Figure 2:

Illustration of the process to which the preprocessed signals are reshaped and fed through the proposed MAD network. The top plot shows a 10 second window of signals with an arousal annotated at 5 seconds. Arousal and wake probabilities are computed from the reshaped input being fed through the network structure consisting of CNN (convolutional neural network), LSTM (Long Short-Term Memory), and FC (fully connected) layers. The three network types (CNN, LSTM, FC) each have a main function, (1) extraction of features describing the 1 second segment of the signals, (2) computing features with temporal information (e.g. frequency change), (3) mapping temporal features to predict wake and arousal labels.