| Autor: |
F. M. Azevedo, Geovani R. Scolaro, Christine Fredel Boos, M. C. V. Pereira |
| Rok vydání: |
2013 |
| Předmět: |
|
| Zdroj: |
IFMBE Proceedings ISBN: 9783642293047 |
| DOI: |
10.1007/978-3-642-29305-4_146 |
| Popis: |
Artificial Neural Networks (ANN) are a common tool for pattern recognition in applications using bioelectrical signals. The different ways that these signals are presented as input stimuli of the neural networks have a direct influence on their performance. Literature shows that good results are usually obtained when the characteristics of the analyzed signal are highlighted and presented in patterns that are easily assimilated by the network. The pre-processing of the input signal is a technique that can be used to highlight these features. In this study we evaluated various types of mathematical methods that are usually applied in the detection of epileptiform events in electroencephalogram (EEG) signals and used them to pre-process the network’s input stimuli. We analyzed the frequency spectrum obtained by Fast Fourier Transform, the spectrogram derived from the Short-Time Fourier Transform, the features extracted from Wavelet Transform decompositions and morphological and statistical features – that in this study will be called morphological descriptors – extracted from the signals. Several neural networks, for each pre-processing method, were implemented using a similar topology to enable an assessment and comparison between their results. The best networks for each method achieved sensitivity results in the range of 73 to 97%, specificity between 63 and 96%, and efficiency between 81 and 93%. After analyzing the results we could determine that the best method for pre-processing the ANN input stimuli, among those analyzed in our study is the feature extraction by morphological descriptors. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
