An automated EEG-based framework using Empirical Mode Decomposition combined with Fuzzy Entropy and K-Nearest Neighbors classifier achieved 92.33% accuracy for multi-class dream emotion recognition from REM sleep signals.
Key Findings
Results
The EMD-FuzzEn method combined with the KNN classifier achieved the best multi-class classification accuracy of 92.33% for positive, neutral, and negative dream emotion recognition.
Accuracy was 92.33 ± 0.82% for multi-class (positive, neutral, negative) classification
Classification was performed over 20-second EEG segments
Four classifiers were compared: K-Nearest Neighbors (KNN), Support Vector Machine, Extreme Gradient Boosting, and Random Forest
The publicly available Dream Emotion Evaluation Dataset (DEED) was used
Results represent 'substantial advancement in classification over prior studies'
Results
The EMD-FuzzEn with KNN classifier achieved 96.47% accuracy for neutral versus non-neutral dream emotion binary classification.
Accuracy was 96.47 ± 0.83% for neutral versus non-neutral classification
This was one of two binary classification tasks evaluated alongside positive versus negative
Classification used 20-second EEG segments
EMD decomposed EEG signals into subbands from which Fuzzy Entropy was extracted
Results
The EMD-FuzzEn with KNN classifier achieved 90.69% accuracy for positive versus negative dream emotion binary classification.
Accuracy was 90.69 ± 1.51% for positive versus negative classification
This binary classification task had higher variability (±1.51%) compared to neutral vs. non-neutral (±0.83%)
Classification was performed over 20-second EEG segments
EEG signals were recorded during the Rapid Eye Movement (REM) sleep stage
Results
ReliefF feature selection identified temporal and frontal EEG regions, particularly T7 and T8 channels, as the most discriminative for dream emotion classification.
The ReliefF algorithm was applied to select the most discriminative features from the extracted Fuzzy Entropy values
T7 and T8 channels (temporal regions) were highlighted as particularly important
Frontal EEG regions were also identified as distinctive
Feature selection was applied after Fuzzy Entropy extraction from DWT and EMD subbands
Methods
The proposed methodology combined EEG signal decomposition via DWT and EMD with single nonlinear feature extraction using Fuzzy Entropy for dream emotion classification.
Two decomposition methods were compared: Discrete Wavelet Transform (DWT) and Empirical Mode Decomposition (EMD)
Only a single nonlinear feature, Fuzzy Entropy (FuzzEn), was extracted from each subband
EMD-based decomposition outperformed DWT-based decomposition across classification tasks
The feature matrix from FuzzEn was fed into four classifiers after ReliefF feature selection
EEG signals from REM sleep were the source data, as dreams are intrinsically linked to emotional processes
What This Means
This research suggests that brain activity recorded during dreaming (specifically during REM sleep) contains patterns that can be automatically identified and classified as positive, neutral, or negative emotions using machine learning. The researchers developed a system that takes EEG (electroencephalogram) brainwave signals, breaks them down into component parts using a method called Empirical Mode Decomposition, extracts a single mathematical complexity measure called Fuzzy Entropy, and then uses a classifier algorithm to sort the signals into emotional categories. This system achieved over 92% accuracy in distinguishing between three emotion types and nearly 97% accuracy when just separating neutral from non-neutral dreams.
A notable finding was that the regions of the brain most useful for identifying dream emotions were the temporal lobes (measured at electrode positions T7 and T8, located on the sides of the head) and the frontal regions. This suggests that these brain areas play a particularly important role in the emotional content of dreams. The study used a publicly available dataset of EEG recordings taken during REM sleep, where dreaming most commonly occurs.
This research suggests that automated EEG analysis could eventually be used to study the emotional content of dreams without requiring people to wake up and report what they dreamed, which could advance understanding of how dreams relate to mental health and emotional well-being. The authors note this represents a significant improvement over previous studies attempting similar dream emotion classification.
Sayad Mojdehbar N, Mohammadzadeh Asl B, Zarei A. (2026). Automatic EEG-based dream-related emotion recognition using fuzzy entropy and efficient signal decomposition methods.. Computers in biology and medicine. https://doi.org/10.1016/j.compbiomed.2025.111425