Abstract
Reliable detection of the onset of epileptic seizures has seen renewed interest over the past few years, owing to several factors including, the global push toward digital health-care, the advancements in signal processing techniques, and the increased computational power of machines. A reliable automatic system could result in tremendous improvement in the quality of life of epilepsy patients. This paper presents dynamic mode decomposition (DMD), a data-driven dimensionality reduction technique, originally used in fluid mechanics, as an instrument for epileptic seizure detection from scalp electroencephalograph (EEG) data. DMD is employed in this paper to measure power of signals in different frequency bands. These subband-powers, along with signal curve lengths, are used as features for training random under-sampling boost decision-tree classifier. Post-processing measures ensure an acceptable balance between false positives and true positives. The proposed algorithm has been tested over a thousand hours of EEG data from two different data sets, the CHB-MIT data set and the KU Leuven data set, giving sensitivity values of 0.87 and 0.88, respectively, and specificity values of 0.99 for both the data sets.