Behavioral and brain responses to verbal stimuli reveal transient periods of cognitive integration of the external world during sleep

Abstract

Sleep has long been considered as a state of behavioral disconnection from the environment, without reactivity to external stimuli. Here we questioned this 'sleep disconnection' dogma by directly investigating behavioral responsiveness in 49 napping participants (27 with narcolepsy and 22 healthy volunteers) engaged in a lexical decision task. Participants were instructed to frown or smile depending on the stimulus type. We found accurate behavioral responses, visible via contractions of the corrugator or zygomatic muscles, in most sleep stages in both groups (except slow-wave sleep in healthy volunteers). Across sleep stages, responses occurred more frequently when stimuli were presented during high cognitive states than during low cognitive states, as indexed by prestimulus electroencephalography. Our findings suggest that transient windows of reactivity to external stimuli exist during bona fide sleep, even in healthy individuals. Such windows of reactivity could pave the way for real-time communication with sleepers to probe sleep-related mental and cognitive processes.

Fig. 1. Experimental design.

Participants with narcolepsy went through five 20-min naps during the same day. In each nap, periods with stimulation (ON) alternated, every minute, with periods when no stimulus was presented (OFF). During the ON periods, participants were presented with words and pseudo-words and asked to either frown (corrugator muscle contractions) or smile three times (zygomatic muscle contractions) in response to the stimuli. Stimuli were presented every 10 s (±1 s). Following each nap, participants were asked to report whether (1) they had any dream, (2) they were lucid and (3) they recalled any words presented during the nap. Immediately after this debriefing, participants performed a forced-choice ‘old/new’ recognition task. Healthy participants went through the exact same procedure except that they had a single 100-min nap.

Fig. 2. Examples of behavioral responses during N2 sleep in healthy participants (top) and during lucid REM sleep in participants with narcolepsy (bottom).

Wake periods corresponding to the same participants are shown on the left side of the figures as a comparison. The orange vertical line on the last channel indicates the stimulus onset. In these examples, we observed the typical markers of N2 sleep: spindles (EEG); and REM sleep: low chin tone (EMG), rapid eye movements (EOG) and θ rhythm (EEG).

Fig. 3. Accurate behavioral responses in both populations.

a, The overall response rate across different sleep stages during OFF (blue) and ON (green) stimulation periods in participants without (left) and with (right) narcolepsy. The total number of trials in a given condition is indicated on top of the bars. We used binomial generalized mixed-linear models with participants as a random factor for statistical analysis. All P values are corrected for multiple comparisons using the Benjamini–Hochberg procedure. Response rates were significantly larger in ON than in OFF periods (pairwise post hoc two-sided comparisons) in HP during wakefulness (P < 0.0001), N1 (P < 0.0001), N2 (P < 0.0001), REM (P = 0.0003) and in NP during wakefulness, N1, N2, REM and lucid REM sleep (all P < 0.0001). b, Accuracy was computed over responsive trials in the lexical decision task for participants without narcolepsy—HP (left) and with narcolepsy—NP (right). Only participants with at least three responses were included in this analysis (number of HP: wake = 21, N1 = 17, N2 = 10; number of NP: wake = 24, N1 = 25, N2 = 24, REM = 12, lucid REM = 15). Each dot represents a participant and dashed lines indicate the 50% chance level. The boundaries of the boxes represent the first and third quartiles (Q1 and Q3, respectively), the midline represents the median and the whiskers depict Q1 − 1.5× IQR and Q3 + 1.5× IQR. One-sided Wilcoxon signed-rank test revealed that both HP and NP were significantly more accurate than chance in all tested sleep stages (corresponding P values are indicated in the figure). IQR, interquartile range. ****, P < 0.0001; ***, P < 0.001. Source data

Fig. 4. Lucidity is associated with longer reaction times and increased responsiveness.

a, Distribution of reaction times from stimulus onset to response in correct trials (words and pseudowords) in NP across sleep stages. Dashed lines indicate medians. A mixed-linear model with participant as a random factor and pairwise post hoc two-sided comparisons revealed slower reaction times in lucid REM sleep compared to N1 (P < 0.0001), N2 (P = 0.0001) and REM sleep (P = 0.002). P values are corrected for multiple comparisons using the Benjamini–Hochberg procedure. b, Flowchart detailing the repartition of naps in participants with narcolepsy—the percentage of naps with at least one behavioral response is indicated and the responsive naps are further divided depending on whether participants reported a lucid dream upon awakening and whether they explicitly recalled responding during the nap. ****, P < 0.0001; ***, P < 0.001. Source data

