DreamSense: A Multi-Task EEG-based Framework for Predicting Dream Occurrence, Polarity, and Nightmare Severity
Ensteih Silvia, K. Dutta, P Anupama V, K Ramya, T Jayashreedevi K, Ashwini Ganesh Ratod
International Conference Intelligent Data Communication Technologies and Internet Things June 10, 2026 DOI: 10.1109/icici68773.2026.11581096 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Computational modeling study Peer reviewed |
|---|---|
| Topics | Dreaming |
| Key points | Task-specific architectures performed best: a CNN-LSTM detected dream occurrence at 93.9% accuracy (AUC 0.977), a residual CNN with self-attention classified emotional sentiment at 94.67% accuracy (Kappa 0.92), and XGBoost assessed nightmare intensity at 81.42% test accuracy with 79.9% cross-validation consistency. The authors argue deep learning excels with large temporal sequences while tree-based methods provide superior interpretability for clinical applications. |
Abstract
Sleep disorders and nightmare disturbances challenge clinical assessment due to subjective self-reporting and lack of quantitative precision. This work presents DreamSense, a multi-task deep learning framework using electroencephalography (EEG) biomarkers that classify dream states on three independent tasks: occurrence detection, emotional sentiment, and nightmare intensity. Via systematic architecture optimization, we identify task-specific models based on dataset complexity. Dream occurrence detection utilizes a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture with bidirectional processing on dual-channel Sleep-EDF data (5,085 epochs), yielding 93.9% accuracy and 0.977 Area Under the Curve (AUC). Emotional sentiment classification is addressed with an improved residual Convolutional Neural Network (CNN) incorporating self-attention on 62-channel SEED recordings (2,250 trials), with 94.67% accuracy and Cohen's Kappa of 0.92. Nightmare intensity assessment determines that XGBoost is optimal among comparative classifiers with 81.42% test accuracy and 79.9% cross-validation consistency using 32 neurophysiological biomarkers, dominated by slow-theta oscillations at 35% weight. Results indicate that deep learning excels with large temporal sequences while tree-based methods provide superior interpretability for clinical applications. We establish a scalable and modular framework for automated sleep disorder assessment, demonstrating that task-specific architectural optimization across independent EEG classification stages yields robust, interpretable results suitable for clinical sleep research.