next up previous contents
Next: Neural Network Modeling of Up: Computational Information Processing Previous: Recurrent Self Organizing Map

Medical Signal Processing

Researchers: Jukka Heikkonen, Markus Varsta and José Millan from the Joint Research Centre of the European Commission Ispra site.

Electro encephalography (EEG) is an important clinical tool in the study, diagnosis and treatment of several neurological disorders such as epilepsy and sleeping disorders. The rapid improvement in the performance ratio of computers and storage devices have made long EEG recordings feasible but experts capable of examining these recordings spanning possibly over twelve hours are in great shortage, hence a system capable of automatic EEG evaluation to reduce the work load of the examining physician is a lucrative prospect.

In this project three different feature sets were compared using EEG recordings containing epileptic spikes and seizures. The goal is to evaluate the class separation of the sets without actually building the classifiers. First of the feature sets is Haralick's co-occurrence features that are traditionally connected with texture analysis in image processing. The second set is wavelet features that capture both spatial and frequency information. The third set of features is Fourier spectral features. This set was chosen as a standard feature set in signal processing to provide a measuring stick for the other sets. The results indicate that while with ``easy'' data all three sets performed adequately the Fourier spectral features were clearly outperformed by the other two in the case of ``harder'' data (see Fig.8), where the epileptic activity was not all as easily distinguished. However none of the sets provided satisfactory results in all of the test cases but the best performance might well be achieved with a ``team of experts'' approach, as the wavelet features and co-occurrence features are different in nature, but this still remains to be investigated.

 
Figure 8:   Epileptic activity in two EEG recordings. The activity is easily spotted in the upper signal but much harder to distinguish in the bottom signal.
Figure 8


next up previous contents
Next: Neural Network Modeling of Up: Computational Information Processing Previous: Recurrent Self Organizing Map
Juha Merimaa
1/2/1998