Information Technology Reference
In-Depth Information
3.2 ANN Results
In the supervised classifications, available EEGs were distributed in two sets,
training and testing sets. The size of the training dataset was 20 EEGs (10 con-
trols and 10 uremic patients). In order to define the most significant parameters,
different combinations of inputs have been used (from one unique band to all
available data) for the training of the ANN. The rest of the EEG records were
used for testing the network (7 uremic patients and 8 controls).
The classification results were not satisfactory when only one frequency band
was selected for input; in those cases the classification accuracy was never higher
than 60%. Using the whole of 55 input parameters, the accuracy was 80% (2
errors in uremic patients and only 1 error in controls). The highest classification
accuracy was obtained when eliminating the MDF parameters. In that case, the
ANN correctly clasified 86% of the cases (2 errors in uremic and no errors in
controls).
These results show that no abnormalities can be found using just one fre-
quency band of the EEG or any isolated parameter. The need of combining
bands and parameters demonstrates that there are some complex correlations
between bands in uremic patients, and that those patterns can be detected by
the ANN. As final conclusions we can state that brain abnormalities in uremic
patients are complex and should not be studied using just one parameter, and
that the pattern of manifestations is not always the same in different patients.
4 Conclusions
In the present study, an automatic system for the classification of CRF patients
employing ANN techniques was developed and implemented. The classification
was performed using features extracted from EEGs. Emphasis was placed on
selection of the characteristic features and for the accurate extraction of these
features.
From the results, it can be seen that the RBF network provided an excel-
lent performance for the studied application, classifying accurately a 100% of
controls and a 86.6% of CRF patients. The performance of the system can be
further enhanced by training with a larger number of training inputs, which
would increase the network ability to classify unknown signals. The system per-
formance can be also improved by considering other features of EEG that were
not included in the current system. Thus, the present study shows the feasibility
of the application of RBF based ANN for the classification of CRF patients.
References
1. Gabor, A.J., Seyal, M.: Automated interictal EEG spike detection using artifi-
cial neural networks. Electroencephalography and Clinical Neurophysiology 83(5),
271-280 (1992), ISSN: 0013-4694
Search WWH ::
Custom Search