DEED: A Multimodel Dataset for Dream Emotion Classification

Abstract

Abstract Although the number of research on exploring the brain neural mechanisms has been increasing dramatically, the dream-related aspects - especially dream emotion – are not yet well understood. This status is exacerbated by the lack of sufficient EEG dream data with emotion labels. To accelerate research on decoding the neural mechanisms of dream emotions, we released the Dream Emotion Evaluation Dataset (DEED), which contains multimodel dream-related information. In addition to multiple emotional personal trait assessment questionnaires, the neural data was collected from 38 participants over 82 nights by polysomnography (PSG). After each rapid eye movement (REM) awakening, participants reported their dream contents and affective states. Using the DEED, we implemented the first dream emotion classification algorithms, which are support vector machine (SVM), convolutional neural networks (CNN) and convolutional neural networks (CNN). Meanwhile, power spectral density (PSD), differential entropy (DE), multi-frequency band common spatial pattern (MCSP) is employed for feature extraction. Our results indicated that the combination of MCSP and SVM has the highest accuracy (83.6%). To summarize, the open-source DEED provides valuable data for addressing the neural mechanism of the dream, and several algorithms have verified its credibility in this work. We encourage researchers to use it to investigate the relationship between dreams and neural activities and develop the AI algorithms with biological intelligence.