Abstract
Lucid dreaming (LD) is a state of conscious awareness of the ongoing oneiric state, predominantly linked to REM sleep. Progress in understanding its neurobiological basis has been hindered by small sample sizes, diverse EEG setups, and artifacts like saccadic eye movements. To address these challenges in characterizing the electrophysiological correlates of LD, we introduced an adaptive multistage preprocessing pipeline, applied to human data (male and female) pooled across laboratories, allowing us to explore sensor- and source-level markers of LD. We observed that, while sensor-level differences between LD and nonlucid REM sleep were minimal, mixed-frequency analysis revealed broad low alpha to gamma power reductions during LD compared with wakefulness. Source-level analyses showed significant beta power (12–30 Hz) reductions in right central and parietal areas, including the temporoparietal junction, during LD. Moreover, functional connectivity in the alpha band (8–12 Hz) increased during LD compared with nonlucid REM sleep. During initial LD eye signaling compared with the baseline, source-level gamma1 power (30–36 Hz) increased in right temporo-occipital regions, including the right precuneus. Finally, functional connectivity analysis revealed increased interhemispheric and inter-regional gamma1 connectivity during LD, reflecting widespread network engagement. These results suggest that distinct source-level power and connectivity patterns characterize the dynamic neural processes underlying LD, including shifts in network communication and regional activation that may underlie the specific changes in perception, memory processing, self-awareness, and cognitive control. Taken together, these findings illuminate the electrophysiological correlates of LD, laying the groundwork for decoding the mechanisms of this intriguing state of consciousness.
Significance Statement
Lucid dreaming (LD) is a unique state of oneiric awareness, where individuals recognize they are dreaming while still in the dream. LD neural correlates remain elusive, as it is very rare and difficult to reproduce in the laboratory. Using an advanced preprocessing pipeline, we harmonized diverse EEG datasets to analyze the largest LD sample to date. We observed gamma power increases in the precuneus during initial eye lucidity signaling; beta power reductions in parietal areas, including the temporoparietal junction; and enhanced alpha and gamma connectivity during LD over nonlucid REM sleep. These findings shed light on how the brain generates self-referential awareness and volitional action even during sleep.
Introduction
During lucid dreaming (LD) subjects know they are dreaming and sometimes are able to control the oneiric content. Lucid dreamers are aware of the hallucinatory character of ongoing dream perceptions (Windt, 2010, 2021; Hobson et al., 2014; Gross et al., 2021; Waters et al., 2021), which often allows them to voluntarily interact with their internal world models (Dresler et al., 2014) via fully immersive simulations (Windt, 2010; Hobson et al., 2014). LD happens predominantly during rapid-eye-movement (REM) sleep, but can also occur during sleep onset (N1) and light sleep (N2) stages (Brylowski et al., 1989; Stumbrys and Erlacher, 2012; Mota-Rolim et al., 2015).
Electroencephalographic (EEG) research of LD was initiated in the late 1970s with the introduction of the eye signaling technique: experienced lucid dreamers were instructed to perform intentional left→right→left→right (LRLR) ocular movements in their dreams (Hearne, 1978; LaBerge, 1980b; La Berge et al., 1981), which in turn can be objectively assessed via electrooculography (EOG), as eye muscles are not affected by the general muscle atonia experienced during REM sleep (Aserinsky and Kleitman, 1953; Dement and Kleitman, 1957). This technique has since become the gold standard method in LD research, allowing for electrophysiological recordings of LD in multiple studies by different research groups (Mota-Rolim, 2020).
Despite five decades of EEG research on the topic, these studies still paint an inconclusive picture of LD neurophysiology (Baird et al., 2019, 2022; Zerr et al., 2024): earlier findings such as increased occipital alpha (8–12 Hz; Ogilvie et al., 1978), parietal beta (13–19 Hz) (Holzinger et al., 2006), or frontal gamma (36–45 Hz) underlying LD (Voss et al., 2009) could not be substantiated after adequate removal of non-neural artifacts (Baird et al., 2022). One independently replicated finding is that reduction in delta-band [0.5–2.9 Hz (Dodet et al., 2015); 2–4 Hz (Baird et al., 2022)] activity differentiates LD from non-LD, potentially pointing to alterations in local sleep depth.
LD is a learnable skill (La Berge, 1980a; Stumbrys et al., 2012); however it still cannot be reliably induced (Tan and Fan, 2023), leading to very small sample sizes in most studies. The rarity of LD in general population (Mota-Rolim et al., 2013; Saunders et al., 2016) and thus lack of statistical power in the typical nonrepresentative studies may be overcome by aggregating data from multiple studies or recording sites, which however requires considerable effort in harmonization of datasets. Here we developed an adaptive preprocessing pipeline for valid EEG data aggregation tailored to LD, addressing the field's inherent confounds such as saccadic artifacts and assumptions of signal nonstationarity in distinct EEG layouts, which we detail next.
A particular concern in LD recordings is saccadic artifacts: higher eye movement density during LD (LaBerge, 1990) will be reflected in frontal potentials if not properly accounted for (Baird et al., 2019). A recent study (Baird et al., 2022) demonstrated that a unique signature of LD found in an earlier study (Voss et al., 2009), namely, frontolateral gamma activity at 40 Hz, could be explained by improper cleaning of saccadic potentials (SPs), which are resistant against traditional, regression-based cleaning approaches (Keren et al., 2010; Baird et al., 2022). SPs are minor spike deflections in the horizontal EOG (Yuval-Greenberg et al., 2008; Keren et al., 2010; Baird et al., 2022) and noticeable just before saccadic eye movements, especially in the direction opposite to the saccade's trajectory. While independent component analysis (ICA) has been shown to clean SPs that happen during LD (Baird et al., 2022), ICA may overclean data when applied to few EEG channels (Liu et al., 2020). As an alternative, we introduce a preprocessing solution that works for low-density EEG setups and hence is applicable for analyzing data from wearable devices (as long as the device includes at least six EEG channels) intended for home use (Jafarzadeh Esfahani et al., 2024) as well as polysomnographic recordings in the laboratory.
Another issue for deriving electrophysiological LD correlates is signal nonstationarities likely encountered in the cognitively complex computations of the LD state (Mota-Rolim et al., 2010). REM sleep, in particular, is inherently nonstationary, as it can be subdivided into phasic and tonic periods with distinct spectral signatures (Cantero et al., 2003; Simor et al., 2020). While nonstationarity exists to some degree in all EEG recordings, it becomes particularly pronounced in transient cognitive states including psychedelic experiences and transitions under anesthesia (Li et al., 2022; Ort et al., 2023). However, LD presents a unique challenge beyond cognitive fluctuations: voluntary eye signaling. These preinstructed eye movements introduce abrupt signal deviations, adding an additional layer of nonstationarity not typically observed in other transient conditions. Many established tools, such as ICA, often assume stationarity (Delorme and Makeig, 2004) and risk overcleaning when spectral aspects shift dynamically. Our preprocessing protocol is designed to mitigate this issue.
After harmonizing the preprocessed datasets originating from diverse EEG setups, we investigated spectral correlates and functional connectivity patterns across traditional frequency bands, as well as entropy, complexity, and fractal dimension measures, to discern how LD differs from both REM sleep and the waking state. Additionally, we explored the evolving dynamics of LD over time, focusing on the neural changes around the eye movements signaling lucid insight (Baird et al., 2019).
Importantly, our study pioneers the use of electrophysiological source localization in LD, leveraging high-density sensor data to uncover the neural correlates of the LD state through spectral analysis of neuroanatomical structures. Given the inconclusive state of the field so far, we approached our analyses without specific hypotheses about spectral bands or brain regions affected, thus providing a comprehensive exploration of the electrophysiological correlates of LD. Figure 1 provides a graphical summary highlighting the key findings and significance of this study.
Summary of key electrophysiological findings during LD, including increased gamma activity in the precuneus around initial lucidity eye signaling activation, as well as enhanced frontal gamma and posterior alpha connectivity compared with REM sleep.
Finally, it is important to highlight that since LD is characterized by an increase in self-awareness during dreaming, its study has potential to clarify the mechanisms of self-awareness during waking, in mental diseases, as well as in other states of consciousness (Mota-Rolim and Araujo, 2013), such as altered states induced by psychedelics, which has many similarities with dreaming, and specially with LD (Kraehenmann, 2017). LD also presents metacognitive abilities that are found during meditation, which explains why LD and meditation are related (Gerhardt and Baird, 2024). Another relevant aspect of LD research is its clinical potential for treating recurrent nightmares (for review, see de Macêdo et al., 2019; Ouchene et al., 2023) and understanding consciousness, such as during anesthetic procedures or disorders of consciousness (Chennu and Bekinschtein, 2012), because during LD subjects perform a voluntary signal to indicate they are aware, as said before (Mota-Rolim, 2020).
