Background EEG monitoring provides important information about the neurological status of the preterm infant but is difficult to interpret for most. We aim to automatically detect the typical bursting pattern (trace discontinu) of the preterm EEG and compare the detections with expert manual annotations.
Methods The method was based on the single channel EEG method of Palmu et al. but extended to 8-channel recordings for the first time. The EEG signal was first filtered with a Kaiser-window filter and the output of a non-linear energy operator (NLEO) was calculated. The NLEO signal was smoothed and corrected for baseline artefacts. A burst was identified if the resulting signal remained above 1.5µV2 for longer than 1s. Each EEG channel was processed separately and the final outcome was considered to be an activity burst if a burst was detected in 2 or more channels.
The method was tested on a database of 24 babies born before 30 weeks gestation (avg 26.5 + –1.7 weeks). For each baby 10 minutes of 8-channel EEG signal was analysed.
Results Agreement with the expert burst annotations was on average 77.2% over the 24 subjects (Min:59.6%, Max:90.3%, std-dev:8.5%). Most errors consisted of disagreement over the precise start and end points of a burst.
Conclusions Automatic burst detection has been applied for the first time to a large database of preterm 8-channel EEG. Promising results were obtained for automated EEG interpretation. Future work will attempt to reduce the error by use of more sophisticated methods to merge the per-channel detections.