KurisutinaSiclari, F., Baird, B., Perogamvros, L., Bernardi, G., LaRocque, J. J., Riedner, B., Boly, M., Postle, B. R. and Tononi, G. Nature Neuroscience 20, 872–878. DOI and publisher record; complete accepted manuscript in PMC.
Read on 29 September 2026. Full reading complete for the recorded version/package: the complete PMC author manuscript, including Methods, captions, acknowledgments, disclosures and 72 references; all five main figures visually; the publisher's 13-page supplement, including all eight figures and five tables visually; and the entire 11-page reporting checklist, with its statistical tables also inspected visually. Main text was read as archived HTML, not the final typeset PDF. The two publisher-listed supplement files are the authority for supplementary coverage. PMC also lists three generically named manuscript attachments; their binary identities were not independently reconciled with the published package. Data and code were not requested or reanalyzed, and no experiment was run. See provenance.
Question and approach
Can EEG distinguish reported experience from reported absence of experience within sleep stages, and does its location relate to subsequently reported dream contents? This avoids equating REM with dreaming or unresponsiveness with absence of consciousness. It does not remove reliance on reports: labels come from structured interviews after awakening.
The interview distinguished experience with recalled content (DE), experience whose content could no longer be recalled (DEWR), and no experience (NE). Participants described their most recent experience and, where applicable, rated thought versus perception and the presence of faces, spatial setting, movement and speech. The separation of DEWR from NE is a useful test against the narrow explanation that all effects merely reflect successful recall of content; it cannot independently establish that every NE was truly devoid of experience.
Design and measurement
| Experiment | Sample and selection | Measurement and comparison |
|---|---|---|
| 1 | NREM analysis: 32 people selected from a larger pool of 69, requiring both DE and NE in the same night; 233 retained awakenings after seven exclusions. Twenty had all three report categories. REM analysis included six people from the larger pool with both DE and NE. | 256-channel EEG, usually the final 20 s before awakening; source power contrasts within participants. NREM primary contrast used N2. |
| 2 | Seven separate, trained participants, five to ten nights each; 836 awakenings, 21 excluded, 815 retained overall. Supplementary Table 2 tabulates 586 N2/N3 and 170 REM reports, not all 815 retained awakenings. | Replication of sleep-experience contrasts, with N2 and N3 included; detailed REM content comparisons. REM DE–NE analysis pooled eligible participants across experiments 1 and 2, total n=10. |
| 3 | Seven additional participants, one uninterrupted baseline night and two prediction nights; 84 triggered awakenings. | 64-channel EEG; individualized power thresholds over posterior electrodes, followed by awakening and report. The post-exclusion counts have reporting discrepancies described below. |
The first two experiments used artifact removal including ICA and an atlas-based four-shell forward model with 2,447 cortical dipoles on a 7-mm grid, sLORETA and regularization. Thus the localization is an EEG inverse estimate, not a direct measurement of neurons, synapses or connectivity. Source smoothing, volume conduction and propagating slow waves limit anatomical precision and interpretation of absolute power.
Main low-frequency contrasts used 1–4 Hz; high-frequency comparisons used 20–50 Hz in NREM and 25–50 Hz in REM. There are small band-label inconsistencies in the supplement. Whole-cortex experience contrasts used paired tests and cluster-based SnPM correction. Dream-content maps used more permissive uncorrected tests, followed by directional ROI tests. The ROIs came from prior literature, but the stated direction of the tests was informed by the preceding analysis. No sample-size calculation was performed. Collection and analysis were explicitly not blinded.
Findings, with figure coverage
Main Figures 1–2: DE relative to NE was associated with lower low-frequency power in posterior parieto-occipital cortex in NREM and REM. The DEWR–NE contrast also showed lower posterior low-frequency power; DE–DEWR did not show a significant low-frequency difference. Supplementary Figures 1–2 provide the second-sample result and overlap of report-category contrasts. These are group contrasts, not a universal absolute threshold: Supplementary Figure 5 shows substantially different absolute levels between NREM and REM.
Main Figure 3: higher-frequency power was greater for DE than NE, extending beyond posterior regions into frontal and temporal cortex. DE versus DEWR involved medial/lateral frontoparietal differences; DEWR versus NE had no corrected significant high-frequency effect (Supplementary Figure 4). The latter null does not establish equivalent physiology. Supplementary Figure 3 supplies a common 25–50-Hz NREM comparison. Supplementary Figure 8 displays unthresholded low-frequency effects beyond the most prominent posterior clusters. Supplementary Table 5 mostly shows no significant hemisphere-by-report interaction, with a marginal right-sided REM high-frequency interaction, p=.046.