Materials and Methods
Participants
The intrinsic rarity of LD necessitates collaboration across laboratories to gather a sample of adequate size (Zerr et al., 2024). We collected and gathered LD data from multiple laboratories: the Donders Center for Cognitive Neuroimaging at Radboud University (RU; 12 recordings), University of Osnabrück (UO, 5 recordings), Max Planck Institute (MPI; 10 recordings), Brain Institute at Federal University of Rio Grande do Norte (UFRN; 2 recordings), and Department of Psychology, Stanford University (SU; 27 recordings). We obtained written signed informed consent prior to recording from all subjects, and ethical approval was acquired from the corresponding university institutional review boards (approval codes CMO 2014/288, 4/71, 043.5, 094–13, 061/2008 for RU, UO, MPI, and UFRN, respectively). Details regarding the dataset from SU can be reached in the previous study (S. LaBerge et al., 2018). We excluded data with <6 EEG channels, continuous eye signaling during LD, unclear lucidity eye signals, having no possibility to capture two distinct REM conditions, and containing persistent noise that we could not clean up. This led to the exclusion of 12 recordings, with the final sample consisting of 26 subjects (25.0 ± 5.1 years, 20 females). Of those, five subjects contributed two nap and night sessions each, one contributed three nap sessions and the other one subject contributed 12 nights, yielding a sample size of 44 sleep recordings: 5 recordings consist of 6 channels, 20 recordings consist of 29 channels, and the remaining 19 are high-density recordings, with 10 having 64 electrodes and 9 having 128 electrodes. While 44 recordings were initially selected, one recording was later identified as an outlier based on its spectral properties [see power spectral density (PSD) analysis and Extended Data Fig. S1 for details]. Thus, subsequent analyses were conducted on the remaining 43 recordings.
Data recording
Polysomnography recordings started when participants went to bed. Bed times differed across labs: with regard to data collected at RU, participants went to bed in the morning hours after slight sleep deprivation, whereas at UO, MPI, UFRN, and SU, full nights were recorded. For this, we used different EEG devices and setups, namely, at UO, a six-channel SOMNOscreen PSG device utilizing a 10/20 layout (F3, F4, C3, C4, O1, O2); at RU, a 64-channel actiCAP active electrode EEG system (Brain Products) with a 10/10 layout; at MPI, a 128-channel easyCAP passive electrode EEG system (Brain Products) featuring a 10/05 layout; at UFRN, a 32-channel EMSA BNT-36 passive electrode EEG system with a 10/20 layout; and at SU, a 32-channel passive electrode system (Neuroscan) containing 29 EEG electrodes placed in a 10/20 layout. After each sleep session, free reports of the last remembered dream were collected via audio recordings and later transcribed by researchers.
Task
Before all sleep sessions, we instructed participants to move their eyes to the left, then to the right, and repeat this sequence at least once as soon as they became aware that they were currently dreaming. If successful, this well established maneuver yielded a LRLR signal during sleep (Mota-Rolim, 2020) which can be easily discerned in horizontal EOG recordings (Fig. 2). Beyond this common task, protocols varied between subjects and labs, encompassing flying, counting, two-way communication, motor movement (hand clenching), and mind-wandering tasks. Given the diversity of event protocols across datasets during LD, there is a wide range of eye signaling patterns, from initial and final LRLR sequences to multiple intermediate eye movements that mark the beginning and end of specific events. The duration of these eye signals also varies, ranging from brief movements signaling event onsets to longer eye movements that are considered voluntary ocular events. See Extended Data Figure S2 for EEG segments capturing LD episodes contrasted with wakefulness across all recordings.
A, Time table of conditions over sessions. The top section shows the individual distributions of condition on- and offsets (colored bars) with respect to a standardized timeline, where 0% denotes recording onset and 100% recording offset. Individual recordings are sorted along the y axis with regard to the LD onset, showing how most LD episodes are collected toward the end of a recording session. Also note that waking condition intervals occur early in the timeline as this was the main selection criterion for this condition (see text for details). The bottom section depicts the distribution of each condition segment's center over the standardized timeline using box plots and individual data points. A notable gap between early and late REM states, accounting for ∼50% of total sleep duration, indicates the passage of one or more sleep cycles during both nap and full-night experiments. The LD segment is positioned close to the later REM phase, with selection slightly earlier on average. B, Overview of the EEG preprocessing, postprocessing and analysis protocol employed in this study. Note that this protocol can handle and integrate diverse EEG setups to allow robust neural correlation analyses. C, Illustration of amplitude and waveform characteristics of EEG, EOG, and EMG activities between LD and wakeful states, showcasing representative segments from four participants. LD, lucid dreaming; REM, rapid eye movement sleep; PREP, early-stage EEG processing pipeline; ASR, artifact subspace reconstruction; SSP, signal-space projection; PSD, power source density; topo, topography.
Selection of conditions
All sleep data were scored either by at least one human expert or by the U-Sleep v2.0 (Perslev et al., 2021) automated model, using 24 sparse EEG channels along with an additional EOG channel. We trained human raters to score the data according to the criteria of the second version of the sleep scoring manual by the American Association of Sleep Medicine. Accordingly, we labeled continuous polysomnographic recordings within each 30 s nonoverlapping interval. Sleep scoring was done blind with no auxiliary information other than the PSG data.
REM and waking segments were chosen for quantitative comparisons with LD segments within the same sleep session. We selected time intervals from each session recording that matched the duration of lucid REM segments for comparison with the states of waking, nonlucid early REM (E. REM), and nonlucid later REM (L. REM). Two such temporally distant REM intervals were selected to test the stability of potential LD correlates against confounding effects of nighttime (e.g., due to changes in sleep pressure; Preller and Vollenweider, 2016). Where possible, we aimed to minimize potential confounding effects of nighttime variation by selecting REM intervals from distinct points in the night, ideally separated by at least one sleep cycle or intervening stages. However, in some recordings, REM segments did not have significant temporal distance, and E. REM and L. REM segments were selected accordingly. Noise levels were assessed by calculating the standard deviation of each segment, with higher standard deviations indicating greater signal variability likely due to transient artifacts (e.g., muscle activity, movement). To ensure stable neural signal extraction, we selected segments with lower standard deviation, as they represent more consistent brain activity with reduced contamination from non-neural sources. We further conducted a visual inspection of the selected segments in the temporal domain to ensure no significant artifacts were included, namely, muscle activity, movement-related distortions, and transient high-amplitude noise. This simplified approach aimed to maintain the natural spacing between early and later REM phases for accurate comparison, prioritizing intervals that offered clear distinctions in REM sleep stability and duration.
A schematic overview of the project including the visual depiction of conditions for each subject with the overall distributions across normalized sleep duration is shown in Figure 2, A and B, depicts an overview of the preprocessing and analysis protocol. Note that for the waking condition, we primarily selected data at the beginning of the experiment, specifically during periods when the subjects' eyes were closed. This strategy aimed at choosing a waking state where subjects were relaxed, not conversing, and free from physically induced artifacts. Ensuring the waking represented a relaxed state was crucial for the accuracy of our analysis.
Multistage preprocessing protocol
Our multistage preprocessing protocol was specifically designed to mitigate saccadic artifacts and signal nonstationarity, which are potential confounds that are especially important to consider in electrophysiological measures of LD. SP artifacts are particularly prevalent and problematic in the context of LD research due to the eye movements characteristic of this state. The ocular movements during LD are both involuntary (Baird et al., 2022), the rapid eye movements that define REM sleep, and voluntary by instructing the subjects to perform eye signals in order to be able to determine their lucidity state and subsequent events (Mota-Rolim, 2020). A recent study (Baird et al., 2022) suggested that ICA successfully cleansed the SP artifacts during LD, but this was achieved with relatively high-density EEG data (32 channels). Importantly, ICA is sensitive to spatial sampling (Liu et al., 2020) and therefore does not provide robust cleaning for low-density variations (e.g., six channels). In this study, we have developed an alternative multistage preprocessing protocol that addresses nonstationarity of LD EEG and allows for inclusive analyses with low-density EEG setups (Fig. 2B). The pipeline is adaptive given that with minimal parameter tuning effort, it offers semiautomatic signal cleaning in a data-driven manner. By doing so, we could standardize the data across diverse EEG setups (in both low- and high-density layout variations) and, in later stages, unify them to enable large-sample-size analyses of the neural correlates of LD.
After standardizing the channel labels across datasets and assigning channel types (EEG, EOG, EMG, ECG), we applied the three-stage preprocessing protocol consisting of the PREP pipeline (Bigdely-Shamlo et al., 2015), artifact subspace reconstruction (ASR; Mullen et al., 2013; Miyakoshi, 2023), and signal-space projection (SSP; Uusitalo and Ilmoniemi, 1997), each taking unsegmented, continuous data as input in order to avoid condition-based biases. Note, however, that before running the algorithms, we cropped the data between the onset of the earliest condition (which most often was E. REM; Fig. 2A) and the offset of the latest condition (most often LD) in order to save computation time and focus on the signals of interest.
Stage 1 (PREP pipeline)
The PREP pipeline offers standardized early-stage (preartifact rejection) EEG preprocessing ensuring consistency across diverse datasets (Bigdely-Shamlo et al., 2015). Apart from its robust line-noise removal capabilities, PREP effectively resolves a major circular problem of EEG re-referencing (identifying bad channels depends on a good average reference, while a good average reference presupposes bad channel exclusion), which is accomplished by iteratively detecting and excluding bad channels and thus refining the average reference with each iteration. We further improved bad channel detection in high-density setups via enabling the included RANSAC (random sample consensus) options which requires at least 19 channels. Prior to employing PREP, we applied a 1–49 Hz IIR bandpass filter to the data. For applying these functions in this study, we employed PyPREP (Appelhoff et al., 2022) within a Python 3.9 environment.
