Dream research is vital for brain science development. Investigating how the brain works during dreaming and how a brain switches the mode between consciousness and unconsciousness is super attractive to neuroscientists. Also, dream research can help humans to perform better in daily life, because dreams are considered to have the ability for memory consolidation and emotional recovery [1][2]. Therefore, research on exploring dream-related neural mechanisms has been increasing dramatically in recent years [3][4].
However, doing dream research is difficult since scientists don't have the tools to observe them directly - most studies rely on qualitative approach such as dream reports (a person writes out her dreams when she wakes up) or questionnaires (a person answers questions like "How many dreams have you recalled in the past month?"). When dream research comes to modeling neural mechanisms, such research tools seem to be insufficient. Scientists require quantitative EEG-based dream data for in-depth research.
To solve this dilemma, we establish a Dream Emotion Evaluation Dataset (DEED), which has the key features below: 1) Accessible EEG-based dream data with a three-class dream emotion labels (positive, neutral, and negative), this will serve as a benchmark for developing dream emotion classification algorithms; 2) A comprehensive EEG sleep data that contains identical REM stage labels, and this provides great flexibility to do dream and sleep-related mechanisms research. For artificial intelligence researchers, this unique dataset provides samples with non-stationary characters, obscure spatiotemporal features, and time-series signals encoded with dream mood information. This is suitable for testing algorithms with biological-like intelligence performances such as few-shot learning, generalization, and anti-interference ability. For neuroscience researchers, it provides a high quality dataset that enabling the potential exploration of the biological characteristics of dreams, the mechanisms of sleep-emotion interactions, and the neural manifestations of the transition between consciousness and unconsciousness, among others.
The paper organizations are as follows: Section 1 gave a general introduction about the current status of dream research and why we released the DEED dataset. Then we compared some related datasets and gave the foundation of DEED dataset. In section 2, we described the details of the data collection procedure, including four parts: the participant information, the experimental design, the data format, and the data preprocessing. To evaluate our data quality, we employed K-Nearest Neighbor (kNN), support vector machine (SVM), and convolutional neural networks (CNN) to classify the emotion valence (negative, neutral and positive). The methods and results are provided in sections 3 and 4 separately. The conclusion of the manuscript presented in the final section.
In summary, the main contributions of this paper are:
-
We developed the Dream Emotion Evaluation Dataset (DEED) for dream and sleep related research.
-
We employed the DEED for the development of dream emotion classification algorithm. The baseline of EEG net and CSP with SVM is 58.9%, 83.6% respectively.
1.1. Related work
There are some EEG datasets relate to the emotion behaviors. In such datasets, subjects typically complete an emotion-evoking task (watching movie clips that could induce different types of emotion) while awake and then measure their EEG [5][6]. Alternatively, subjects were asked to rate several video clips on emotion valences as a label for their clips [7].
Within these datasets, researchers have provided many new insights into the EEG related emotion research. For example, results from SEED (SJTU Emotion EEG Dataset) proved that neural signatures associated with different emotions exist, and share identical features across sessions and individuals [8]. Using the same dataset, Duan et al. [6] compared the emotional classification accuracies using different features (DE, DASM, RASM, and ES). They indicated that proposed features yielded acceptable classification performance. Moreover, the DE showed a more suited ability for emotion classification than other traditional features (DASM, RASM, and ES) and approached to 85% accuracies. Meanwhile, the EEG signals on frequency band Gamma relates to emotional states are more closely than other bands. The results of SEED has demonstrated the validity of EEG based emotion classifications. However, this elicitation by induced clips cannot extend to spontaneous emotional responses.
Another emotion-related dataset, DEAP (the Dataset for Emotion Analysis Using Physiological signals) presented the EEG and peripheral physiological signals of 32 participants. Participants rated some videos in terms of the levels of arousal, valence, like/dislike, dominance, and familiarity aspects, which are severed for the analysis of spontaneous emotions. Within the dataset, Koelstra et al. [9] found significant correlates between the participants and EEG frequencies, and the single-trial classification is better than the random level (average 57.6% for emotion valance).
Although much has been discovered through these datasets about the neural mechanisms behind emotions, these datasets were limited in the sober state[10]. In fact, neural mechanisms between dream and sorber has already shown a significant difference [11][12]. Recent studies have highlighted that dream experience is promoted by specific brain activation, characterized by reduced low frequencies and increased rapid frequency [13]. A high-quality dataset about emotion evaluation during dream worthwhile.
1.2. DEED foundation
How can we capture dreams with maximum efficiency and how can we be sure that the time period we mark is indeed a dream? To address these two questions, we used a method of waking subjects during the REM (rapid eye movement) period and asking them to self-report the presence or absence of dreams. There are two reasons for this: 1) Sleep processing can be dichotomies of REM and NREM (Non-rapid eye movement) [3]. While dreaming occurs during all stages of sleep, intense dreaming is largely confined to REM period, which is characterized as wake-like high-frequency EEG pattern. Around 80% of awakenings from REM sleep are followed by a confirmed dream report [14][15]. 2) To rule out the situation of dreams absent during the REM period, we awakened the subjects after they had entered the REM period for a certain period (5 am to participants wake up naturally), and asked them to report whether they had dreams or not. This combined approach has the ability to improve the viability of the data.
Given the pattern of brain neural activity, sleep during REM periods is believed to be the optimal timing for investigating the processing of emotionally related neural characters. Therefore, the dream emotion has been employed for DEED dataset. As we all know, dreaming is a succession of thoughts, images, and sensations occurring involuntarily in a person's mind during sleep, often accompanied by rich affective experience. From one point of view, emotional processes endure despite a major reorganization of brain activity patterns during sleep. In fact, the nature of the dream state has been proved to be highly subjective and affective [16], making the investigation of emotions is an excellent entry point for dream interpretation. This could also be an interesting observation from another point of view that the neural correlates of emotion has been, so far, only investigated in awake subjects.
Some features of REM period instead of NREM period make it a good object for emotional experience. In detail, compared to dreams detected in the NREM period, dreams during REM period are especially characterized by high vividness, hallucinations, and emotional load [17][18][19]. From a neurological point of view, most of the regions involved in emotional memory encoding and consolidation are highly activated during REM periods [2][20][21][22][23]. Some early fMRI studies [24][25] has comprehensively described cerebral activity distribution during REM periods. It demonstrated that brain activity in REM periods is associated with general activation of limbic and paralimbic areas, such as amygdaloid complexes, left thalamus, hippocampal formation, and pontine tegmentum, which contribute largely to the acquisition and adjustment of emotionally influenced memories [26][27][28]. By contrast, a relative deactivation of several cognitive-related cortices (inferior parietal, dorsolateral prefrontal, and orbitofrontal cortices, posterior cingulate gyrus, and precuneus) paralleling the aminergic downtoning and downregulation of noradrenergic input from the locus coeruleus has been linked REM periods to high arousal emotion [29][30].