Leveraging Unlabeled Electroencephalographic Data to Predict Neurological Recovery for Comatose Patients Following Cardiac Arrest.

Abstract

In response to the 2023 George B. Moody PhysioNet Challenge, we propose an automated, unsupervised pre-training approach to boost the performance of models that predict neurologic outcomes after cardiac arrest. Our team, (BrownBAI), developed a model architecture consisting of three parts: a pre-processor to convert raw electroencephalograms (EEGs) into two-dimensional spectrograms, a three-layer convolutional neural network (CNN) encoder for unsupervised pre-training, and a time series transformer (TST) model. We trained the CNN encoder on unlabeled five-minute EEG samples from the Temple University EEG Corpus (TUEG), which included more than 20x the patients available in the PhysioNet competition training dataset. We then incorporated the pre-trained encoder into the TST as a base layer and trained the composite model as a classifier on EEGs from the 2023 PhysioNet Challenge dataset. Our team was not able to submit an official competition entry and was therefore not scored on the test set. However, in a side-by-side comparison on the competition training dataset, our model performed better with a pretrained (competition score 0.351), rather than randomly initialized (competition score 0.211) CNN encoder layer. These results show the potential benefits of leveraging unlabeled data to boost task-specific performance of predictive EEG models.