Stage 2 (ASR)
ASR, a nondestructive, dynamic method for cleaning multichannel EEG artifacts, was selected for its data-driven, adaptive nature (Mullen et al., 2013) that effectively managed the discrepancies from diverse artifact sources and referencing setups in our dataset. Before applying this algorithm, we high-pass–filtered data at 2 Hz in accordance with Baird et al. (2022) and following empirical recommendations for nondestructive EEG cleaning procedures [ASR (Chang et al., 2020) and ICA (Winkler et al., 2015; Bigdely-Shamlo et al., 2020; Klug and Gramann, 2021)]. ASR trains on a clean EEG segment to establish a baseline model of clean signals, which is then used to identify deviations in test data. Training data were selected as follows: First, we divided the data into 300 s consecutive intervals with 75% overlap. For each interval, standard deviations were calculated within each channel and summed across channels. The four intervals with lowest sums were then concatenated, yielding 20 min of training data for each sleep EEG recording. We chose to include multiple intervals in order to offer varying levels of stationarity for the training procedure, which should yield more robust and versatile results. Using sample covariance as an estimator, detected artifacts were cleaned within 2 s intervals with 33% overlap. These parameters offer enough precision to clean artifacts in a sleep stage-specific manner. We decided against using a newer, Riemannian version of ASR (rASR; Blum et al., 2019) for reasons of efficiency and visual inspection which revealed remarkably better cleaning of SPs in the standard version. ASR was applied using respective functions in the Python package meegkit within a Python 3.9 environment.
Stage 3 (SSP)
While PREP and ASR handled the majority of artifacts resulting from SP and other miscellaneous ocular and heart rate-related activities (e.g., potentially increased heart rates due to motor performance during LD; Erlacher and Schredl, 2008), visual inspection of the data suggested they did not entirely address all aspects of EOG and ECG artifacts (Fig. 3C). To resolve this problem, we then used SSP, a spatial filtering technique that leverages all available biopotential sensors including EOG and ECG for nondestructive artifact cleaning. More specifically, it minimizes artifact interference in EEG data by utilizing EOG and ECG channels to identify distinct EOG and ECG artifact patterns, computing orthogonal projection vectors for these, and transforming the EEG data into a subspace where artifact influence is reduced. SSP was applied via respective functions in the MNE-Python toolbox (Gramfort, 2013).
A, Example comparison of differently regularized Hilbert series of HEOG signals with regard to SP detection sensitivity and specificity. Note that derivative-based transforms are sensitive to non-SP–related activity as shown by the highlighted peaks. B, Thresholding of polynomial regularized Hilbert series of HEOG detects SP events. In this example, a 99.9 percentile was selected. C, Exemplary effects of SP events during LD on ERSP effects at channel F7. Better cleaning of broadband SP artifacts, especially below 48 Hz, is shown with each cumulative preprocessing step. D, Grand-average ERSPs of SP events to demonstrate SP cleaning across cumulative preprocessing steps. The left two columns depict recordings from 64-channel active electrode setups, the right two columns from 128-channel passive electrode setups. The left and right columns of each column pair show data from channels F7 and F8, respectively. E, Top row, Grand-average butterfly plots and GFPs around SPs for each cumulative preprocessing step. Bottom row, Topographies of SP artifacts for each cumulative preprocessing step in the ∼40 Hz band [36–45 Hz (27, 30)]. Left panel, 64-channel active electrode setup. Right panel, 128-channel active electrode setup. A remarkable flattening of SP-related GFPs can be shown in all of these setups. HEOG, horizontal electro-oculogram; SP, saccadic potential; LD, lucid dreaming; ERSP, event-related spectral perturbation; GFP, global field power.
Saccadic spike potential detection
Cleaning saccadic spike potential artifacts (SP) represents a necessary step toward identifying reliable electrophysiological markers of LD, requiring a valid identification of SPs (for details, please refer to Methods and Materials). The utilization of different horizontal electrooculogram (HEOG) data transforms for SP detection on example data is shown in Figure 3, A and B. The introduction of the Hilbert transform enhanced our ability to detect saccadic spikes by amplifying SPs and reducing the impact of nonsaccadic high-frequency spike artifacts. After visual inspection of all HEOG data in pre-LD and LD intervals and after comparing first- and second-order derivatives and second-order polynomial values, we determined that polynomial fitting yielded the best approximation as a regularizer of Hilbert peaks. Certain artifacts resembling SPs, but exhibiting slightly lower amplitude, appeared in between and independent of saccadic activities. Slope regularization was less prone to such false-positive SPs, which is why we opted for polynomial fitting as the regularizer for the Hilbert series of the HEOG data (Fig. 3A). In general, we found that thresholding data above the 95% percentile could optimally detect false positives as above-threshold peaks while keeping the number of misses minimal. Often, however, the actual percentile cutoff was set much higher (Fig. 3B). The formula of the proposed SP detection algorithm is shown in Equation 1 as follows:
Saccadic spike potential cleaning
Before evaluating the efficacy of our preprocessing steps with regard to cleaning SP artifacts, we first required a valid identification of SPs. For this, we investigated continuous 100 s HEOG data intervals from every night recording that included both initial and final LRLR signals, ensuring equal padding before the first and after the last signal. This selection was chosen to provide prominent eye movements that included LRLRs for later validation of SP detection. Several filters were applied to these HEOG data in order to test which of them best help identify time points of SPs. Specifically, we applied a bandpass filter between 30 and 100 Hz as suggested previously (Keren et al., 2010) with a subsequent Hilbert transform; a Hilbert transform regularized with a second-order polynomial fit; a first-order derivative; and a second-order derivative. We chose the first- and second-order derivatives and the second-order polynomial fitting as regularization factors after Hilbert transform, normalizing the factor domains by dividing them by their maximum value, which scaled from 0 to 1. The polyfit regularization is less likely to get affected by other miscellaneous artifact-based peaks. This is because polyfit performs a slope approximation to regularize the Hilbert series by quantifying the negative or positive dumping factors that occur after the SP event. Accordingly we selected a polyfit regularizer to find SP events. This was done via visual inspection, opting for conservative selection in order to retain potential signals of interest. Once this filter was selected, we manually selected a threshold percentage that most precisely identified SPs. Systematic differences in HEOG electrode placement between recordings from different datasets necessitated setting individual threshold values for each dataset. The maximum value of data above the threshold then pointed to the time point of putative SPs. Before further validation of our preprocessing/cleaning pipeline, we manually discarded false positives (i.e., above-threshold filter peaks that are not associated with a saccade).
To evaluate the steps of our preprocessing pipeline with regard to cleaning SP artifacts, we investigated the F7 and F8 channels as a scalp position common to all datasets with high frontolateral ocular reflections. Event-related potentials were calculated for each cumulative step in the pipeline (i.e., PREP; PREP followed by ASR; PREP followed by ASR followed by SSP). In the time domain, we investigated Hilbert envelopes across all EEG channels as well as SP-related GFP fluctuations. In the time–frequency domain, we calculated event-related spectral perturbation (ERSP) with multitapers (frequency resolution, 0.1 Hz; time bandwidth value, 8). No logarithmic transformation was applied in order to preserve high contrast values. While our statistical analyses did not exceed 50 Hz, for this validation step, we calculated frequencies up to 100 Hz for a more sensitive picture of the effects of our preprocessing steps.
Epoching and postprocessing
We segmented the cleaned subject data into 4 s windows with a 2 s overlap (i.e., 50% padding) for four conditions (E. REM, L. REM, LD, and waking). Despite our multistage preprocessing protocol largely eliminating SP artifacts, we conducted manual inspections of the epoched data for residual miscellaneous and movement-related artifacts, discarding contaminated epochs. We restricted all analyses regarding neural correlates of LD to the spectral domain between 2 and 48 Hz in order to avoid spectral leakage from bandpass filtering.
Nontopographical global analyses
For the nontopographical analysis, we analyzed six common channels (F3, F4, C3, C4, O1, O2) from all 43 subject data, ensuring a maximum sample size and consistent scalp position contributions. Since all other (topographical) analyses require higher spatial sampling, we selected a subset of 19 participants that share 59 channel positions and limited our analysis to these 59 channels in order to balance topographical detail with the sample size and comparability. This decision necessitated the exclusion of electrodes “TP10,” “Iz,” “FT10,” “FT9,” and “TP9” from 64 channels EEG layout, as some subjects' data did not include readings from these electrodes.
Spectral power analysis
We selected epochs unified across different EEG setups. For each of the six channels and epochs, we extracted multitaper PSD (mPSD) values ranging from 2 to 48 Hz and applied a dB conversion (10 times the base-10 logarithm of the mPSD). We averaged the values across channels and epochs and extracted the 95% confidence bands for the conditions. Furthermore, we averaged PSD values within six different frequency ranges (delta, 2–4 Hz; theta, 4–8 Hz; alpha, 8–12 Hz; beta, 12–30 Hz; gamma1, 30–36 Hz; gamma2, 36–45 Hz). We split the gamma band into gamma1 (30–36 Hz) and gamma2 (36–45 Hz) to account for the differential effects of saccadic artifacts and to isolate cleaner neural activity related to LD. Previous research has demonstrated that 40 Hz EEG activity can be strongly influenced by ocular saccades (Weinstein et al., 1991; Yuval-Greenberg et al., 2008), which may influence REM sleep and in particular LD recordings (Voss et al., 2009). By separating gamma2 (36–45 Hz) denoted as 40 Hz band (Voss et al., 2009), which includes this artifact-prone range, from gamma1 (30–36 Hz), we aimed to more precisely distinguish intrinsic gamma activity linked to lucidity from artifact-related contributions. In addition to predefined band-wise analysis, we performed a mixed-frequency analysis by directly computing dB-based power differences between LD and each of the other three conditions (E. REM, L. REM, and wake) across the full frequency spectrum, allowing for a data-driven exploration of spectral changes without constraints of predefined frequency bins (Donoghue et al., 2022).
