2026 ECNP Congress
PS02-2253Dream features as a predictor of subjective sleep features: a comparison with standard EEG activity metrics
Introduction
Despite sleep being essential for well-being and mental health, approximately one third of the population is estimated to experience poor sleep quality [1]. Instrumental examinations based on standard EEG metrics have been found to be not fully aligned with subjective sleep feature [2]. Interestingly, though, some studies pointed to an association between dream characteristics and subjective sleep quality or duration.
Aims
In this study, we compared EEG data and dream features to explore their potential in predicting changes in subjective sleep depth and sleepiness perceived upon awakening.
Methods
The study was conducted at IMT School for Advanced Studies Lucca. Forty-four subjects were monitored overnight for up to four nights using high-density EEG. Participants were awakened several times during the night while in non-REM stage 2 sleep. Information about subjective sleep features (sleepiness and sleep depth), as well as dream features (including auditory, visual, and tactile dream contents), was collected at each awakening. All variables were coded on a 1–5 Likert scale (1 = low, 5 = high), while mean power values were used for EEG activity. Different machine learning models (including Random Forest, Support Vector Machines, and Neural Networks) were trained to test EEG features, dream features, and their combination in predicting subjective sleep characteristics. Data were splitted into training and test set (80 vs 20%) and 5-fold cross validation was applied. Statistical analyses were performed using the CaretEnsemble R package.
Results
A total of 432 dreams were analyzed. Subjective sleep depth and subjective sleepiness were recoded into two categories (low and high) based on their distribution to obtain more homogeneous classes (52.5% vs 47.5% and 62.7% vs 37.3% for sleep depth and sleepiness, respectively). The ROC metric was used to select the best model for each group of variables. The best performing models and their tuning parameters are reported in table 1. For subjective sleep depth, dream features showed the best AUC (0.72), while the model was only slightly improved by adding EEG features (AUC = 0.72), whereas EEG alone showed lower discriminative capability (AUC = 0.62). For subjective sleepiness, EEG turned out to be the best classifier (AUC = 0.74 vs 0.68 for dream features and 0.62 for the combination of variables).
Conclusions
In this study, we explored the potential of dream features in describing subjective sleep characteristics compared to EEG data. While EEG remains the best predictor of subjective sleepiness, it only slightly improves the prediction of subjective sleep depth when combined with dream features. However, standard metrics may not capture subjective sleeping features, while alternative metrics might prove to be more sensitive. Overall, this study highlights the potential of dream features for studying sleep and related disorders.
Table 1
| Sleep Depth | Sleepiness | |||||
| Model | Tuning | ROC | Model | Tuning | ROC | |
| Dream Features | Support vector machine | C = 0.01 | 0.72 | Neural network | Size = 3Decay = 0.1 | 0.68 |
| EEG | Support vector machine | C = 0.01 | 0.61 | Random forest | Mtry = 2 | 0.74 |
| Dream Features + EEG | Support vector machine | Sigma = 0.05C = 1 | 0.75 | Neural network | Size = 5 Decay = 0.1 | 0.2 |
References
1.https://www.cdc.gov/sleep/data-research/facts-stats/adults-sleep-facts-and-stats.html 2.Stephan AM, Lecci S, Cataldi J, Siclari F. Conscious experiences and high-density EEG patterns predicting subjective sleep depth. Curr Biol. 2021 Dec 20;31(24):5487-5500.e3. doi: 10.1016/j.cub.2021.10.012. Epub 2021 Oct 27. PMID: 34710350.