Quadratic time-frequency distribution selection for seizure detection in the newborn

Several, recently proposed, newborn EEG seizure detection techniques use quadratic time-frequency distributions (QTFDs) to generate the time-frequency representations (TFRs) at their core. The specific type of QTFD that provides the best discrimination between the TFR of nonseizure and seizure epochs of EEG, however, has yet to be thoroughly investigated. This paper proposes the selection of an optimal QTFD that maximises the the absolute error between seizure and nonseizure QTFDs calculated on a database of newborn EEG. The optimisation procedure is a data driven process that selects the optimal QTFD based on the distribution of the absolute error between nonseizure/nonseizure QTFDs and the seizure/nonseizure QTFDs. Several non-adaptive QTFDs were selected for comparison and those selected were subjected to a restriction on the kernel's volume to ensure that the QTFD can accurately represent the time-frequency distribution of signal energy. The results show that a lag independent or narrowband QTFD such as the modified B distribution provides a QTFD that best highlights the difference in time-frequency signal energy between newborn EEG seizure and nonseizure.