Aims: In intensive care units, patients who have fallen into a coma due to cardiac arrest are at risk of succumbing to brain injury. Accurate prognoses help doctors decide the appropriate treatment modalities. To avoid "false positives” caused by subjectivity, we proposed an approach to predict neurologic prognostication of cardiac arrest patients by combining data augmentation and a graph neural network.
Methods: We adopted a two-stage preprocessing strategy to achieve data augmentation. Firstly, EEGs of all patients were shuffled and the inputs of model were defined as EEG segments in 1 hour but not 72 hours to solve the problem of data missing during monitoring. Secondly, new EEG segments were generated by randomly selecting 4-minute or 3-minute EEG segments to balance the numbers of samples with different Cerebral Performance Category (CPC) scores. After data preprocessing, the inter-channels features of EEG segments were represented by adjacency matrices calculated by samples in different channels. A graph neural network (GNN) with 3 graph convolutional layers was adopted to learn these features and predicted prognoses from EEG segments. Because of data augmentation, the outputs of GNN were the prognoses of EEG segments in 1 hour. But final prognoses were decided by EEGs at 12, 24, 48, and 72 hours which should have 12, 24, 48, and 72 EEG segments respectively. So the final prognosis for a patient was decided by the weighted average results of all EEG segments via the time information.
Results: Our team, Aircas, achieved the 5-fold cross-validation Challenge scores of 0.21(±0.08), 0.29(±0.10), 0.43(±0.22) and 0.44(±0.19) at 12, 24, 48, and 72 hours, respectively. Unfortunately, due to technical issues, we finally achieved the Challenge metric scores of 0.02, 0.02, 0.00 and 0.00 on the hidden validation set.