Deep Learning for EEG Seizure Detection in Preterm Infants

EEG is the gold standard for seizure detection in newborn infant, but interpretation preterm group particularly challenging; trained experts are scarce and task of interpreting real-time arduous. Preterm infants reported to have a higher incidence seizures compared term infants. morphology differs from that infants, which implies algorithms on may not be appropriate. The developing specific becomes extra-challenging given limited amount annotated data available. This paper explores novel deep learning (DL) architectures neonatal study tests compares several approaches address problem: training full-term infants; age-specific transfer learning. system performance assessed large database continuous recordings 575[Formula: see text]h duration. It shown accuracy validated term-trained algorithm, based support vector machine classifier, when tested falls well short achieved An AUC 88.3% was obtained as 96.6% EEG. When re-trained EEG, marginally increases 89.7%. alternative DL approach shows more stable trend cohort, starting with an 93.3% algorithm reaching 95.0% by model using available data. proposed avoids time-consuming explicit feature engineering leverages existence model, resulting accurate predictions minimum