Biomedical Engineering Reference
In-Depth Information
TABLE 4.2
Performance of Various Neural Networks-Based ECG Beat Recognition
Systems
Number of
Classifier Type
Beat Types
Accuracy (%)
MLP1 (Hu et al., 1993)
2
90.0
MLP1 (Hu et al., 1993)
13
84.5
MLP2 (Izeboudjen and Farah, 1998)
12
90.0
MLP-Fourier (Minami et al., 1999)
3
98.0
SOM-LVD (Yu et al., 1997)
4
92.2
MLP-LVQ (Oien et al., 1996)
2
96.8
Hybrid Fuzz NN (Osowski and Tran, 2001)
7
96.06
1997), and a summary of the performance of some successful beat recognition
system is provided in Table 4.2.
The success of NNs attracted further research in applications such as detec-
tion of MIs where in several reported studies, the ST-T segment of the ECG
waveform yielded substantially discriminative information for the recognition
of MI and ischemic episodes based on large amounts of ECG data. Edenbrandt
et al. (1993a,b) studied two thousand ST-T segments from a 12-lead ECG and
visually classified them into seven groups. These constituted labeled data,
which were then divided into a training set and a test set where computer-
measured ST-T data for each element in the training set with the correspond-
ing label formed inputs for training various configurations of NNs. It was found
that the networks correctly classified 90-95% of the individual ST-T segments
in the test set. However, it should be noted that the success of NNs depends
largely on the size and composition of the training set making it dicult to
design a generalized system. Edenbrandt and coworkers also note that careful
incorporation of NNs with a conventional ECG interpretation program could
prove useful for an automated ECG diagnosis system if the aforementioned
factors could be dealt with consistently.
In more specific work, Maglaveras et al. (1998) examined the European
ST-T database using an NN-based algorithm to detect ST-segment elevations
or depressions, which could signify ischemic episodes. The performance of
their system was measured in terms of beat-by-beat ischemia detection and
the number of ischemic episodes. The algorithm used to train the NN was an
adaptive BP algorithm and reported to be 10 times faster than the classical BP
algorithm. The resulting NN was capable of detecting ischemia independent
of the lead used and gave an average ischemia episode detection sensitivity
of 88.62%, whereas the ischemia duration sensitivity was 72.22%. ST-T seg-
ment information was also used by Yang et al. (1994) to detect inferior MI. A
total of 592 clinically validated subjects, including 208 with inferior MI, 300
normal subjects, and 84 left ventricular hypertrophy cases, were used in their
study. From this, a total of 200 ECGs (100 from patients with inferior MI
Search WWH ::
Custom Search