Bridging the source-target mismatch with pseudo labeling for neonatal seizure detection

The mismatch between training and testing conditions is a known problem in the machine learning community. In this work, we outline a process of how a model which was trained under one set of conditions can be adapted to a new set of conditions by means of pseudo labeling. This is shown for the domain area of neonatal seizure detection. A previously developed deep learning architecture is first trained on a publicly available source dataset. It is then evaluated on another target dataset which was recorded in a different center, with different equipment, and annotated by a different expert. This model is then used to create pseudo labels on a sample of the target dataset, fine-tuned with the created pseudo labels, and re-evaluated on the target dataset. The results show a relative improvement of 13.5% and 28.8% in AUC and the number of seizures detected respectively. Various factors of the pseudo labeling procedure such as the amount of data vs confidence in pseudo labels are analyzed and presented.