Attention-Based Network for Weak Labels in Neonatal Seizure Detection.

Seizures are a common emergency in the neonatal intesive care unit (NICU) among newborns receiving therapeutic hypothermia for hypoxic ischemic encephalopathy. The high incidence of seizures in this patient population necessitates continuous electroencephalographic (EEG) monitoring to detect and treat them. Due to EEG recordings being reviewed intermittently throughout the day, inevitable delays to seizure identification and treatment arise. In recent years, work on neonatal seizure detection using deep learning algorithms has started gaining momentum. These algorithms face numerous challenges: first, the training data for such algorithms comes from individual patients, each with varying levels of label imbalance since the seizure burden in NICU patients differs by several orders of magnitude. Second, seizures in neonates are usually localized in a subset of EEG channels, and performing annotations per channel is very time-consuming. Hence models which make use of labels only per time periods, and not per channels, are preferable. In this work we assess how different deep learning models and data balancing methods influence learning in neonatal seizure detection in EEGs. We propose a model which provides a level of importance to each of the EEG channels - a proxy to whether a channel exhibits seizure activity or not, and we provide a quantitative assessment of how well this mechanism works. The model is portable to EEG devices with differing layouts without retraining, facilitating its potential deployment across different medical centers. We also provide a first assessment of how a deep learning model for neonatal seizure detection agrees with human rater decisions - an important milestone for deployment to clinical practice. We show that high AUC values in a deep learning model do not necessarily correspond to agreement with a human expert, and there is still a need to further refine such algorithms for optimal seizure discrimination.