Entropy and complexity analysis
In accordance with previous research (Baird et al., 2022), we attempted to extract consciousness levels in EEG time series via Lempel–Ziv complexity (LZc; Zhang et al., 2001) and two selected entropy markers—permutation entropy (PE; Bandt and Pompe, 2002) and sample entropy (SE; Delgado-Bonal and Marshak, 2019). LZc is a nonlinear dynamic measure that quantifies the rate of new pattern generation in a time series, making it a valuable tool for assessing the complexity of EEG signals, which are known for their nonlinear and nonstationary nature (Liyanage et al., 2013). PE offers a robust and computationally efficient way to capture the order relations between values in a time series, providing insights into the dynamical changes in EEG data (Bandt and Pompe, 2002). We derived condition-specific estimates via averaging over the channel-specific values. The formulas of PE and SE are shown in Equations 2 and 3, respectively, as follows:
Fractal analysis
The Higuchi fractal dimension (HFD; Harne, 2014) is a mathematical method used to quantify the fractal properties of time series data. HFD measures the complexity and self-similarity of a signal, providing a single value that represents the fractal dimension of the data. The HFD
D of a time series
Outlier verification test
Extreme values were identified in two samples across five distributions in both PE and SE. These outliers were detected through visual examination of smoothed density estimates in raincloud plots, showing significant deviations from the expected distribution patterns. Notably, extreme values were found in PE during E. REM and L. REM, along with SE. Additionally, wake conditions also exhibited outlier values in both PE and SE. To statistically validate these outliers, we applied a z-score–based tail probability test, yielding significant results (p < 0.001). The z-score and tail probability calculations are shown in Equations 5 and 6, respectively, as follows:
Sensor-level topographical power analysis
We applied current source density (CSD; Perrin et al., 1987) transformation with stiffness value of 0.081 to raw data prior to performing multitaper spectral density estimation for topographical power analysis. The CSD method enhances spatial resolution by minimizing volume conduction effects and improving the localization of underlying neural activity on the scalp level. For each subject and condition, we computed mPSD across various frequency ranges. We derived subtracted power values from these contrasts and normalized each contrasted dataset across all channels, standardizing amplitudes by maximum division.
Surface-based source reconstruction
To further investigate neural activation distinguishing LD from waking and non-LD, we employed source reconstruction. To benefit from its excellent noise suppression characteristics, we planned to use beamforming approaches (Van Veen et al., 1997; Gross et al., 2001); however, given that we are combining EEG data with template MRIs, there are concerns whether our forward models would be of sufficient quality for beamforming (Westner et al., 2022). Indeed, comparing the results to the more robust source reconstruction approaches dynamic statistical parametric mapping (dSPM; Dale et al., 2000) and exact low-resolution electromagnetic tomography (eLORETA; Jatoi et al., 2014) consolidated these concerns, with the beamformer portraying a strong bias toward the center of the brain. Thus, we only report the results of the two minimum norm estimation approaches, which are less impacted by forward model inaccuracies. We opted to employ both methods in a complementary manner. We recognized that dSPM, with its noise normalization feature, excels in highlighting brain regions with significant activity by effectively reducing background noise. This makes it particularly useful for detecting subtle changes in brain activity across different conditions. On the other hand, eLORETA offers enhanced spatial accuracy in minimizing localization errors assuming a smooth distribution. Furthermore, applying both dSPM and eLORETA allowed us to cross-validate our findings, increasing our confidence in the robustness of our results. See Extended Data Figure S4 for grand-average distributions of cortical power per condition and frequency band.
The initial step in our pipeline involved applying average re-referencing to all subject datasets. For cortical mapping, we selected the “fsaverage-ico-5-src.fif” from the boundary element model as our source space template, ensuring a high-resolution cortex representation. The forward solution was calculated with a minimum distance (mindist) parameter set to 5.0 mm minimum distance constraint. To calculate noise covariance, we developed an algorithm to identify baseline segments from the entire sleep dataset for each subject. This algorithm evaluated the standard deviation (STD) of 10 s windows, with an 80% overlap, across the raw data. The segment with the lowest STD, sufficiently distanced from all experimental conditions to avoid overlapping neural processes, was selected as the baseline. This baseline segment was then used to compute the noise covariance, which informed the creation of the inverse operator, applying a depth parameter of 0.8 to fine-tune source localization. For source PSD estimation, we set the signal-to-noise ratio (SNR) parameter at 3, determining the lambda value for regularization. We utilized both dSPM and eLORETA as our inverse methods to achieve cortical power estimations. These analyses resulted in high-resolution cortical power maps, with a total of 20,484 spatial points.
Functional connectivity analysis
In this analysis, we adopted a distributed source estimation approach, focusing on the cortical surface using both dSPM and eLORETA methods and the same standard fsaverage source template we used for surface estimation. The inverse solution involved creating an inverse operator with the forward solution and noise covariance model (calculated from the selected baseline segments of each raw data), setting the loose parameter to 0.2 and the depth parameter to 0.8, to balance localization accuracy and depth weighting. We then computed the inverse solution with an SNR of 3.0, applying dSPM and eLORETA with the orientation set to normal, targeting the component perpendicular to the cortical surface. Subsequently, we partitioned the source coordinates into 68 cortical labels (34 per hemisphere) using FreeSurfer (Fischl, 2012), aparc (Desikan et al., 2006), and cortical parcellation, calculating the time courses for these labels in the mean-flip mode to minimize signal cancelation. For the connectivity analysis, we utilized both the phase lag index (PLI; Stam et al., 2007) and the debiased weighted PLI (wPLI debiased; The BASIS Team et al., 2014) to measure phase synchronization, thereby enhancing our investigation of spectral connectivity. This analysis was conducted through a multitaper approach, incorporating adaptive multitapering to improve the accuracy of spectral estimation. In the absence of a predefined adjacency sparse matrix to use for cluster-permutation testing, we used a radius neighbor graph (79) algorithm to define parcel adjacencies. This algorithm constructs a graph based on the neighborhood radius around each parcel.
Oscillatory activation extent of lucidity
In addition to our contrast-based analyses, we focused on the LD segment to explore the spatiotemporal extent of spectral activities concomitant with LD that extend to both sides of the initial eye signaling. To examine activations at both sensor and source levels, we selected high-density EEG data (59 common channels from 19 subjects). We first identified sensor-level time intervals containing significant activations, which then informed search of these activations at the volumetric source level.
Sensor-level global field power around eye signaling onset
For each subject, we extracted a 30 s temporal window centered around the initial lucidity eye signal (LRLR), using the first 5 s as a baseline. We derived temporal power fluctuation patterns via computing absolute global field power (GFP) in this window for each channel and frequency band, normalizing the 25 s GFP patterns to this baseline with z-score standardization. To increase the sensitivity of detecting the potential significant time intervals of GFP fluctuations, we further applied downsampling to the GFP data by factor of 5. This adjustment aimed to reduce the complexity of the temporal domain, thereby enhancing the computational efficiency of the spatiotemporal permutation cluster test. Additionally, in the context of a small sample size (n = 19), downsampling contributed to a more robust statistical analysis by diminishing the impact of temporal noise and reducing the risk of Type 1 errors, ultimately making it easier to identify periods of significant neural activity. Since GFP represents absolute values, sensitivity to detect any activation pattern regardless of polarity is enhanced. Using region of interest (ROI) separation, we also analyzed GFP data from channels in frontal left, frontal right, central left, central right, parietal left, and parietal right regions separately within each frequency range. ROIs defined with specific channels: frontal left (Fp1, AF7, AF3, F7, F5, F3, F1, FT7, FC5, FC3, FC1), frontal right (Fp2, AF4, AF8, F2, F4, F6, F8, FC2, FC4, FC6, FT8), central left (T7, C5, C3, C1, TP7, CP5, CP3, CP1), central right (C2, C4, C6, T8, CP2, CP4, CP6, TP8), parietal left (P7, P5, P3, P1, PO7, PO3, O1), and parietal right (P2, P4, P6, P8, PO4, PO8, O2). These 25 s GFP data were normalized to the baseline as well. To determine the common extent of lucidity within the 25 s interval of each frequency band, we performed spatiotemporal cluster-based permutation tests on preprocessed EEG data.
Source-level decomposition of lucidity activation
We analyzed source-level activity within frequency bands that corresponded to time intervals of significant sensor-level activations. To accomplish this, we utilized dSPM and eLORETA for surface-based source reconstruction, focusing on cortical estimations during the 5 s baseline and the GFP-defined temporal segments that exhibited significant activation relative to the baseline. The normalization of these estimates involved subtracting baseline values from activation segments, followed by division by the maximum absolute value of the contrasted data, thus scaling the differences within a normalized range.
Additionally, we conducted source-level functional connectivity analysis using directed PLI (dPLI; Stam and Van Straaten, 2012) on the significantly activated temporal segments. Unlike PLI or wPLI, which we used for state comparisons with clearer distinctions (e.g., LD vs other conditions), dPLI provides directional phase relationships, making it particularly suitable for capturing transient changes in connectivity dynamics. Additionally, dPLI is less sensitive to volume conduction and is particularly suited for capturing reconfigurations in functional connectivity that accompany cognitive state transitions. As noted in previous work (Stam and Van Straaten, 2012), dPLI has been shown to effectively characterize connectivity shifts between resting and activation states, supporting its application in analyzing the transition from prelucid to LD. To quantify these changes, we applied percent normalization by subtracting baseline connectivity values from activation segment values and dividing by the baseline, ensuring a standardized measure of connectivity modulation.
Statistical analyses
Statistical verification of SP cleaning across three different EEG setups (with 6, 64, and 128 channels) involved comparing the averaged short-window (600 ms) time–frequency power (dB) evoked by SPs at the beginning (i.e., raw data) and end of the pipeline (i.e., PREP followed by ASR followed by SSP). We selected channel F3 for the six-channel setup and F7 for both the 64- and 128-channel setups. Our null hypothesis posited that there would be no significant increase in SP power in the raw data compared with the preprocessed data. We used the Shapiro–Wilk test to check normal distributions of each condition. If no normal distributions can be shown, we would use the Wilcoxon signed-rank test and Student's paired t test otherwise.
