To identify neural correlates distinguishing conscious from unconscious states during sleep using high-density EEG and machine learning.

Sleep represents a key condition to study the neural correlates of consciousness, with alternating conscious and unconscious periods within the same behavioral state. Serial awakening paradigms combined with high-density EEG (HDEEG) allow to identify differences in pre-awakening brain activity linked to subjective reports while minimizing behavioral confounds.

We recorded HD-EEG (185 electrodes) in 115 healthy participants overnight, who were repeatedly awakened (n = 698 awakenings) and asked if they were dreaming. One-minute pre-awakening HD-EEG was classified as conscious experience if patients reported dream activity (CE, n = 341) or no conscious experience (NCE, n = 357) if no dreaming was reported. Functional connectivity matrices were computed via lagged coherence across delta, theta, alpha, beta, and gamma bands. Feature selection was performed sequentially on power and connectivity measures using nonparametric testing (p<0.05, effect size>0.6), recursive feature elimination via Random Forest, and LASSO regression for multicollinearity. The final set with 20 features was used in three classifiers (Support Vector Machine, Elastic Net, Decision Tree) via five-fold cross validation at subject level.  Model performance was evaluated with area under the receiver operating characteristic curve (ROC-AUC).

All three methods achieved ROC-AUC=0.82 distinguishing CE from NCE. Frontal delta power was greater in CE, while posterior delta and both frontal and posterior alpha were greater in NCE (p<0.05). Delta power was the strongest differentiator, while alpha changes were also informative. Posterior region outperformed the frontal, but predictive power was highest when both were integrated.

HD-EEG features can distinguish conscious from unconscious states during sleep. Models achieved reliable performance, with posterior delta power emerging as the key differentiator of conscious experience. These posterior-dominant patterns align with Integrated Information Theory predictions and support the feasibility of scalp EEG in dream detection.