fuzzy rules may be
overshadowed by the number of rules required. To overcome this, the feature
set should be studied a posteriori to select the best feature set containing
the most useful features. They then applied the information gain measure
for a fuzzy model on 14 different physical ECG features obtained from ECG
samples in the MIT-BIH database. It was discovered that the QRS width was
the most useful in the discrimination of a three-class arrhythmia problem as
compared to ST-segment levels or the negative of the QRS area.
Ifeachor et al. (2001) described fuzzy methods for handling imprecision and
uncertainty in computer-based analysis of fetal heart rate patterns and ECG
waveshapes during childbirth. CI models, based on fuzzy logic techniques,
that explicitly handle the imprecision and uncertainty inherent in the data
obtained during childbirth were proposed. The ability to handle imprecision
and uncertainty in clinical data is critical to removing a key obstacle in elec-
tronic fetal monitoring.
The use of fuzzy logic has also been reported in the design of a knowledge-
based approach to the classification of arrhythmias. Kundu et al. (1998) have
put forward a knowledge-based system to recognize arrhythmia based on the
QRS complex, P waveshape, and several other physical features. Several fuzzy
rules based on fuzzification of these features were implemented, for example,
If average heart rate of the ECG cycle is high , then diagnosis is
sinus tachycardia .
It was argued that the fuzzy classification increased the granularity of the
diagnosis as opposed to the hard classification methods. In addition, it was
argued that this method could deal with incomplete knowledge of the input
data, however, experiments on 18 datasets demonstrated only limited improve-
ments in diagnostic accuracy. Beat recognition was also pursued by Osowski
and Tran (2001), who applied an FNN model on statistical ECG features
shown in Figure 4.33. In their design, the second- to fourth-order cumulants
were extracted from the ECG signal because these features were less sensitive
of minor variations in waveform morphology. These cumulants were calculated
for the QRS segment of the ECG signal. They reported a decreased error rate
of 3.94% over the classification of seven ECG beat types.