Sleep unbinds memories from their emotional learning context, protecting them from emotional interference due to a change of mood between learning and recall. According to the ‘sleep to forget and sleep to remember’ model, emotional unbinding takes place during rapid eye movement sleep. To test this hypothesis, we investigated emotional contextual interference effects after early versus late post-learning sleep periods, in which slow wave and rapid eye movement sleep, respectively, predominate. Participants learned a list of neutral word pairs after induction of a happy or a sad mood, then slept immediately afterwards for 3 h of early or late sleep under polysomnographic recording, in a within-subject counterbalanced design. They slept for 3 h before learning in the late sleep condition. Polysomnographic data confirmed more rapid eye movement sleep in the late than in the early sleep condition. After awakening, half the list was recalled after induction of a similar mood than during the encoding session (non-interference condition), and the other half of the list was recalled after induction of a different mood (interference condition). The results disclosed an emotional interference effect on recall both in the early and late sleep conditions, which does not corroborate the hypothesis of a rapid eye movement sleep-related protection of recent memories from emotional contextual interference. Alternatively, the contextual demodulation process initiated during the first post-learning night might need several consecutive nights of sleep to be achieved.
Recollection of personally experienced events involves the retrieval of two main aspects of episodic memory, namely the retrieval of what events happened (item memory) and the retrieval of the context in which these events have been encoded (context memory) (Graham and Cabeza, 2001). Context memory stores various information including, for instance, details of the physical environment in which the events occurred, the voice in which information was presented or the learner's emotional state. Context memory allows distinguishing one stored event from similar events co-existing in memory, and provides cues for item retrieval. Consequently, recollection of memorized items is enhanced when recall of the learned information occurs in the same context than at learning. This ‘context-dependent memory’ phenomenon was observed for multiple dimensions of the contextual background of learning, including physical environment (Godden and Baddeley, 1975; Smith, 1979), room size, odours and background music (Parker et al., 2007) and mood state (Bartlett and Santrock, 1979; Kenealy, 1997).
It is known that sleep contributes to the offline consolidation of item memories (Peigneux and Smith, 2010). However, its role in the processing of contextual information has received less attention, with inconsistent results. In one study (van der Helm et al., 2011), participants had to create an association between a poster on the wall (context) and a list of words (item). Two lists were learned, each associated with a different context. Although item memory was similar, context memory was enhanced selectively after the nap, suggesting sleep-dependent consolidation of the context. Conversely, in another study (Cairney et al., 2011) participants learned two lists of words, each list in a room differing in size, odour and background music. Here, recall was better after a 12-h sleep than wakefulness interval, but only when testing occurred in a different room than at learning. These results cast doubts about whether post-learning sleep consolidates contextual memories. Alternatively, it is possible that sleep contributes to the decontextualizing of recently learned information, eventually protecting it from the negative impact of contextual variations at retrieval. Accordingly, we reported that three nights of sleep unbinds items memory from their implicit emotional learning context (i.e. an induced mood), protecting them from a retroactive emotional interference effect at delayed recall (Deliens et al., 2012). In this study, subjects learned word pairs after induction into a specific mood, then slept or stayed awake during the post-learning night. Three days later, half the list was recalled after induction into a similar mood than at the encoding session, the other half after induction into a different mood. Performance was higher for items recalled in the same mood (i.e. a mood-dependent retrieval effect), but only in participants deprived of sleep after learning. This finding was in agreement with the ‘sleep-to-forget and sleep-to-remember’ (SFSR) proposal (van der Helm and Walker, 2009) that sleep facilitates the decoupling of memories from their emotional context, eventually strengthening item memory (the word pairs) while removing their affective blanket (the emotional context created by the mood induction procedure).
