A seizure warning system based on EEG signals is developped, which can be used to 'warn' the medical team that a seizure is coming by detecting it at its onset.(beginning of the seizure). The system is based on a patient specific classifier. The original EEG data are preprocessed in two steps. First a band-pass filter is applied to suppress non-seizure related activity and secondly the half wave decomposition algorithm(a data reduction method) is applied to the filtered signal. Subsequently, features in both time and frequency domain are extracted. Using the latter features and training data a classifier is built. The classifier is a modified nearest neighbor algorithm, which based on a similarity degree between a new, possible seizure and the training data obtained above, classifies the new EEG as seizure or non-seizure. The programs were carried out in MATLAB. In this talk we will present the classifier and comment on its performance and possible future work as well as alternative strategies that can improve it.
Nearest Neighbor Classifier, Feature extraction, Decomposition in Halfwaves, Band pass filtering, Template Matching, Channels
Hao Qu and Jean Gotman. (1997)
Patient-Specific Algorithm for the Detection of Seizure Onset in Long-Term EEG Monitoring : Possible use as a Warning Device,
IEEE Transactions on Biomedical Engineering, No 2,.