Fig. 5. Participants exhibit sleep activity in responsive trials, with local brain activations associated with responsiveness.

a, Normalized PSD values in α (PSD |α|) and δ (PSD |δ|) frequencies in responsive trials across different sleep stages in prestimulus and poststimulus periods. Prestimulus marker values are computed over the 1 s-period before the stimulation, whereas poststimulus marker values are calculated in the 8s-period following the stimulation. Data are presented as mean values ± 95% confidence intervals. n depicts the number of datapoints included in the statistical analysis, taken from 25 NP (22 in wake, 24 in N1, 23 in N2 and 15 in REM) and 21 HP (21 in wake, 21 in N1 and 20 in N2). Please note that marker values in different sleep stages were never at wake level as revealed by a linear mixed model with participant ID as random effect (****P < 0.0001 in pairwise post hoc two-sided comparisons, adjusted for multiple comparisons), indicating that participants were indeed asleep while they were responding. b, Time–frequency analysis (TFA) performed on the Fp1 (top) and the O1 (bottom) electrodes in N2 (23 nonresponsive participants and 21 responsive participants) and REM sleep (15 nonresponsive participants and 14 responsive participants) of NP. Left and middle, stimulus-locked TFA in nonresponsive and responsive trials, respectively. Right, response-locked TFA. Transient and spatially localized increases in α and β frequencies were associated with behavioral responsiveness to the task. Source data

Fig. 6. EEG markers of high cognitive states computed before stimulus presentation vary with responsiveness to stimuli.

After the z score transform of marker values, we subtracted the marginal estimated mean of nonresponsive trials (NR) from responsive (R) trials for each marker and each stage (represented by bars). Statistical comparisons between the responsive and nonresponsive trials were made using linear mixed models with participant ID as random effect. Asterisks represent statistical significance in pairwise post hoc two-sided comparisons. ****, P < 0.0001; ***, P < 0.001; **, P < 0.01; P < 0.05; red stars indicate significance after FDR correction for 72 comparisons. All adjusted P values and averaged marker values can be found in Supplementary Tables 7 and 8. Almost all markers showed a variation in the direction corresponding to increased cognitive states when contrasting responsive trials to nonresponsive trials (for example, increased EEG complexity and decreased EEG δ power), both in participants with (left) and without narcolepsy (right). Note the similarity with Fig. 3 in ref. that contrasted conscious to nonconscious states in patients suffering from disorders of consciousness. R, responsive; NR, nonresponsive. Source data

Fig. 7. EEG markers of high cognitive states computed before stimulus presentation predict responsiveness to stimuli in each nonlucid sleep stage.

a,b, We fed a random forest classifier with our nine EEG markers and trained it to classify R trials versus NR ones using a tenfold cross-validation method. We conducted this analysis considering all responses (in blue), then separately for both correct (in green) and incorrect responses (in red). A confusion matrix for correct REM sleep trials in nonlucid naps of participants with narcolepsy is shown in a, with a description of the balanced accuracy measure that we computed to take unbalanced datasets into account. The confusion matrix for each stage and group can be found in Supplementary Table 9. Balanced accuracy scores are plotted in b for different sleep stages, in function of response accuracy, both for participants with narcolepsy (wake, N1, N2, REM sleep; left) and without narcolepsy (N2, right), with the corresponding statistical significance against chance level computed with a 500 permutations test (all trials: all NP P values = 0.002, HP P value = 0.006; correct trials: all P values = 0.002; incorrect trials: wake P = 0.002, N1 P = 0.04. Note that 0.002 is the smallest P value obtainable via 500 permutations). Data are presented as mean values ± 95% confidence intervals. n represents a number of datapoints in all (correct + incorrect) trials, taken from 22 NP in wake, from 24 NP in N1, from 23 NP in N2, from 15 NP in REM and from 20 HP. TP, responsive trials classified as responsive; TN, nonresponsive trials classified as nonresponsive; FP, false positives (NR trials classified as responsive). FN, false negatives (R trials classified as nonresponsive). Source data

Fig. 8. Effect of lucidity on EEG markers and response to stimuli in participants with narcolepsy.

