Background: Epileptic seizures evolve through several states, and in the process the brain signals may change dramatically. Signals from different states share similar features, making it difficult to distinguish them from a time series; the goal of this work is to build a classifier capable of identifying seizure states based on time-frequency features taken from short signal segments.
Methods: There are different amounts of frequency components within each Time-Frequency window for each seizure state, referred to as the Gabor atom density. Taking short signal segments from the different states and decomposing them into their atoms, the present paper suggests that is possible to identify each seizure state based on the Gabor atom density. The brain signals used in this work were taken form a database of intracranial recorded seizures from the Kindling model.
Results: The findings suggest that short signal segments have enough information to be used to derive a classifier able to identify the seizure states with reasonable confidence, particularly when used with seizures from the same subject. Achieving average sensitivity values between 0.82 and 0.97, and area under the curve values between 0.5 and 0.9.
Conclusions: The experimental results suggest that seizure states can be revealed by the Gabor atom density; and combining this feature with the epoch's energy produces an improved classifier. These results are comparable with the recently published on state identification. In addition, considering that the order of seizure states is unlikely to change, these results are promising for automatic seizure state classification.
Keywords: ECoG; Epilepsy; Gabor atoms density; Kindling model; Matching pursuit; Seizure states.
Copyright © 2014 Elsevier B.V. All rights reserved.