Neonatal EEG seizure detection using a time-frequency matched filter with a reduced template set

Electroencephalographic (EEG) recordings are an important diagnostic resource in determining the presence or absence of clinical seizures in neonates. These nonstationary signals require some form of nonstationary analysis to detect seizures in the EEG data. A time-frequency (TF) matched filter has been previously proposed to detect seizures in both adult and newborn EEG. A method which constructs a reference or template set from a feature of EEG seizures, rather than the whole EEG seizure, displayed the most promising results. However this method suffered from an inability to adequately represent patient variability in the template set while simultaneously maintaining a low false detection rate. A new method of the TF matched filter is proposed that halves the template set required by approximating the templates with a more general ambiguity domain function representation. This proposed method is also less sensitive to false detections when a larger reference set is used, as evidenced by the findings on both simulated and real neonatal EEG.