In the present study, we tested an additional prediction of the SFSR model, i.e. that the specific neurobiological conditions of REM sleep promote this decontextualization process, a hypothesis that could not be tested in our previous study (Deliens et al., 2012) in which polysomnographic recordings were not obtained. To test this hypothesis, we compared recall performance in the same or in a different mood than at learning after early versus late post-learning sleep periods dominated, respectively, by slow wave sleep (SWS) and REM sleep. According to the SFSR hypothesis, suppression or reduction of the effects of contextual emotional interference on recall was expected in the late (REM sleep-dominant) but not in the early (SWS-dominant) sleep condition.
Twenty young adults gave their informed written consent to participate in a partial sleep deprivation experiment approved by the local Ethics Committee. Due to technical problems during polysomnography recording or floor effects in the memory task (recall score at delayed phase <10%), two subjects were excluded from further analyses. The remaining 18 subjects [mean age ± standard deviation (SD) 22.39 ± 2.99 years; seven males] met the following criteria: native French speakers, non-smokers, no history of neurological disorders or of sleep disorders (Pittsburgh Sleep Quality Index total mean score ± SD 4.61 ± 1.72, Buysse et al., 1989; Fatigue Severity Scale 2.73 ± 0.69, Krupp et al., 1989), absence of mood disorders (Beck's Depression scale <7, Beck et al., 1988), intermediate or neutral chronotype (morningness–eveningness questionnaire: range 38–64, Horne and Ostberg, 1976) and verbal intelligence within normative values (Mill Hill Vocabulary Scale 23.39 ± 3.74, Deltour, 1993). Participants were required to keep regular sleep patterns during the week before and throughout the experiment, and to abstain from drugs, alcohol or stimulant (e.g. caffeine, energizers, etc.) consumption throughout the duration of the study. To control for the regularity of sleep habits, they were asked to wear an actimeter (Daqtometer; Daqtix GbR, Oetzen, Germany) and to complete daily sleep logs during the entire experiment. All subjects participated both in the early sleep (ES) and late sleep (LS) conditions spaced 2 weeks apart in a counterbalanced randomized order.
Word pairs memory task
Two parallel lists of 40 French unrelated word pairs [e.g. cheval – salon (horse – lounge)], composed of emotionally neutral nouns, were assigned randomly across individuals to the early and late sleep conditions (for detailed information see Deliens et al., 2012).
Mood induction procedure
The mood induction procedure is the translated French version of the original combined imagery vignettes and music procedure created by Mayer et al. (1995). Subjects listened for 1 min to one of six classical music pieces (Deliens et al., 2012), inducing either sadness or happiness. While listening, subjects had to perform a guided imagery task on eight vignettes presented on the computer screen. Each vignette was a descriptive sentence of a sad or joyful situation (e.g. ‘A pet you were really fond of has died’). Subjects were asked to imagine themselves emotionally engaged in each situation. The affect grid (Eich and Metcalfe, 1989), the Brief Mood Introspection Scale (BMIS; Mayer and Gaschke, 1988) and a visual analogue scale for valence (VAS)-mood, were administered before and after each mood induction procedure to assess changes in the emotional state attesting the effectiveness of the mood induction procedure. The affect grid is a 9 lines × 9 columns matrix in which the horizontal axis represents emotional valence and the vertical axis indicates arousal level, on a scale ranging from –4 (very sad/non-aroused) to + 4 (very happy/aroused). The square in the centre of the matrix represents neutral feelings. Subjects were asked to tick the square that fitted their current feelings most closely. The BMIS is a self-evaluation questionnaire composed of 16 adjectives (four adjectives specific to each mood state: happiness, serenity, sadness and anger). To each adjective, subjects allocated a score ranging from 1 (disagree) to 4 (agree). Mean scores were calculated for the four groups of adjectives. For the assessment of the valence of the VAS-mood, subjects were asked to place a mark on a 100-mm horizontal line to indicate their mood perception between the two extreme positions, labelled ‘very sad’ and ‘very happy’.
After 1 week of actimetry recording, subjects were tested during two nights according to the two experimental conditions: early (ES) or late (LS) sleep. Each night started at 21:30 hours with the preparation of the subject for polysomnographic recording.
