The purpose of this work is to comprehensively evaluate the performance of convolutional and recurrent neural network architectures on the PhysioNet/Computing in Cardiology Challenge 2018. The competition dataset was randomly split into 793 training subjects, 97 test subjects, and 102 validation subjects. Time-series sampled at 2 Hz were derived from the EEG, EOG, CHIN, CHEST, and ABDOMINAL signals and normalized on a per subject basis. For instance, time-series derived from the EEG consisted of spectral energies in the delta (0.5-3 Hz), theta (3-8 Hz), alpha (8-13 Hz), beta (13-25 Hz), and gamma bands (25-50 Hz) computed over a 5-second window shifting by 0.5 seconds. Time-series derived from the CHEST and ABDOMINAL channels consisted of breath-amplitude, breath-width, and inter-breath intervals. Next, 1-minute windows of these time-series were associated with the subject’s arousal state at the center of the window thus forming the pairs necessary for model training. More than 1 million target-arousal and 17 million no-arousal pairs were used for model training, and more than 200K target-arousal and 4 million no-arousal pairs were used for testing and validation. Finally, Google Cloud ML Engine was used to search, evaluate, and select model hyperparameters using only data from the training and test subjects (e.g. size of convolutional kernels, number of convolutional and fully-connected layers, absence or presence of recurrent layers etc). The best hyperparameter combination was then evaluated on the validation set to confirm model generalizability. At the end of Phase 1 of the competition, a leading architecture relies on a convolutional-recurrent architecture and achieves a test-set AUC-ROC score of 0.86 and a validation set AUC-ROC score of 0.85 and AUC-PR score 0.32.