Automatic EEG-based dream-related emotion recognition using fuzzy entropy and efficient signal decomposition methods

Dreams can reflect our profound needs and desires, intrinsically linked to emotional processes. In recent years, research on dream emotions has been made possible by capturing electroencephalogram (EEG) signals from the Rapid Eye Movement (REM) sleep stage. This study aims to develop an automated framework for classifying dreams with positive, neutral, and negative emotions using the publicly available Dream Emotion Evaluation Dataset (DEED).

The proposed methodology of this research involves an initial decomposition of the EEG signal into various subbands using the Discrete Wavelet Transform (DWT) and Empirical Mode Decomposition (EMD). Subsequently, the single nonlinear feature of Fuzzy Entropy (FuzzEn) is extracted from each subband, followed by the selection of the most discriminative features through the ReliefF algorithm. The resulting feature matrix is then fed into four classifiers: K-Nearest Neighbors (KNN), Support Vector Machine, Extreme Gradient Boosting, and Random Forest for classification.

Superior classification performance was achieved using the EMD-FuzzEn method combined with the KNN classifier over the 20s segments. This combination yielded 92.33 ± 0.82 % accuracy for multi-class, 96.47 ± 0.83 % for the neutral versus non-neutral, and 90.69 ± 1.51 % for the positive versus negative dream emotion classification. The results of ReliefF feature selection further highlighted the distinctive importance of temporal and frontal EEG regions, particularly the T7 and T8 channels.

Consequently, this methodology demonstrates strong potential for identifying dream emotions, offering insights into signal decomposition and feature extraction efficacy, and representing substantial advancement in classification over prior studies.