In the ES condition, subjects learned word pairs between 22:30 and 23:30 hours. Lights were turned off at 23:30 hours and subjects were then allowed to sleep under polysomnography (PSG) recording for 3 h. The recall session took place at 03:30 hours, 1 h after awakening, to ensure dissipation of sleep inertia (Achermann et al., 1995).
In the LS condition, subjects first slept for 3 h (from 23:30 to 2:30 hours). One hour after awakening, between 3:30 and 4:30 hours, they learned the word pairs. They then returned to bed for 3 h more of sleep, after which they were reawakened. Recall performance was tested 1 h later (between 7:30 and 8:30 hours).
In the study phase, subjects were asked to complete the affect grid, BMIS and VAS-mood (mood baseline). They were then administered the mood induction procedure (either sadness or happiness, counterbalanced). At the end of the induction procedure, they again had to complete the three mood questionnaires before learning the AB list. In this study phase, the 40 unrelated word pairs were displayed one by one on a computer screen. Immediately after the learning session, there was an immediate cued recall procedure: the first word of each pair appeared on the computer screen and subjects had to write down the associated word. After the subject's response, the computer provided the correct answer to help error-free consolidation of items. This cued recall procedure was repeated until subjects were able to recall at least 30 word pairs (i.e. a 75% learning criteria). When a pair was completed correctly, it was not presented in subsequent trials to avoid over-consolidation through repetition. The presentation order of the word pairs changed after each presentation of the list to prevent serial learning. During the learning session and immediate recall phases, the music piece used in the induction procedure was played in a continuous loop to help to keep subjects in the induced mood. Finally, emotional state was again assessed using the three mood scales.
In the recall session, subjects were administered a cued recall procedure in which half the pairs were recalled after a mood induction procedure congruent with the emotional context at the encoding session [non-interference context (NIC)]. The other half was then recalled after induction into a different mood [interference context (IC)]. Although this might induce an order effect, a fixed order (NIC then IC) was chosen to prevent contamination in NIC by prior presentation of an interfering contextual emotional background in IC. In the NIC, subjects underwent the same mood induction procedure as at encoding, except that the piece of music was different from the encoding session but aimed at inducing the same mood. Changing the music aimed to preclude the hypothesis that any interference effect would be inherent to the musical piece per se. After having completed the mood induction procedure, subjects continued listening to the music piece and provided the associated word upon presentation of the first word of the pair for 20 pairs from the learned list AB. In the IC, the mood induction procedure was aimed at inducing the opposite mood to that at encoding, after which the 20 other word pairs were recalled in the same condition. The words subjected or not to interference were allocated randomly at delayed recall.
The two recall phases were separated by a 20-min break to minimize interference from the first mood induction onto the second one. During this break, subjects had to complete the Karolinska Sleepiness Scale (KSS; Akerstedt and Gillberg, 1990) and were administered the psychomotor vigilance task (PVT; Dinges and Powell, 1985) and two occupational tasks. The efficacy of the mood induction procedure was assessed using the three mood scales before and after each induction and after each recall session.
The following week, subjects repeated the same procedure using a parallel version of list AB, while being engaged in the other post-training sleep condition (i.e. ES or LS). Mood induction at learning was also opposite to the mood induced during the preceding week (i.e. sadness versus happiness). In other words, participants were tested once with the sad mood followed by LS and once with the happy mood followed by ES; or once with the happy mood followed by LS and once with the sad mood followed by ES. Moreover, half of the participants began with the LS condition and the other half with the ES condition. To control alertness levels throughout the experiment, subjects were administered the PVT and the KSS before each learning and test session. An overview of the experimental design is illustrated Fig. 1.
