Classification of newborn seizure EEG time-frequency signatures

Background: The newborn EEG presents distinct signatures in the joint time-frequency domain during seizure events. This study is a preliminary step towards classification of newborn EEG seizure timefrequency signatures for future correlation with clinical seizure signatures. Automatic classification of seizure signatures will aid in the development and administration of new anti-epileptic drugs. Methods: A T-class time-frequency distribution, with hyperbolic timeonly kernel, was applied to 191 newborn EEG seizure epochs from 12 neonate patients. A method for extracting the instantaneous frequency (IF) using image processing techniques, (developed by the authors),was then applied. The extracted IFswere then classified using a large-scale optimization algorithm, which is a subspace trust region method, based on the interior-reflective Newton method. Results: The classes of time-frequency signatures observed from this study included piecewise linear, sinusoidal and hyperbolic frequency modulated components. It was observed that newborn EEG seizures contained one signature (monocomponent) in 52.8% of seizure epochs, while 47.2% contained multiple signatures (multicomponent). A large proportion of seizures, 72.8%, contained the piecewise linear frequency modulated signatures. 44.5% of seizures exhibited a sinusoidal frequency modulated signature, while only 4.7% contained a hyperbolic frequency modulated signature. Conclusions: This study shows that newborn EEG seizure can be classified using its time-frequency signatures. It also demonstrates that automatic classification of newborn EEG seizure is achievable and can work in parallel with an automatic seizure detection algorithm for improved administration of anti-epileptic drugs.