In all further tests, we included “condition” (LD, E. REM, L. REM, waking) and “frequency band” (delta, 2–4 Hz; theta, 4–8 Hz; alpha, 8–12 Hz; beta, 12–30 Hz; gamma1, 30–36 Hz; gamma2, 36–45 Hz) as independent variables. For low-density analyses (regarding PSD, entropy, and complexity), we utilized linear mixed models (LMMs; Bolker, 2015). Model estimation was performed in JASP (Love et al., 2019), version 0.17.3. Initially, both random intercepts and random slopes were considered; however, model diagnostics indicated singular fits when including random slopes, necessitating their exclusion for model stability. Consequently, the final model retained random intercepts only, effectively accounting for intersubject variability while ensuring a robust fit. To mitigate the risk of Type 2 errors, we applied a parametric bootstrap method with 10,000 sample estimations, integrating subject IDs as random effect grouping factors. Subsequently, we conducted post hoc pairwise comparisons, adjusting for multiple comparisons using the Bonferroni’s correction. Effect sizes were reported using marginal R2, capturing the variance explained by fixed effects, and conditional R2, incorporating both fixed and random effects. For high-density (topographical) analyses, we constrained statistical testing via paired one-sample cluster–permutation tests (Maris and Oostenveld, 2007) with two-tailed approach. For mixed-frequency power analysis, we applied pairwise permutation cluster testing with an adaptive cluster-forming threshold. We applied threshold-free cluster enhancement (TFCE; Smith and Nichols, 2009) to inductively determine cluster thresholds [start/step of 2/0.2, 1.4/0.02, 2/0.2 for surface-based power estimation of a priori intervals, functional connectivity analyses (to assess both intercondition differences and activation patterns during LD eye signaling relative to the baseline), and GFP-guided surface reconstruction, respectively], except for sensor-level power analyses and spatiotemporal clustering test on GFP, for which we set the thresholds of 2.878 and 2.095 based on the T-distribution percent point function for α < 0.025. Furthermore, for sensor-level topographical power analysis, we used a step-down function of 0.001 and 10,000 permutations, whereas for all other analyses, 1,000 permutations and no step down were applied. While TFCE enhances our capacity to detect true effects by integrating both the spatial extent and the intensity of signals—thereby minimizing the risk of Type 1 errors—in analyses conducted with a two-tailed approach, we further evaluated the clusters with the adjusted significance threshold. This adjustment was specifically implemented to control the false alarm rate, thereby introducing an additional layer of Type 1 error correction for the selected tests. LD was contrasted with E. REM, L. REM, and waking. In addition, we contrasted both nonlucid REM segments against each other, leading to a total of four comparisons per high-density analysis and frequency band. For PSD and entropy analyses, false discovery rate correction with Benjamini–Hochberg method was applied to control for multiple comparisons. To identify outliers in our entropy measures, we employed a z-score–based tail probability test. This test calculates the z-score for each value, representing the number of standard deviations a data point is from the mean. By determining the tail probabilities associated with these z-scores, we could quantify the likelihood of these values being extreme deviations from the overall distribution. A p value threshold of 0.001 was used to designate significant outliers. Cluster-based permutation tests were implemented in Python between versions of 3.0 and 3.9 with the MNE package (Gramfort et al., 2014).
Results
Saccadic spike potential cleaning
Figure 3C displays an example time series of a 100 s interval around left→right→left→right eye signaling (LRLR) around one LD episode. The raw HEOG is shown in the top row, while ERSPs of progressive preprocessing steps are displayed in the rows below. From the raw data (first ERSP row), three distinct types of SP-induced artifacts are visible in the time–frequency domain: First, higher SP-correlated power can be shown above our analysis cutoff of 50 Hz, compared with the power below 50 Hz. Second, even higher SP artifact-related power in low-frequency bands (below 5 Hz) can also be seen. Lastly, artifactual broadband effects are visible as vertical stripes of elevated activity. After the first preprocessing step [PREP (Bigdely-Shamlo et al., 2015); second ERSP row in Fig. 3C], we see the eradication of artifactual broadband activity and some diminishing of artifactual low-frequency activity. Since PREP is not designed to clean artifacts in EEG, this initial cleaning should be improvable by further preprocessing steps. Indeed, the addition of ASR (Mullen et al., 2015; third ERSP row) visibly reduced artifactual low-frequency activity even further but leaves some bursts of activity that are correlated with SP events (Fig. 3C, vertical dashed lines), especially between 30 and 45 Hz. Considering that up to this point only EEG-based cleaning had been performed, subsequent SSP based on HEOG patterns could remove all remaining types of SP artifacts as expected.
Figure 3D shows the effects of the cumulative preprocessing steps on grand-average SP–induced ERSPs of frontal channels. In both medium- (64 channels; active electrodes; 16 SP events) and high-density (128 channels; passive electrodes; 63 SP events) EEG setups, each additional preprocessing step progressively cleans phasic activities correlated with SPs but retains unrelated tonic activity, which is most apparent in the retained contrast between frequencies below and above 10 Hz. Statistical analyses support these observations: Since the values are normally distributed [Shapiro–Wilk, p = 0.106; 0.996; and 0.208 for 6-channel (14 SP events), 64-channel, and 128-channel data, respectively], we employed t tests. The SP-evoked power values are significantly lower after the final preprocessing stage for 6-channel (t(13) = 3.012; p = 0.005; d = 0.805), 64-channel (t(15) = 10.319; p < 0.001; d = 2.580), and 128-channel setups (t(62) = 2.971; p = 0.002; d = 0.374).
Visually, the application of ASR on top of PREP exhibits the most drastic changes in artifact removal (compare second and third rows in Fig. 3D). Note, however, that right-hemispheric artifacts in high-density passive setups appear more resistant to cleaning (Fig. 3D, rightmost column). This pattern is also reflected in the ∼40 Hz band topographies in Figure 3E (bottom row): Increased cleaning along the preprocessing pipeline is shown with the addition of ASR visibly decreasing the most artifacts. Again, some right-hemispheric artifactual components are retained even after SSP (Fig. 3E, right side). This might be due to different datasets having systematically different placement of HEOG electrodes in a way that rightward saccades could not be detected as easily by our algorithm and SSP. GFPs (Fig. 3E, top row, gray areas) again confirm that the combination of PREP and ASR have the biggest effect on SP-related activity reduction. The same is shown for Hilbert envelopes across EEG channels (Fig. 3E, top row, colored lines). Note that in 64-channel active setup data, fluctuations visibly oscillate more often (indicative of previous and upcoming saccades), more slowly and with higher amplitude than in 128-channel passive setups. This might be because instructions for participants from different datasets were systematically different. For example, in 64-channel datasets, participants might have been instructed to do extensive LRLR signals, which take longer to perform, and to perform these signals multiple times in a row. Crucially, note that most of these oscillations of saccadic origin are resolved after applying PREP and ASR, which highlights ASR in handling nonstationary artifacts. However, the lowest-density EEG dataset (six-channel) was excluded from the PREP pipeline, as its performance relies on adequate channel density for effective bad channel detection and interpolation. For this dataset, preprocessing began with ASR, followed by SSP.
PSD analysis results
With the validity of our preprocessing demonstrated, we can now move on to the electrophysiological correlates of LD. In all analyses, we investigated the following within-subject and within-session conditions: waking, LD, earlier nonlucid REM sleep (E. REM), and later nonlucid REM sleep (L. REM, i.e., similar times as LD periods; Fig. 2A; see Materials and Methods for further details). Results of global, sensor-level power analyses are shown in Figure 4A–C. As expected, the waking condition showed a pronounced power increase in the alpha band (8–12 Hz) compared with nonwaking stages. A robust increase in power can also be seen from 20 Hz upward (Fig. 4A) during waking.
Sensor-level PSD, complexity, entropy, and topographical power analysis results. A, Statistical comparison of PSDs across conditions for each frequency band. The colored shapes represent smoothed density estimates. Overlaid box plots display the median and interquartile range, with whiskers indicating the data spread. The magenta line connects the mean values across conditions. B1, Grand-average PSDs by across six common channels with shaded areas showing 95% confidence intervals. Error bars show standard error of the mean. B2, Pairwise mixed-frequency power comparisons between LD and E. REM, L. REM, and wake. The top row displays absolute power differences in dB, while the bottom row illustrates relative power changes (%). The shaded regions indicate 95% confidence intervals. A significant power drop (blue contour) is observed in the lucid versus wake comparison, spanning from 8.78 Hz across the entire frequency range. No significant differences were detected between lucid and E. REM or L. REM across any frequency range. C, Statistical comparisons of LZc, entropy measures, and HF across conditions. Significant differences are indicated by black lines for p < 0.01 and a red line for p < 0.05. Density estimates are shown in color to represent the data distribution, complemented by box plots that indicate the median and interquartile range. D, Sensor-level topography maps contrasting tonic power in LD, REM, and waking conditions within each frequency band. Colors show T values representing relative changes in power. White dots point to electrode locations that belong to significant channel clusters (p < 0.02). PSD, power spectral density; LZc, Lempel–Ziv complexity; HF, Higuchi fractals; LD, lucid dreaming; REM, rapid eye movement sleep.
