Efficient frameworks for statistical seizure detection and prediction

This paper presents two efficient frameworks for seizure detection and prediction that depend on statistical analysis. The common thread between them is the selection of certain attributes extracted from the electroencephalography (EEG) signals and the derivation of probability density functions (PDFs) of these attributes in two different types of activities of EEG signals. The first framework is for seizure detection based on scale-invariant feature transform (SIFT). Its idea is to choose some segments for normal and seizure activities. These segments are transformed to 2D matrices to be treated with the well-known SIFT. The feature key-points are extracted from those 2D matrices. The parameter that is used to discriminate between normal and seizure activities is the number of key-points. The PDFs of the number of key-points in cases of normal and seizure activities are estimated and a threshold value is used to classify any new segment as either a seizure or a normal segment. The other framework is for seizure prediction. The discrimination in this case is between normal and pre-ictal activities. A statistical treatment is performed on five attributes extracted from the wavelet transforms of different EEG segments for normal and pre-ictal activities. These attributes are amplitude, local mean, local variance, median and derivative. All PDFs of these attributes are estimated for normal and pre-ictal activities. Thresholds are set for all attributes. Decisions are taken based on these thresholds. Finally, a majority-voting strategy is used to merge decisions taken for all attributes.