Given the EEGs for this Challenge, we extract, from hour-long traces for each patient and channel, (a) Pearson's correlation coefficients Cij between the signals from the 18 channels, (b) the fraction time in which suppression bursts occur (identified by calculating the fraction of time the signal spends below a specified threshold, in 10-second intervals, for each of the channels, and then mean-centering and normalizing the data), and (c) the background activity (quantified via Fourier amplitudes of alpha, beta, gamma, and delta waves). We feed the outputs for each hour-long trace for each patient in the dataset, from (a), (b), and (c), to three different 1DCNNs - CNNa, CNNb, and CNNc. These CNNs are trained for a fraction of the data provided to yield (i) the probabilities of the outcome and (ii) the CPC per hour per patient. We then feed the outputs from these CNNs to three variable-input LSTMs, and additional patient information (e.g., age, etc.) to a fourth variable-input LSTM. The outputs of these LSTMs are averaged to obtain the final (i) probabilities of the outcome and (ii) the CPC per patient in the training dataset. We now use this combination of trained CNNs and LSTMs with EEGs that are not included in the training dataset to test the efficacy of our predictions. The Python program that we have submitted to the Physionet Challenge shows the following scores on the Leaderboard: 0.13 (for 12 hours); 0.48 (for 24 hours); 0.49 (for 48 hours); and 0.61 (for 72 hours). We are fine-tuning our deep-learning scheme to obtain better scores by extracting other features, e.g., multifractal spectra of the given EEGs.