From high- to low-density EEG for automatic classification of dream experiences during stage 2 of NREM

Machine learning models trained on high-density electroencephalography (EEG) signals can automatically detect whether a person is dreaming during the N2 stage of non-rapid eye movement (NREM) sleep with high accuracy. Using permutation-based channel selection, the models achieved up to 0.94 accuracy, F1 score, precision, and recall, an area under the receiver operating characteristic curve of 0.97, and a kappa of 0.88 on a balanced dataset of dream experience and no experience reports. Performance remained similar when using 30–40 EEG channels, and removing occipital channels slightly improved accuracy by 0.02. On a separate set of dream reports without recall, accuracy was 0.7. Reducing the number of channels makes portable, low-cost devices for real-time dream detection feasible.

Study at a glance

Abstract

Abstract This study proposes a method to automatically identify dream experience (DE) and no experience (NE) during the sleep stage N2 of nonrapid eye movement (NREM). We investigated the use of machine learning (ML) to automatically identify when a subject is having a dream during NREM from electroencephalography (EEG) signals. We use permutation-based channel selection to identify the most informative EEG channels for the classification of DE and NE and to select a set of channels that allow focus on the most important areas of the brain at the scalp level. The results show that when the ML models are trained on a balanced dataset containing both DE and NE reports, along with high-density EEG, they can achieve a classification performance of up to 0.94 in accuracy, F1 score, precision and recall, an Area Under the Receiver Operating Characteristic of 0.97, and a kappa of 0.88. Performance decreases while we reduce the number of channels, but it remains like using up to 30–40 EEG channels. We show that ML models trained on high-density EEG to classify NE and DE can identify whether a subject was dreaming, achieving an accuracy of 0.7 on a separate set of dream reports where subjects reported a dream experience without recall, and channel selection methods have shown that performance could increase by 0.02 when EEG channels are removed from the occipital area. Our results show a high classification performance for automatic dream detection and the need to reduce the number of EEG channels needed to create the ML models, thus obtaining low-cost portable devices that can be used in real-life scenarios. Statement of Significance Dreaming during non-REM sleep is difficult to detect objectively. Here, we show that machine learning models leveraging high-density EEG can accurately classify the presence vs absence of dream experiences during N2 sleep. A key finding is that this high performance is maintained with a substantially reduced channel set, establishing the feasibility of portable, low-density EEG systems for real-time dream detection. Our work provides a critical objective correlate for subjective dreaming and opens new avenues for practical applications in sleep research and clinical diagnostics.