EEG recordings and analysis
Three EEG leads (C3–A2, O1–A2, FP1–A2), submental electromyography (EMG) and two electro-oculograms (EOGs) were recorded at a sampling rate of 200 Hz. A low-pass filter (45 Hz) and a high-pass filter (0.75 Hz) were applied to EEG channels. In addition, we performed an EEG power spectral analysis derived from the central lead (C3–A2) by computing fast-Fourier transformation (FFT) with a frequency resolution of 0.25 Hz, using the PRANA® software package (PhiTools™, Strasbourg, France). Artefacts due to eye, muscle and movement activities were detected automatically prior to the spectral analysis. The EEG signal was then controlled visually for accuracy of the computerized artefact detection, and additional epochs of EEG recording containing additional artefacts were further discarded. After FFT, power spectra were averaged over 30-s intervals to match the sleep stage scoring and subsequent power of standard frequency bands was computed (δ: 0.75–4.5 Hz; θ: 4.5–8.5 HZ; α: 8.5–12.5; σ: 12.5–15.5; β: 15.5–22.5; γ: 22.5–45). Intervals containing more than 50% of artefacts were considered as missing data.
Statistical analyses were performed using Statistica version 7.0 (Statsoft Inc., Tulsa, OK, USA). Repeated-measures analysis of variance tests (anovas) were conducted and followed by Tukey's post-hoc tests when appropriate. Pairwise correlations were computed using Pearson's product–moment r. Tests were two-sided and carried out at a 5% significance level.
Sleep and vigilance parameters
Results from sleep recordings are summarized in Table 1. Mean sleep duration was similar during ES (2 h 45 min ± 20.17 min) and LS (2 h 44 min ± 13.64 min) conditions. Mean latency duration and percentage of time spent in Stage 1 did not differ between ES and LS conditions. As expected, REM sleep proportion was higher in the LS condition (28.04%, mean duration: 46.08 min) than in the ES (7.96%, 13.03 min; t(17) = 12.08; P <0.000) condition. Conversely, the proportion of SWS (N3) was higher in the ES condition (45.05%, 74.78 min) than in the LS [15.95%, 26.17 min; t(17) = 7.97; P <0.000] condition. Time spent in Stage 2 (N2) was also longer in the LS condition than in the ES condition [51 versus 42.8 min; t(17) = 3.01; P =0.008].
Table 1. Sleep parameters
Sleep during early versus late retention interval
Sleep during early retention interval
Sleep during late retention interval
Values are means ± standard deviation. Right columns show results for pairwise statistical analysis by t-test (***P <0.001; †P <0.01).
Separate repeated-measures anovas were conducted on subjective (KSS scores) and objective (PVT mean reaction-time) vigilance parameters with within factors sleep (ES versus LS) and session (IC versus NIC) (see Table 2). Analyses on self-reported sleepiness scores (KSS) revealed a main effect of sleep (F1,17 = 8.17; P = 0.01), a main effect of session (F1,17 = 4.73; P =0.015) and a trend for an interaction effect (F1,17 = 3.11; P =0.057). Tukey's post-hoc tests disclosed that subjects felt sleepier before the non-interference recall session than before the learning session (P =0.02), and a trend for more sleepiness before the interference recall than during the learning session (P =0.053). No differences in sleepiness levels were reported before interference or non-interference recall sessions (P =0.9).
Table 2. Vigilance parameters
Values are means ± standard deviation. Recall in the same emotional context (IC recall) or in a different emotional context (NIC recall).
Early sleep condition
330.5 ± 45.6
360.8 ± 69.2
376.0 ± 78.6
Late sleep condition
326.8 ± 29.7
343.3 ± 50.7
361.7 ± 62.9
Early sleep condition
5.5 ± 1.6
5.1 ± 2.0
5.5 ± 1.9
Late sleep condition
5.2 ± 1.5
3.8 ± 2.1
3.7 ± 1.9
Analyses on objective vigilance parameters revealed a main effect of session (F1,17 = 11.49; P =0.001) with lower vigilance levels before the non-interference and interference recall sessions than before the learning session (all Ps < 0.02). Vigilance levels before interference and non-interference recall were similar (P =0.13). There was no main effect of sleep (F1,17 = 1.17; P =0.3) nor interaction effect (F1,17 = 0.84; P =0.4).
