Multi-Label Neonatal EEG Segmentation Enriched with Semi-Synthetic Data

Electroencephalography (EEG) is widely used in routine clinical practice. In particular, it is of the utmost importance for monitoring newborns suffering from acute neonatal encephalopathy. In such a pathological context, continuous acquisitions during the first days of life must be analyzed, which represents an insurmountable workload for clinicians. Machine learning approaches are a way of helping human experts in their work. More specifically, deep learning models have the ability to segment specific physiological and pathological patterns. The main obstacle to the development of these models is their correct training from a sufficient number of annotations. Indeed, building annotated EEG datasets involves a time cost that is generally unacceptable. In this paper, a deep learning model dedicated to the segmentation of multiple patterns in newborn EEG is proposed. We study the impact of training this model either with only human-defined annotations on the real EEG, or with augmented datasets that also include synthetic patterns embedded in the real signal, endowed with their annotations. The mixed use of semi-synthetic and real data leads to insights about future development of effective AI-based assistive tools for newborn EEG assessment.