On the Selection of Time-Frequency Features for Improving the Detection and Classification of Newborn EEG Seizure Signals and Other Abnormalities

摘要

This paper presents new time-frequency features for seizure detection in newborn EEG signals. These features are obtained by translating some relevant time features or frequency features to the joint time-frequency domain. A calibration procedure is then used for verification. The relevant translated features axe ranked and selected according to maximal-relevance and minimal-redundancy criteria. The selected features improve the performance of newborn EEG seizure detection and classification systems by up to 4% for 100 real newborn EEG segments.