Automatic seizure detection in newborns and infants

Abstract

Current used methods of seizure detection were not designed to detect seizures in newborns and infants. Since seizure patterns in this age group art very different from patterns in older subjects, the standard method performs poorly. The authors present a method aimed at detecting these specific patterns, which an often very focal, slow and of particularly gradual onset. The EEG of each channel is broken down into overlapping 10-second epochs, for which the spectrum is computed. Features, including frequency, amplitude and width of the dominant spectral peak, are compared to a constantly updated background. The method was evaluated on 10 recordings from the Montreal Children's Hospital (average age 32 days) and 9 recordings from the Texas Children's Hospital (average age 5 days). Recording lasted an average of 5.5 hours and there was an average of 8.2 seizures/recording. Analysis by the existing "adult" method yielded an average detection rate of only 28% and an average of 1.9 false detections/hour. The new method yielded an average detection rate of 74% and an average of 2.4 false detections/hour. Although the false detection rate should be reduced the detection performance is now comparable to that in adults and is sufficiently encouraging to envisage clinical implementation.