As part of The George B. Moody PhysioNet Challenge 2023, we developed a method to predict neurological recovery from coma based on the patient's EEG recordings. We utilized a large Challenge training set of EEG recordings with sufficient signal quality and applied self-supervised learning techniques using contrastive learning on short EEG segments of about 46 seconds. This involved randomly selecting two disjoint segments from each 5-minute EEG recording and applying different augmentations. Using a deep residual 1D convolutional network and contrastive loss function, we solved the task to pair the corresponding segments within the minibatch to obtain their embedding onto a unit hypersphere in a 512-dimensional space. We then used the obtained latent factors and pre-trained neural network for downstream tasks, specifically binary classification of individual EEG recordings based on neurological prognosis. The final prediction for the patient was obtained by aggregating predictions from available EEG recordings with sufficient quality in the last six hours. In the internal evaluation, we achieved a challenge score of 0.53 for 72 hours using 10-fold cross-validation on the training set. In the unofficial round, our submission (team "CeZIS”) was ranked 6th out of all submissions with a challenge score of 0.67 for 72 hours on the hidden validation set. For the official round, we aim to expand the training set by incorporating additional EEG datasets and improve prediction accuracy by considering the time evolution of the patient's EEG and other factors. Our self-supervised learning method on EEG could also be applicable in solving other downstream tasks in automatic EEG analysis.