Efficacy of the mood induction procedure (MIP)
Separate repeated-measures anovas were computed on affect grid, BMIS and VAS-mood scores with within-subject factors sessions (at encoding, congruent mood recall and incongruent mood recall) and mood-induced sessions (happy versus sad). These analyses failed consistently to disclose a main effect of session (all Ps > 0.6), but revealed a main effect of mood with a more positive valence after happy MIP than after sad MIP (Ps < 0.03 for affect grid and BMIS scores, P =0.07 for VAS-mood), and interaction effects between session and mood (all Ps < 0.001). Tukey's post-hoc comparisons for the three mood scales confirmed similar mood at encoding and retrieval in the congruent mood condition for sad and happy mood (all Ps > 0.98) and different moods in the incongruent mood condition for sad and happy mood (all Ps < 0.01).
Effect of mood type at learning
Finally, we tested whether the mood type (happiness versus sadness) induced at learning had an impact on ulterior memory retrieval in the NIC. An independent-sample anova (n =18) with factors sleep (ES versus LS) and mood (happiness versus sadness) failed to disclose mood-related differences in recall performance (F1,32 = 0.07; P =0.79) or a sleep × mood interaction (F1,32 = 0.02; P =0.88).
Effects of sleep and emotional interference on memory consolidation
A repeated-measure anova was computed on the number of correctly recalled pairs with within-subject factors sleep (ES versus LS) and interference (IC versus NIC). This analysis disclosed a main effect of interference (F1,17 = 7.16; P =0.016), with higher recall in the NIC (15.92 ± 3.05 word pairs) than in the IC condition (14.83 ± 3.89 word pairs) (Fig. 2). However, results disclosed no main effect of sleep (F1,17 = 0.04; P =0.85) or interaction between the moment of sleep (ES versus LS) and interference (NIC versus IC) (F1,17 = 0.02; P =0.88). The size of the interference effect (NIC minus IC) was not different after early versus late sleep (t17=0.15; P =0.88). Also, no correlation was observed between REM or SWS duration and the size of the interference effect (all Ps > 0.16).
Finally, we tentatively computed the correlation coefficients between the size of the interference effect (NIC minus IC recall score) and the time spent in each sleep stage, and between interference and power in delta, theta, alpha and sigma frequencies. All correlations were non-significant (coefficients of correlation range −0.26 to +0.20; all Ps > 0.1). In the ES condition, a positive correlation was found between the size of the interference effect and (a) the rate of fast spindles (>13 Hz) in N2 sleep (r =0.671; P =0.002) and (b) fast spindles duration in N2 (r =0.675; P =0.002), whereas these correlations were non-significant in the LS condition (all Ps > 0.3).
We investigated how and whether post-training sleep modulates mood-dependent memory effects on the delayed recall of verbal memories, using an early versus late sleep protocol aimed at highlighting the effect of SWS- versus REM-dominant periods of post-learning sleep on resistance to emotional interference. Based on the SFSR model (van der Helm and Walker, 2009), we predicted a superiority of REM sleep for unbinding the learned item information from its emotional learning context. In line with previous studies (Bartlett and Santrock, 1979; Kenealy, 1997) and the Network Theory of Affect (Bower, 1981), we found a mood–state-dependent retrieval effect with higher recall when participants learned and recalled word pairs after the same mood induction than when the mood induction was different. In contrast to our working hypothesis, however, the mood–state-dependent effect was modulated neither by the early versus late sleep factor nor by its interaction with interference (NIC versus IC). Compared to previous studies (Baran et al., 2012; Groch et al., 2013), our study presents an original method to probe consolidation for neutral stimuli learned in an emotional contextual background, rather than for intrinsically emotional material. Indeed, using emotional stimuli makes it difficult to segregate changes pertaining to the consolidation of their intrinsic emotional value compared to the impact of their emotional envelope. Therefore, using neutral stimuli learned in an emotional background allowed us to test the impact of an incidentally learned emotional association, preventing the aforementioned bias.
