Part of MONSTER: https://arxiv.org/abs/2502.15122.

Dreamer is a multimodal dataset that includes electroencephalogram (EEG) and electrocardiogram (ECG) signals recorded during affect elicitation using audio-visual stimuli [1], captured with a 14-channel Emotiv EPOC headset. It consists of data recording from 23 participants, along with their self-assessments of affective states (valence, arousal, and dominance) after each stimulus. For our classification task, we focus on the arousal and valence labels, referred to as DreamerA and DreamerV respectively.

The dataset is publicly available [2], and we utilize the Torcheeg toolkit for preprocessing, including signal cropping and low-pass and high-pass filtering [3]. Note that only EEG data is analyzed in this study, with ECG signals excluded. Labels for arousal and valence are binarized, assigning values below 3 to class 1 and values of 3 or higher to class 2, and has been split into cross-validation folds based on participant.

[1] Stamos Katsigiannis and Naeem Ramzan. (2017) Dreamer: A database for emotion recognition through EEG and ECG signals from wireless low-cost off-the-shelf devices. IEEE Journal of Biomedical and Health Informatics, 22(1):98–107.

[2] Stamos Katsigiannis and Naeem Ramzan. (2017). Dreamer: A database for emotion recognition through EEG and ECG signals from wireless low-cost off-the-shelf devices. https://zenodo.org/records/546113.

[3] Zhi Zhang, Sheng-Hua Zhong, and Yan Liu. (2024). TorchEEGEMO: A deep learning toolbox towards EEG-based emotion recognition. Expert Systems with Applications.