a, Top, Kolmogorov complexity (left), normalized γ PSD (norm-γ; middle) and normalized δ PSD (norm-δ; right) before stimuli onset as a function of whether the stimulus will be followed by a behavioral response (in blue) or not (in orange), for lucid and nonlucid REM sleep in participants with narcolepsy. Data are presented as mean values ± 95% confidence intervals. Statistical differences computed via linear mixed models, adjusted for multiple comparisons are indicated (NS, nonsignificant). A number of datapoints in the model are 229 (from 13 participants) for responsive REM sleep, 358 (from 15 participants) for nonresponsive REM sleep, 353 (from 15 participants) for responsive lucid REM sleep and 333 (from 16 participants) for nonresponsive lucid REM sleep. Kolmogorov complexity and norm-γ were significantly higher for responsive trials compared to nonresponsive trials in nonlucid naps for all participants. Conversely, the norm-δ was significantly lower in responsive trials in nonlucid naps. No such differences were observed in lucid naps, suggesting a ceiling effect for markers of high cognitive states in lucid naps (see Supplementary Table 10 for statistical details). Overall, Kolmogorov complexity and norm-γ were higher, and norm-δ was lower in lucid naps compared to nonlucid naps irrespectively of the responsiveness. b, Time-generalization decoding of stimulus-related brain activity compared to baseline brain activity, in trials with (top) and without (bottom) response, in wake (left) and lucid REM sleep (right). The logistic regression classifier was trained on each time point and then tested on all the time points to obtain a generalization pattern. Each intersection point of a training time and a testing time shows the AUC of the classifier. Time points with an AUC > 0.5 and that are statistically significant are outlined in black (two-sided nonparametric sign test across participants with FDR correction for 41,616 comparisons, P < 0.05). Source data

Extended Data Fig. 1. Examples of response during N2 (upper panel) and during N3 sleep (lower panel) in participants with narcolepsy.

Wake periods corresponding to the same participants are shown on the left side of the figures. The orange vertical line on the last channel indicates the stimulus onset. Responses to stimuli corresponded to contractions of the zygomatic or corrugator muscles. All raw EEG and behavioral data are available on OSF (see Data Availability statement).

Extended Data Fig. 2. Mixed contractions to signal lucidity.

(a) Number of mixed contractions (objective lucidity code) in different sleep stages during naps that participants reported to be lucid (sky blue) or non-lucid (green). The relative number of lucidity codes can be found on top of the bars. (b) Number of lucidity codes exhibited by each participant in lucid REM sleep. We did not observe any mixed contractions in participants without narcolepsy (HP). On the other hand, we observed a total of 117 mixed contractions from 12 participants with narcolepsy (NP) in 19 different naps. Importantly, all 19 naps contained responses to the stimuli during N2 and/or REM sleep. Among the 117 mixed contractions, 93 were observed in REM sleep, 92 being in naps that were reported (upon awakening) to be lucid (Supplementary Fig. S7). Moreover, 18 contractions were observed in N2 sleep (12 being in lucid naps) and 6 contractions were observed in N1 sleep (5 being in lucid naps).

Extended Data Fig. 3. Automatic detection of the response contractions.

Upper panel: Examples of EMG traces showing corrugator (pink) and zygomatic (green) muscles contractions in Wake, N2 and REM sleep in healthy participants (HP). Lower panel: EMG variance modulations computed by the response detection algorithm in the corresponding trials. EMG variance drastically increases in the very time when contractions are visible on the EMG signal and only for the contracted muscle. Note that this method is robust to the slow drifts in the EMG signal (as shown in Wake, left panel) and only detects sudden modulations in the signal such as muscle contractions.

Extended Data Fig. 4. Automatic response detection algorithm detects significantly more contractions during ON stimulation periods compared to OFF stimulation periods in different sleep stages and mirrors the manual scoring.

(a) Statistical significance of differences in contraction rates between ON and OFF stimulation periods in healthy participants (left) and participants with narcolepsy (right) found by the algorithm using different parameter combinations [Threshold k: 5, 7, 10; Window size: 1 or 2 seconds]. Significant differences revealed by post-hoc two-sided pairwise comparisons following linear mixed models are indicated (****: p < 0.0001, ***: p < 0.001, **: p < 0.01, *: p < 0.05, dot: p < 0.1. (b) Response rates in ON and OFF stimulation periods found by the algorithm using the strictest parameter combination: threshold k = 10 and window size = 2 seconds. The algorithm labeled a trial as responsive if the variance of a 2 second window exceeded 10 times the baseline variance. Post hoc two-sided pairwise comparisons revealed significant differences in HP during wake (z = 33.82; p < 0.0001), N1 (z = 9.49; p < 0.0001), N2 (z = 4.45; p < 0.0001) and REM sleep (z = 1.98;p = 0.048); and in NP during wakefulness (z = 26.34; p < 0.0001), N1 (z = 15.73; p < 0.0001), N2 (z = 7.27; p < 0.0001), N3 (z = 2.11; p < 0.0001); REM (z = 10.16; p < 0.0001) and lucid REM sleep (z = 9.85; p < 0.0001) after correction for multiple comparisons using Benjamini-Hochberg procedure. Please note the similarity between this figure and Fig. 3a, indicating that response rates found by the algorithm followed the same trend as the manual scoring. Source data

