Dream Emotions Identified Without Awakenings by Machine and Deep Learning from Electroencephalographic Signals in REM Sleep.

We explored the automatic classification of dreams with emotional content, which were collected by awakening 38 subjects after they had entered to Rapid Eye Movement (REM) sleep, and the dreams were recorded using 6 electroen-cephalographic (EEG) channels. We used the discrete wavelet transform for feature extraction and well-known classification algorithms, such as gradient boosting and random forest, as well a convolutional neural network for creating subject-independent models in different experimental setups. When creating a model to classify dreams with neutral emotion versus a dream with positive/negative emotion, we obtained accuracies of up to $0.66\pm 0.02$ . We classified dreams with positive versus negative emotional content, obtaining accuracies of up to $0.64 \pm 0.03$ . We were also able to classify dreamless sleep versus sleep with dreams with accuracies of up to $0.85 \pm 0.02$ , and obtained similar accuracies using 2–3 channels selected by the Non-dominated Sorting Genetic Algorithm II. Our results indicate that the proposed methods can classify dream -containing EEG signals with high accuracies. These are encouraging results towards the development of automatic methods that can facilitate the study of emotions in dreams and provide insight into the human psyche to address symptoms of psychiatric and sleep disorders.