Frequency-band–specific LMM analyses revealed significant differences between conditions across the delta (p < 0.001), theta (p = 0.023), alpha (p < 0.001), beta (p < 0.001), gamma1 (p < 0.001), and gamma2 (p < 0.001) frequency bands. However, no significant post hoc comparisons were found in the theta band. Variance explained by fixed effects (marginal R2) was as follows: 0.466 for delta, 0.424 for alpha, 0.342 for beta, 0.458 for low gamma, and 0.562 for high gamma. The total variance explained by both fixed and random effects (conditional R2) was 0.641 for delta, 0.542 for alpha, 0.495 for beta, 0.604 for low gamma, and 0.690 for high gamma. Post hoc analyses indicated that the waking condition drives the observed effects, as waking power significantly differed from all other conditions (Fig. 4C). In the delta band, waking power (−10.506 ± 0.655) was lower than E. REM (−8.272 ± 0.744; p < 0.001) and L. REM (−8.551 ± 0.707). In the alpha band, waking power (−9.452 ± 0.627) was higher than E. REM (−13.435 ± 0.671; p < 0.001), L. REM (−12.952 ± 0.749; p < 0.001), and LD (−12.712 ± 0.781; p < 0.001). For the beta band, waking power (−17.810 ± 0.577) was higher than E. REM (−20.227 ± 0.681; p < 0.001), L. REM (−20.424 ± 0.669; p < 0.001), and LD (−20.415 ± 0.637; p < 0.001). In the low gamma band, waking power (−21.609 ± 0.640) was higher than E. REM (−25.481 ± 0.853) (p < 0.001), L. REM (−25.325 ± 0.884; p < 0.001), and LD (−25.061 ± 0.816; p < 0.001). Lastly, in the high gamma band, waking power (−24.860 ± 0.649) was higher than E. REM (−29.588 ± 0.949; p < 0.001), L. REM (−29.679 ± 1.021; p < 0.001), and LD (−29.586 ± 0.964; p < 0.001). Notably, no significant differences were observed within REM variations (E. REM and L. REM) or between the REM variations and LD in any frequency band.
In the mixed-frequency analysis, we performed pairwise permutation cluster tests to compare spectral power between LD and each of the three conditions (E. REM, L. REM, and wake) individually. Using an alpha threshold of p < 0.01 to identify significant frequency regions, we found no significant differences in power between LD and either E. REM or L. REM. However, a substantial significant cluster emerged in the comparison between LD and wake, spanning the frequency range from 8.78 to 45 Hz, indicating a pronounced power reduction in LD relative to wake within this range. To further illustrate these findings, we visualized both absolute power differences and relative power changes (%) between LD and the other conditions (Fig. 4B1, B2). In particular, relative power comparisons (Fig. 4B2) highlight that power in LD closely resembles that of both E. REM and L. REM, particularly from 8 Hz onward, whereas a clear divergence is observed in the LD versus wake contrast.
Entropy, LZc and fractal analyzes results
Markers of LZc (Aboy et al., 2006), SE (Delgado-Bonal and Marshak, 2019), PE (Bandt and Pompe, 2002), and HFD (Harne, 2014) have been shown to covary with conscious awareness (Zhang et al., 2001; Seth et al., 2011; Sarasso et al., 2021), which makes them potential candidates for metrically capturing lucidity (Baird et al., 2022). The differences between waking, LD, and nonlucid REM sleep with regard to entropy and complexity markers are shown in Figure 4C. Two samples with extreme values were identified across five different distributions in PE and SE. These outliers were flagged through visual inspection of smoothed density estimates in raincloud plots, indicating significant deviation from the typical distribution. Specifically, extreme values were noted in PE for E. REM and L. REM, as well as in SE; outlier conditions were also identified for wake in both PE and SE. A z-score–based tail probability test confirmed the statistical significance of these outliers (p < 0.001). To maintain consistency across analyses, these samples were removed from all markers, resulting in a final dataset of 41 samples per condition. LMM results reveal that all measures differ significantly between conditions (LZc p < 0.001; PE p = 0.002; SE p < 0.001; HF p < 0.001). Variance explained by the fixed effects (marginal R2) was 0.396 for LZc, 0.396 for PE, 0.464 for SE, and 0.505 for HF. The total variance explained by both fixed and random effects (conditional R2) was 0.466 for LZc, 0.556 for PE, 0.586 for SE, and 0.570 for HF. Post hoc analyses revealed that the waking condition drives these effects. In terms of LZC, waking values (0.127 ± 0.002) were significantly higher than those of E. REM (0.110 ± 0.002; p < 0.001), L. REM (0.112 ± 0.002; p < 0.001), and LD (0.116 ± 0.002; p < 0.001). Additionally, LD was found to be significantly higher than E. REM in LZC (p = 0.003). In SE, waking values (0.866 ± 0.013) were significantly higher than E. REM (0.782 ± 0.013; p < 0.001), L. REM (0.786 ± 0.013; p < 0.001), and LD (0.810 ± 0.012; p < 0.001). For PE, waking levels (0.829 ± 0.003) were significantly higher than E. REM (0.820 ± 0.003; p = 0.003) and L. REM (0.819 ± 0.002; p < 0.001), while no significant difference was observed between waking and LD (0.825 ± 0.003; p = 0.709). Moreover, LD was observed to be significantly higher than L. REM (p = 0.048). For HF, waking values (1.955 ± 0.003) were significantly higher than E. REM (1.918 ± 0.004; p < 0.001), L. REM (1.922 ± 0.004; p < 0.001), and LD (1.929 ± 0.004; p < 0.001).
Topographical power analysis results
A sufficient sample size of high-density EEG recordings allowed us to explore channel clusters showing significant differences between conditions across frequency bands of interest (Fig. 4D). Firstly, we observed no clusters that could significantly distinguish between L. REM and E. REM in any of the frequency bands. However, when contrasting LD with waking, we note significant topographical differences in all frequency bands but theta. Specifically, in the delta band, we observed higher power in the frontal and central regions during LD; in the alpha band, a significant decrease in the left-occipital region; in the beta band, a near-global decrease except the central region; in the gamma1 band, significant decreases in the frontal-left and occipital regions; and in the gamma2 band a significant power drop particularly in the frontal-left and extending to the frontal and left-parietal and occipital regions. Sensor-level topography of theta band power differs significantly between LD and nonlucid REM (i.e., both E. REM and L. REM) with an observed theta decrease in LD compared with L. REM over central and extending to parietal and occipital channels and an observed theta decrease in LD compared with E. REM in the central-left region. We also found significant differences in LD delta-band power compared with E. REM delta-band power, with stronger power differences over the central-left channels, but no differences in LD delta compared with L. REM delta.
Surface-based source reconstruction results
We applied source reconstruction to uncover neural activation patterns that set LD apart from both wakefulness and non-LD. Results are shown in Figure 5A for dSPM and eLORETA, respectively. Comparing the two methodologies, we note substantial overlap in the findings of the two methods as well as with sensor-level topographical results (Fig. 4D) in frequency bands from 8 to 45 Hz, underscoring the robustness of our findings. Specifically, within the alpha, beta, and gamma frequency ranges, we find power reductions in LD relative to waking conditions. Both methodologies identified alpha power reductions primarily in the occipital region, whereas beta power showed a more extensive decrease spanning the bilateral temporal, frontal, and occipital lobes. In the gamma frequency bands, eLORETA indicated a widespread cortical power decline with notable decreases in the left and particularly the right frontolateral regions, while dSPM suggested a more localized reduction, especially in the right frontal region for gamma2. Furthermore, the findings revealed localized significant variations between LD and L. REM sleep for both dSPM and eLORETA. dSPM pointed to a theta decrease in the right occipital region with extension from posterior aspects toward the parietal lobe encompassing precuneus in LD, a pattern similar to the beta reduction observed mainly in the precuneus with traces in the occipital area. eLORETA identified beta power reductions as well but showing a more distributed profile. Additionally, an increase in gamma1 power during LD in the left frontolateral region as compared with L.REM was observed via dSPM only. No significant differences in cortical surface power were observed between LD and E. REM or between L. REM and E. REM, across all frequency bands.
A, Thresholded surface-based source–level power for each frequency band derived from dSPM and eLORETA and plotted over cortex images. Each surface significance is shown with five different views (coronal, dorsal, left sagittal, right sagittal, and ventral). Power values of statistically significant clusters contrasting LD, REM, and waking for both dSPM and eLORETA methods. All shown clusters were significant with p < 0.025, except clusters marked with *, which were significant with p < 0.05. The label “n.s.” denotes regions where the results were not statistically significant. B, Grand-average contrasted connectivities between L. REM and E.REM, LD and E. REM, LD and L. REM, and LD and wake across frequency ranges, analyzed using dSPM and eLORETA. Increases in PLI connectivity patterns, significant through cluster-permutation tests, are highlighted in black color (dSPM, LD–E. REM alpha and gamma1; eLORETA, LD–E. REM and LD–L. REM alpha), with no significant reductions detected. All statistical tests were conducted with and without additional bilateral significance threshold adjustments, with the symbols “*” and “**” indicating the p threshold (0.025 and 0.05, respectively) for each visualized significance test. Significant connections employing wPLI-debiased measures strongly overlap the PLI findings. LD, lucid dreaming; REM, rapid eye movement sleep; PLI, phase lag index.
Functional connectivity results
Cortical measures of functional connectivity allowed us to explore correlates of LD even deeper. Grand-average phase lag indices (PLIs; Fig. 5B) reveal robust increases in alpha connectivity during LD as compared with E. REM across analysis methods (i.e., across dSPM, eLORETA with PLI/wPLI-debiased approaches). Inspecting shared connections across all methods, robust hub areas become evident, most prominently the left superior temporal gyrus, but also right superior parietal gyrus, right superior frontal gyrus, and left lingual gyrus (Fig. 5B). Significantly increased alpha connectivity during LD could be shown against L. REM as well, however only in PLI measures and when using eLORETA (Fig. 5B). Frontal regions were not identified in this connectivity pattern, which overlaps with the auditory resting state network and partly the visual and posterior parts of the default mode network as well. Apart from alpha, we noted increased gamma1 band connectivity during LD as compared with E. REM when employing PLI measures from dSPM, almost exclusively in frontotemporal regions. No other significant connectivity effects could be shown. Specifically, we observed no differences between LD and waking or between temporally distant nonlucid REM intervals. See Extended Data Figure S5 for grand-average PLI and wPLI-debiased results, source-localized using dSPM and eLORETA, as well as significant alpha-band connectivity increases during LD compared with E. REM and L. REM.