Extended Data Fig. 5. Number of trials during ON and OFF stimulation periods in different sleep stages in participants with (left) and without (right) narcolepsy.

The partition of trials containing a response is filled with dark red color in both stimulation periods. Source data

Extended Data Fig. 6. Time-frequency analysis.

Time-frequency analysis (TFA) performed on the Fp1 (upper panel) and the O1 (bottom panel) electrodes in N1 and N2 sleep of HP. The Left and middle panels are stimulus-locked TFA in nonresponsive and responsive trials respectively. The right panels show response-locked TFA.

Extended Data Fig. 7. Timing of statistical differences in the time-frequency analysis.

Differences in power across time in delta, alpha, low beta/sigma and high beta bands, in responsive (R) and unresponsive (NR) trials during N2 and REM sleep in participants with narcolepsy (upper panel) and during N1 and N2 sleep in healthy participants (lower panel). Time-frequency analysis was performed over Fp1 and O1 electrodes. The error bands depict 95% confidence intervals. Significant differences are indicated by yellow shade (FDR corrected p-value < 0.05, mass-univariate analysis on time dimension using mixed-linear models with responsiveness as the explanatory factor and participant ID as a random effect). Source data

Extended Data Fig. 8. Response-locked ERP analysis in participants with narcolepsy (upper panel) and healthy participants (bottom panel).

Dashed vertical lines indicate response start. Please note the motor preparation potential over frontal electrodes in all tested sleep stages.

Extended Data Fig. 9. Evolution of electrophysiological markers across sleep stages in participants with narcolepsy (NP).

Three complexity measures (the Kolmogorov Complexity -KC, the Permutation Entropy -PE, and the Sample Entropy -SE), one connectivity measure (weighted symbolic mutual information (wSMI) in the theta band), and five spectral measures (normalized power spectral densities (PSD) of delta, theta, alpha, beta and gamma frequency bands) were computed separately for the wake (N = 961 from 22 NP), N1 (N = 505 from 24 NP), N2 (N = 1186 from 23 NP), N3 (N = 435 from 16 NP), and REM (N = 587 from 15 NP) sleep stages in participants with narcolepsy. The results in healthy participants can be found in Supplementary Fig. S3. Each dot indicates marginal means estimated by a mixed-linear model including sleep stage as an independent variable, EEG marker as the dependent variable, and participant ID as a random variable. Error bars depict 95% confidence intervals. Complexity and high-frequency PSD decreased in sleep compared to wake (wake > N1 > REM sleep > N2 > N3), whereas delta PSD increased with sleep (N3 > N2 > REM sleep > N1 > wake). Details of the statistical comparisons can be found in Supplementary Table S5. Source data

Extended Data Fig. 10. Evolution of electrophysiological markers across sleep stages in healthy participants (HP).

Three complexity measures (the Kolmogorov Complexity -KC, the Permutation Entropy -PE, and the Sample Entropy -SE), one connectivity measure (weighted symbolic mutual information (wSMI) in the theta band), and five spectral measures (normalized power spectral densities (PSD) of delta, theta, alpha, beta and gamma frequency bands) were computed separately for the wake (N = 981 from 21 HP), N1 (N = 373 81 from 21 HP), N2 (N = 1339 from 20 HP), N3 (N = 166 from 8 HP), and REM (N = 451 from 10 HP) sleep stages in HP. Each dot indicates marginal means estimated by a mixed-linear model including sleep stage as an independent variable, marker as the dependent variable, and participant ID as a random variable. Error bars denote 95% confidence intervals. Complexity and high-frequency PSD decreased in sleep compared to wake (wake > N1 > N2 ≈ REM > N3), whereas delta PSD increased with sleep (N3 > N2 ≈ REM > N1 > wake). Theta PSD was higher in N1 and lower in N3 sleep. Details of the statistical comparisons can be found in Supplementary Table S6. Source data