Emotion Recognition Across Wakefulness and Sleep Using High-Density EEG and Machine Learning Techniques

Dreams, occurring during Rapid Eye Movement (REM) and non-REM (NREM) sleep, are essential for emotional processing and memory consolidation. However, decoding emotional content from dreams remains challenging due to their subjective nature and reliance on self-reports. This study explores the feasibility and differences of classifying emotions during wakefulness and sleep using electroencephalography (EEG) data. EEG recordings were collected from 14 participants exposed to 40-s videos before sleep and during sleep through serial awakenings in REM and NREM stages. Preprocessing included artifact rejection, band-pass filtering, and feature extraction with Common Spatial Patterns (CSP). Machine learning (ML) models, including Gradient Boosting, eXtreme Gradient Boosting (XGB), and Random Forest (RF), were trained to classify low versus high levels of Arousal and Valence. The highest Arousal classification performance during wakefulness was achieved with XGB on 40-s segments filtered at 4–45 Hz (precision: 0.73; Area Under the Receiver Operating Characteristic curve (AUROC): 0.80), and for the case of sleep, RF performed best on 40-s or 5-s segments filtered at 0.1 45 Hz (precision: 0.65; AUROC: 0.65). Valence classification proved to be more challenging, particularly for sleep, with a maximum precision of 0.60 and an AUROC of 0.69. These findings demonstrate the potential of EEG-based emotion classification during dreams, highlighting the need for robust methods and wearable EEG devices for real-time dream decoding. This approach could improve our understanding of emotional processing during sleep and support the clinical management of psychiatric conditions. Future work should investigate optimal segment durations, sleep stage-specific effects, and minimal EEG channel configurations to improve dream emotion recognition.