Temporal dynamics of lucidity activation results
In a last analysis, we tested the temporal dynamics of LD around the initial LRLR eye signaling (within a symmetrical search window of 30 s) via spatiotemporal clustering of GFPs. The first 5 s of the search window served as the baseline (for further details, refer to Materials and Methods). All activation patterns are shown in Figure 6. According to the results of the spatiotemporal cluster tests, we found significant activation in the alpha band, which exceeded the cluster threshold in between 1.75 s before and 3.25 s after LRLR onset. However, no significant differences were identified at the source level when contrasted with the baseline in this band. We also detected significant activation patterns at both the sensor and source levels in the gamma1 and gamma2 bands. Specifically, in the gamma1 band, spatiotemporal clusters can be observed between 2 s before and 7 s after initial LRLR, leading to sensor-level activation primarily in the occipital-left region, as determined by the topographic layout of GFP activations. In the gamma1 frequency range, both MNE-dSPM and eLORETA methods identified marked increases in cortical activity. MNE-dSPM reveals bilateral frontal lobe engagement, with notable extensions into the central and parietal regions, alongside a significant activation in the left temporal lobe. eLORETA echoes these findings but with more pronounced and extensive activations across the frontal and parietal lobes, notably highlighting the right precuneus and enveloping the temporal lobes bilaterally. Additionally, in the gamma2 frequency band, significant sensor-level activation was detected that exceeded cluster thresholds between 1 s before and 5 s after initial LRLR, particularly over occipital-left channels. When examining the same time interval at the source level, robust activations were discerned at the sensor level that surpassed cluster significance thresholds within the period extending from 1 s prior to 5 s following the onset of LRLR eye signals, predominantly over the occipital-left channels. This sensor-level activity is mirrored at the source level with predominant activations in the left frontolateral and left temporal regions as delineated by both dSPM and eLORETA analyses.
A, ROI and frequency-band–specific GFP fluctuations of EEG data during LD around the time point of LRLR onset. Error shadings represent 95% confidence intervals. GFPs are normalized on the baseline interval which is shown in light red shading 15–10 s before the LRLR onset. Time intervals of significant (p < 0.05) spatiotemporal activation of EEG data are marked in gray shading and with a dash-dot line. White dots on topographies indicate significant (p < 0.05) cluster channels. Inlays show clusters exceeding the cluster threshold at the sensor- and source-level of this time area. B, Source-level activation maps showing significant differences in gamma1 and gamma2 bands, identified using dSPM and eLORETA, relative to the baseline. The label “n.s.” denotes regions where the results were not statistically significant. C, Source-level functional connectivity changes during initial eye signaling compared with the baseline, estimated using dPLI. Connectivity patterns are visualized for both dSPM (left) and eLORETA (right) in gamma1 and gamma2 bands. Circular plots with black lines illustrate significant (p < 0.01) connectivity changes across cortical regions, while heatmaps display T distributions of significant pairwise connections between major brain regions. ROI, region of interest; GFP, global field power; LRLR, left→right→left→right eye signal; dPLI, directed phase lag index.
During the initial eye signaling period of LD, compared with the presignaling baseline, source-level functional connectivity analysis using dPLI revealed significant connectivity increases across the cortex. Both dSPM and eLORETA source reconstructions showed a more extensive connectivity enhancement in the gamma1 band (30–36 Hz), with a notable predominance of frontal centrality while also involving widespread connectivity across all major cortical regions. In contrast, gamma2 (36–45 Hz) exhibited a less pronounced overall connectivity increase, with a more localized dominance in occipital regions and relatively weaker frontal involvement compared with gamma1. Additional analyses and supporting materials are provided in Extended Data.
Discussion
LD is an intriguing state of consciousness that remained poorly understood at the electrophysiological level, largely due to low statistical power, inadequate saccadic artifact correction, and limited sensitivity to source-level activities. Here we employed SP cleaning and cortical source reconstruction in the highest sample size with most densely sampled topographical EEG resolution in the LD field to date. In comparison to nonlucid REM sleep, we observed a reduction in posterior theta power and a beta power decrease in the right temporoparietal junction (TPJ). Besides, we found gamma activity involving the precuneus, left prefrontal cortex, and frontopolar areas around the time of the first LD eye signaling. Finally, we identified enhanced connectivity in the alpha band with a prominent hub in the left superior temporal gyrus, as well as increased gamma connectivity within bilateral frontotemporal regions, as discussed below.
With an inclusive preprocessing pipeline accommodating low-density datasets, we revisited earlier sensor-level spectral findings (Baird et al., 2019) with a well-powered sample. LD significantly differed from the waking state across most frequency bands (Fig. 4D), showing higher power in low frequencies (2–4 Hz) and lower power in higher frequencies (8–45 Hz), consistent with general REM sleep patterns (Zhang et al., 2001; Baird et al., 2022). This finding supports that subjects were not awake, addressing the main critiques in early LD research. No theta band (4–8 Hz) differences were found, but full-spectrum analysis (Fig. 4B1) suggests averaging effects, with notable differences between 3 and 5.5 Hz, where LD spectral power lies between waking and nonlucid REM. While this highlights potential limitations of traditional frequency bands and supports recent calls for inductive spectral analyses (Donoghue et al., 2022), continuous frequency analysis showed no robust differences at the sensor level between LD and nonlucid REM epochs. The pronounced power drop during LD compared with wakefulness, spanning from 8.78 Hz—marking the lower boundary of the alpha band—through the entire analyzed spectrum, suggests a broad spectral divergence between these states (Fig. 4B2). Additionally, the maximum sampling rate across our recordings is 100 Hz, limiting analysis up to the Nyquist frequency of 50 Hz and restricting the ability to explore higher gamma ranges (Uhlhaas et al., 2011; Kucewicz et al., 2017). Future studies employing high-sampling-rate EEG recordings of LD could utilize harmonic enhancement methods to preserve content-based oscillatory information while attenuating high-frequency percussive components (Demirel et al., 2024), enabling a deeper exploration of very high gamma activity. Moreover, EEG complexity and entropy measures distinguished LD from nonlucid REM (Baird et al., 2022), with larger sample data here showing significant differences in LZc between LD and E. REM and in PE between LD and L. REM—albeit less pronounced than differences with waking (Fig. 4C). The similarity of E. REM and L. REM across markers reinforces the utility of information-theoretical metrics in distinguishing states of consciousness.
High-density topographies revealed frontocentral and parietal clusters with reduced theta power during LD compared with REM epochs (Fig. 4D), potentially reflecting attentional processes, as neurofeedback protocols reducing theta power enhance attentional control (Heinrich et al., 2014; Holtmann et al., 2014). Lower theta power might aid lucidity in REM sleep, supported by dSPM results localizing effects to right occipital areas (Fig. 5A). While the previous study (Baird et al., 2022) reported widespread delta reductions in LD compared with nonlucid REM, our study found this only when comparing LD to early REM, suggesting an interaction between sleep duration and lucidity. Cluster tests did not show this effect for later REM, possibly due to temporal proximity (Figs. 2A, 4D). Notably, source-level analyses could not confirm the delta power increase in LD versus waking seen at the sensor level, suggesting delta spatial distributions during LD resemble waking states. A significant gamma1 (30–36 Hz) power increase in LD versus later REM (Fig. 4D) emerged via dSPM but not eLORETA or sensor-level analyses (Fig. 5A), likely due to dSPM's noise normalization enhancing SNRs (Dale et al., 2000). Gamma1 increases, localized to left temporal areas, may reflect verbal insight (Jung-Beeman et al., 2004) or predictive coding processes resolving low-level sensory prediction errors within the middle temporal gyrus (Simor et al., 2022).
Both cortical source localization techniques (MNE-dSPM and MNE-eLORETA) consistently revealed reduced beta band activity in LD compared with temporally close nonlucid REM sleep (Fig. 5A), undetectable at the sensor level, likely due to the enhanced spatial resolution of MNE. Beta activity, associated with maintaining cognitive status quo (Engel and Fries, 2010), may diminish during LD to facilitate reassessment of perceived reality, though further research is needed. dSPM's noise normalization approach reveals more focal source clusters, identifying the right TPJ as a likely origin of beta activity. According to a theoretical framework (Simor et al., 2022), LD assigns higher precision weighting to lower sensory predictions, with increased prediction error resolution leading to heightened TPJ activity. The TPJ integrates visual, auditory, tactile, proprioceptive, and vestibular information, contributing to self-consciousness and body imagery (Blanke and Mohr, 2005). Disrupting TPJ activity during wakefulness via magnetic (Blanke et al., 2005) or electric (De Ridder et al., 2007) transcranial stimulation can induce out-of-body experiences, characterized as sensations of being “outside the own body” (Blackmore, 1982). Interestingly, LD and out-of-body experiences have also been linked (Irwin, 1988; Levitan et al., 1999; Campillo-Ferrer et al., 2024). Around initial lucidity eye signals, we observed no beta effects (Fig. 6), likely due to the brief baseline not capturing the sustained nature of beta-mediated metacognition (Engel and Fries, 2010).
