Aims: The aim of this work was to design a neural network that maximized true positive rate (TPR) of predicting poor outcomes in coma patients following cardiac arrest, given false positive rate less than 0.05. This work compared neural network performance using 18 bipolar EEG channels versus a subset of channels. Completed Methods: An LSTM was used to observe hour-to-hour changes of clinically informed features. The features were median amplitude and standard deviation of EEG (using entire recording within hour), both per channel (collapse across time). Features were calculated for each available hour. To reduce dimensionality, electrodes were grouped by region: left-anterior, right-anterior, left-posterior, right-posterior, and all. The median of each feature was taken per region (collapse across space). Completed Results: Using 5-fold cross-validation within the training set, the LSTM, using 18 channels, yielded a TPR of 0.39 including data from all 72 hours. On the validation set, the model yielded a TPR of 0.31. With 5-fold cross-validation, the LSTM, using the reduced 5 channels, yielded a TPR of 0.40 including data from all 72 hours. On the validation set, the model yielded a TPR of 0.33. Conclusions and Planned Methods: This work demonstrated it is feasible to use a reduced set of EEG channels to predict coma recovery. Improved performance of the reduced model may suggest this approach is effective to reduce noise from artifact in cardiac arrest, which results in whole-brain injury. The next steps are to build a 2-D CNN using reduced-channel EEG or power spectral densities to capture temporal, spectral, and spatial features. The CNN will train on the 72 hours of data separately and then make an outcome prediction with an LSTM combining information from all 72 hours. Lastly the output of the CNN-LSTM will be fused with a random forest trained on patient features.