Main Figure 4: in the seven trained participants, thought-like reports correlated with more anterior activity and perceptual reports with more posterior activity. REM content comparisons associated faces, setting, movement and speech with regions compatible with their waking functional associations. The directional ROI p values were .023, .023, .029 and .048, respectively; spatial setting used six participants, the others seven. These are small-sample, weakly corrected content associations, not blinded single-dream decoding. The image displays both p<.05 and p<.1 maps for some comparisons, although the caption describes a different secondary threshold. Effects vary over the final eight seconds and often peak nearest awakening.
Main Figure 5: posterior EEG drove prospective selection of awakening times. Baseline-night NREM epochs supplied each person's upper/lower 15% thresholds. A DE prediction required low-frequency power below its lower threshold and high-frequency power above its upper threshold, with the reverse conjunction for NE. At least 120 s of NREM was required before awakening. Online high-frequency power used 18–25 Hz to reduce artifact vulnerability; the figure and table label low frequencies 0.5–4.5 Hz, while the baseline-threshold Methods paragraph says 1–4 Hz.
The Results report 33 correct DE predictions out of 36 retained DE-prediction awakenings and 21 correct NE predictions out of 26 retained NE-prediction awakenings, giving 54/62 ≈87%. Those fractions condition on the algorithm's prediction: they are not separately the sensitivity and specificity of a classifier applied to all sleep epochs. The seven participants are the relevant independent units for subject-level tests. Supplementary Table 3 provides individual accuracies and thresholds. Supplementary Figures 6–7 and Table 4 show power/ratio differences in this selected sample; at the more permissive corrected threshold, the HF/LF ratio differed across all vertices, so a posterior maximum does not imply an exclusively posterior effect.
Reporting discrepancies that remain unresolved
The primary sources should not be silently forced into a single clean count or threshold:
- Results: 84 total awakenings, then 36+26=62 after exclusions. Methods: 20 exclusions, which would leave 64. Main Figure 5C and its caption instead label 55 DE and 27 NE, totaling 82. Table S3's awakening counts total 84. These are not interchangeable denominators. The exact trial flow and handling of every report class in experiment 3 require clarification or the original data.
- Figure 4C's image key and Methods show p<.05 and p<.1; its caption says p<.05 and p<.01. The more permissive displayed threshold must not be described as corrected significance.
- Supplementary Figure 1B's image labels 20–50 Hz while its caption says 25–50 Hz. The online low-frequency threshold bands also vary as noted above.
- The reporting checklist header says four supplementary tables, but the actual publisher supplement contains five; Table 5 is read and included. Its statistics table also appears to exchange Figure 5B/C descriptions and gives a different face-ROI t value from the main caption. Treat it as supplementary reporting, not an automatic resolution of main-text discrepancies.
These observations are a reading audit, not an allegation about the validity of the underlying data. They particularly limit exact reconstruction of the prediction analysis.
Interpretation and limits
The strongest contribution is the combination of within-stage comparisons, a distinction between content recall and a residual sense of having experienced something, and prospective EEG-guided awakenings in another group. It supports a relationship between posterior slow activity and subsequent experience reports that cannot be reduced simply to REM versus NREM or to ongoing button presses.
The paper's stronger language about a region being required for dreaming goes beyond this observational design. No selective disruption or rescue establishes necessity or sufficiency. Frontal effects are present in several analyses, and thresholded posterior significance cannot establish that other areas are irrelevant. Dream generation, memory encoding, attention, arousability and ability to describe experience can still contribute to the measured contrasts.
Prospective prediction is meaningful, but it applies to extreme EEG configurations selected by the algorithm, in seven new people with subject-specific baseline calibration. The study does not supply an all-night confusion matrix, coverage of intermediate configurations, an independent blinded replication, or performance in injured patients. Selecting extreme low- and high-frequency configurations also makes large subsequent power differences expected; those differences are not an independent validation of the biological mechanism.
Consequences for consciousness and connectomics research
Our inference: a useful model needs to distinguish state, local physiology, experience labels and the process generating a later report. Grossly similar anatomy is compatible with changing sleep physiology and reports within an individual, but this study neither measures the connectome nor estimates how much physiology anatomy could predict under specified inputs and initial conditions.
Compared with the completed Frässle reading, there is no ongoing perceptual decision or motor report during the EEG window, but report dependence is displaced into a later memory-based measurement. Compared with Cogitate, posterior involvement is another convergent observation under different conditions, not a direct replication of its theory-specific tests. The alternative evidence must remain visible: high-frequency frontal associations and wider unthresholded effects prevent a simple posterior-only conclusion.
Proposed next test, not performed: freeze preprocessing, frequency bands, thresholds and exclusions before blinded evaluation in new participants. Report performance and abstention/coverage over ordinary NREM epochs as well as the selected extremes, preserve DEWR as an explicit outcome, and assess sensitivity to sleep depth, time of night and arousal. Pair this reading with an independent blinded dream-prediction study before increasing confidence in generalization. The cited literature is a search queue, not additional independently read evidence.