Generalized Stockwell transform and SVD-based epileptic seizure detection in EEG using random forest

Abstract

PurposeVisual inspection of electroencephalogram (EEG) records by neurologist is the main diagnostic method of epilepsy but it is particularly time-consuming and expensive. Hence, it is of great significance to develop automatic seizure detection technique. MethodsIn this work, a seizure detection approach, synthesizing generalized Stockwell transform (GST), singular value decomposition (SVD) and random forest, was proposed. Utilizing GST, the raw EEG was transformed into a time–frequency matrix, then the global and local singular values were extracted by SVD from the holistic and partitioned matrices of GST, respectively. Subsequently, four local parameters were calculated from each vector of local singular values. Finally, the global singular value vectors and local parameters were respectively fed into two random forest classifiers for classification, and the final category of a testing EEG was voted based on sub-labels obtained from the trained classifiers. ResultsFour most common but challenging classification tasks of Bonn EEG database were investigated. The highest accuracies of 99.12%, 99.63%, 99.03% and 98.62% were achieved using our presented technique, respectively. ConclusionsOur proposed technique is comparable or superior to other up-to-date methods. The presented method is promising and able to handle with kinds of epileptic seizure detection tasks with satisfactory accuracy.