Preprocessing data with our pipeline supports recent findings (Baird et al., 2022) that LD is not characterized by EEG frontal activity in frequencies ∼40 Hz when SP artifacts are sensibly attenuated (Figs. 3E, 4D). However, theoretically, this does not exclude the possibility that frontolateral gamma plays a role in LD (Voss et al., 2009, 2014) but is overshadowed by SPs. Gamma oscillations are associated with attention, sensory integration, and consciousness (Cavinato et al., 2015; Lu et al., 2021; Lin et al., 2023)—processes that are engaged during LD tasks such as flying and visual exploration. We observed gamma1 (30–36 Hz) activations around the initial eye signals marking lucidity's onset, suggesting they might precede epochs marked as lucid via signaling (Fig. 6). Source-level analysis revealed an increase in gamma1 activity around these signals, likely reflecting metacognitive processing linked to the shift from non-LD to conscious awareness. Importantly, source localization identified this gamma1 activity partly in the right-hemisphere precuneus, a region that facilitates complex cognitive functions through gamma oscillations, particularly during heightened awareness (Dresler et al., 2012; Hebscher et al., 2019). This may indicate memory retrieval and motor intention, as subjects recall the instructed task of signaling upon realizing lucidity, highlighting task-related memory during this transitional moment. Precuneus activity is also modulated during self-related cognition (Cavanna and Trimble, 2006) and closed-eye imagery induced by psychedelics like ayahuasca and LSD (De Araujo et al., 2012; Carhart-Harris et al., 2016), which share similarities with dreaming (Preller and Vollenweider, 2016; Sanz et al., 2018; Zamberlan et al., 2018), and specifically with LD (Kraehenmann, 2017). This voluntary behavior is not merely mechanical but represents a goal-directed action with a strong self-referential component. The dreamer's recognition of their mental state and deliberate intention to signal externally reflect key metacognitive processes. This supports the idea that the gamma1 activation seen during LD onset corresponds to the brain's engagement in mental state recognition and the initiation of intentional behavior aimed at communicating this awareness, consistent with findings that gamma-band synchronization is associated with conscious access and the engagement of large-scale brain networks (Dehaene and Changeux, 2011). Increased functional connectivity was observed across the cortex during initial LD eye signaling compared with the baseline in both gamma1 and gamma2 bands. However, gamma1 exhibited a more pronounced increase, particularly with dSPM, showing extensive frontal connections to temporal, parietal, and occipital areas. This frontal-to-global connectivity (Fig. 6C) suggests heightened frontal engagement in coordinating distributed neural activity, potentially supporting cognitive control mechanisms during LD (Cole et al., 2012). In contrast, gamma2 connectivity was less intense and exhibited a more localized pattern, with relatively stronger connectivity in occipital and temporal regions. This reduced frontal synchronization in gamma2 activity may be partially attributed to saccadic artifacts, which are known to be more prominent in higher gamma frequencies and could disrupt phase synchronicity. Notably, both dSPM and eLORETA yielded highly consistent patterns, with dSPM showing slightly more pronounced frontal connectivity but fewer significant connections overall, reinforcing the reliability of the observed connectivity dynamics.
We further observed increases in gamma2 activity related to initial eye signaling, most prominently in both left prefrontal and frontopolar cortices. This localization, combined with the fact that gamma2 includes 40 Hz—an activity notoriously difficult to clean from eye-movement–related artifacts (Yuval-Greenberg et al., 2008; Keren et al., 2010; Baird et al., 2022)—raises the possibility of an artifactual origin. Disentangling the exact contributions of non-neural components from true neural sources requires imaging methods not sensitive to electric currents, such as fMRI. Previous MRI research has linked frontopolar cortices to LD (Dresler et al., 2012; Filevich et al., 2015; Baird et al., 2018). While the preprocessing pipeline tested in this study satisfactorily removed saccadic traces in EEG (Figs. 3D,E), the removed saccades were presented in short-time windows (Fig. 3D), corresponding to miniature saccades (Yuval-Greenberg et al., 2008) around voluntary eye signalings (Figs. 3A,B). These miniature saccades, which typically occur just before major amplitude shifts (during steep transitions approaching a valley or peak), may explain residual eye signaling-related gamma2 activity. Given the high-amplitude EOG from conscious eye signaling necessarily included in this analysis interval, our findings likely support recent work (Baird et al., 2022) showing that saccadic activity explains frontal 40 Hz signals associated with LD (Voss et al., 2009) via source reconstruction methods.
The spectral connectivity results reveal increased long-range alpha communication during LD compared with nonlucid REM sleep, particularly within the left-hemispheric auditory resting state network (Fig. 5B). Statistically, this effect is most pronounced in earlier nonlucid REM episodes, indicating a potential confound of sleep duration. Across source localization and connectivity methods, we noted an alpha connectivity between the right superior frontal gyrus and left lingual gyrus. The involvement of anterior frontal cortices in LD is significant, as during nonlucid REM sleep, the lateral prefrontal and frontopolar cortices are downregulated (Maquet et al., 1996) and uncoupled from posterior regions, effectively separating executive functions from sensory perception (Pérez-Garci et al., 2001; Cantero et al., 2004). This uncoupling may underlie the lack of judgment and passive acceptance of bizarreness during ordinary non-LD (Hobson et al., 2000). Alpha connectivity increases contrast with the decreases seen in psychedelics (Pallavicini et al., 2019), despite shared subjective features such as vivid imagery, emotional intensity, and increased metacognition (Kraehenmann, 2017). While psychedelics often lead to ego dissolution and reduced self-referential processing within the DMN, LD appears to harness self-awareness and control, potentially reflected in enhanced alpha connectivity. Similarly, alpha reduction during shamanic trance states (Huels et al., 2021) suggests an altered state of consciousness distinct from LD's self-directed focus.
Regarding other limitations of our study, LD cannot yet be fully experimentally induced at will, meaning that some uncontrolled factors may contribute to the observed differences between lucid and nonlucid REM sleep. Relatedly, it should be noted that subjective aspects beyond dream awareness could also influence the differences observed in LD, such as the often-reported increase in perceptual vividness, especially visual (Mota-Rolim et al., 2010). Unfortunately, the lack of nonlucid dream reports precludes a statistical control for sleep mentation, and future studies should include such reports. Additionally, while heterogeneity in dream content remains an inherent challenge in LD research, it is important to emphasize that lucidity itself represents a distinct cognitive state that arises independently of specific dream narratives. The moment of lucid insight—when metacognitive awareness emerges within the dream—serves as a unifying event that is highly similar across different LD episodes and individuals, potentially offering meaningful insights into the mechanisms of LD. Nevertheless, the heterogeneity of LD experiences remains an unresolved challenge, as variations in task conditions, dream environment, and content inherently limit cognitive homogeneity across episodes. Such heterogeneity is also inherently present in non-LD, of course, but remains as a limitation, and options for more homogeneous conditions and standardized protocols should be explored in future studies. To this end, the tonic brain activity identified in our study might serve as a guide to better define the temporal extent of individual LD. Furthermore, despite the advancements our data cleaning pipeline allows, the potential for overcleaning and inadvertently diminishing signals of interest is a risk that can never be entirely eliminated. Still, while artifacts were cleaned by our preprocessing protocol (Fig. 3), signals of interest are also retained as reflected in the spectral characteristics (Fig. 4A,B), which highlight the pronounced differences between the waking state and REM sleep (including LD). These differences comprise a wake alpha peak near-absent in sleep (Cantero et al., 2002) and relatively increased higher power in higher-frequency bands during the waking state (Kozhemiako et al., 2022). Mirroring the spectral results, waking epochs also differentiated from all REM sleep epochs with regard to entropy, complexity, and fractal dimension measures, as expected (Srinivasan et al., 2007; Tononi, 2008; Cavinato et al., 2015; Sarasso et al., 2021; Fig. 4C).
In summary, electrophysiological source localization measured from lucid REM sleep offers important insights into the nature of the brain's ability to determine whether currently active world models are aligned with one's environment (as in functional waking perception or LD) or not (as in nonlucid dreams and hallucinations). LD thus offers a unique avenue for voluntarily interacting with and immersively adjusting internal world models whose dysfunction is often at the root of various mental disorders (Filevich et al., 2015; Baird et al., 2018). The ability to adjust these dysfunctional models during dreaming presents a potentially useful tool for therapy (de Macêdo et al., 2019; Foffani, 2023; Ouchene et al., 2023; Yount et al., 2023). The electrophysiological characterization of LD across a representative dataset in this study enhances our understanding of the cortical activities associated with metacognitive insight into an ongoing state of nonveridical world representation. It may also facilitate the development of neurofeedback and brain–computer interface technologies aimed at inducing LD toward unlocking its full clinical potential.
Data Availability
Dataset from Donders Center for Cognitive Neuroimaging at RU, UO, and Brain Institute at UFRN will be made available upon request. For datasets from MPI/Munich and Department of Psychology/SU, participant agreement for data sharing was not obtained. The source code is available at https://github.com/caghangir-1/Electrophysiological-Correlates-of-Lucid-Dreaming.
Footnotes
This work was supported by a Vici grant from the Dutch Research Council (NWO). We are grateful for all students and assistants involved in spending overnights and earliest mornings to help provide the database for this study and make this project a reality. We also thank Olivia Brunette, Mahdad Jafarzadeh Esfahani, Pedro Reis Oliveira, and Teresa Campillo-Ferrer for their support in data collection and sleep scoring.
↵*N.A. and M.D. shared last authorship.
The authors declare no competing financial interests.
- Correspondence should be addressed to Çağatay Demirel at cagatay.demirel.sci{at}gmail.com.
