Cardiac arrest is a condition with high morbidity and mortality. Many survivors of cardiac arrest subsequently end up in a coma state, and these patients will go onto achieve varying levels of neurological recovery, ranging from brain death to full recovery. Electroencephalogram (EEG) analysis can be used to predict the neurological outcome of a cardiac arrest patient, but the patterns are complex and human analysis is time-consuming and error-prone. Therefore, machine learning could be used to increase the accuracy and efficiency of prognostication. One particular machine learning technique that is applicable is the variational autoencoder. This architecture compresses a high-dimensional feature space down to a low-dimensional latent space by teaching the model to encode the data itself. This can be used to extract a small number of features from a complex dataset, such as a collection of EEGs. Another benefit of this architecture is that the variational constraint naturally promotes the extraction of interpretable features. We trained a variational autoencoder to extract features from EEGs in the I-CARE (International Cardiac Arrest REsearch consortium) database and used those features in a Gradient Boosting Machine to predict neurological outcome after cardiac arrest. The resulting model was able to differentiate between good neurological outcome, defined as Cerebral Performance Category (CPC) of 1 or 2, versus poor neurological outcome, defined as CPC of 3-5. In a 100-fold permutation cross-validation, the model had an area under the receiver operating characteristics curve of 0.70 +/- 0.03. These results demonstrate that it is possible to use variational autoencoders to extract EEG features that are useful for downstream tasks. This opens the door to more interpretable models for EEG analysis in the future.