The SFSR model proposes that the specific neurobiological imbalance dominating REM sleep contributes to the strengthening of recent memories while removing their learning-related affective tag. This emotional unbinding would be the long-term consequence of a sleep-dependent reprocessing that takes place through multiple REM sleep iterations, across the night and/or multiple nights (van der Helm and Walker, 2009). The fact that REM sleep did not dampen the mood–state-dependent retrieval effect in the present study might be due to the reduced number of REM sleep episodes obtained before the testing session during a 3-h sleep period, compared to the number of REM sleep episodes obtained over the several nights of sleep that might be needed to achieve emotional decontextualization. Nevertheless, we previously found a demodulation effect between item memories and their ‘affective blanket’ after three nights of sleep versus 1 night of complete sleep deprivation followed by two recovery nights (Deliens et al., 2012), suggesting an important role for the first night of post-training sleep in the initiation of this unbinding process. Further investigations are needed to clarify the temporal evolution of this phenomenon.
In the present study, we hypothesized that unbinding of the emotional context would occur during REM sleep. We therefore focused our analysis on the comparison between REM- versus SWS-dominant post-learning sleep episodes, an analysis that failed to disclose differences between conditions. Polysomnographic data additionally revealed higher amounts of N2 sleep in the late sleep condition, in addition to the expected higher proportion of REM sleep. A selective benefit of sleep for contextual memory was correlated with the amount of N2 sleep in the van der Helm et al. (2011) study, an effect replicated only partially in this study. Indeed, as a significant correlation between spindles and the size of the interference effect was observed only in the ES condition, it does not allow us to attribute a specific role to fast sleep spindles in the emotional demodulation process. Regardless, the mood-dependent retrieval effect was not higher in the late than in the early sleep condition in our study, contrary to what may have been expected from this association. The apparent discrepancy between studies may be explained considering several factors. First, REM sleep represented only a small proportion of the total sleep time in the van der Helm et al. nap study (± 10% of 120 min), actually no more than that obtained in our early sleep condition. Therefore, the amount of REM sleep in the van der Helm et al. (2011) study may have been insufficient to observe REM sleep-related effects. Secondly, and most importantly, the instructions provided before the encoding session were different. In the van der Helm et al. study, the association between items to memorize and contextual elements was emphasized explicitly: subjects were instructed to create an association between a poster on the wall (context memory) and the word to memorize (item memory). In our study, the mood induction procedure took place before the encoding session and the mood was maintained by the continuous presentation of music excerpts, but it was never mentioned that the emotional context had to be associated with the items to memorize. Consequently, explicit binding between context and item memory in the van der Helm et al. (2011) study makes it a sleep-dependent ‘source memory’ experiment (see e.g. Johnson et al., 2009; Lekeu et al., 2002), whereas our study concerns implicitly acquired contextual memories, as in the Cairney et al. (2011) study. In this context, sleep might strengthen selectively memories that are cued explicitly to be remembered during encoding (Rauchs et al., 2011; Saletin et al., 2011), even when relevance for future utilization is mentioned after the learning session (Van Dongen et al., 2012). In our current and previous studies (Deliens et al., 2012), subjects were instructed before the beginning of the encoding session that a retest would follow after 5 h (or 3 days), but nothing was mentioned about the emotional context. Accordingly, item memory (word pairs) targeted explicitly for consolidation was consolidated during sleep, but context memory (mood state) not pointed out as relevant for further testing was removed after three nights of sleep (Deliens et al., 2012). However, future relevance cannot be the sole explanation, as a mood-related interference effect was still present after 3 h of sleep (either early or late) in the present study, thus suggesting that decontextualization might occur across several NREM-REM sleep cycles as proposed in the SFSR model (van der Helm and Walker, 2009).
Although subjects were sleepier before the recall than the learning sessions in our study, sleepiness and vigilance parameters were similar in the NIC and IC recall sessions, ruling out a confounding impact on contextual interference. Also, mood induction before the IC recall session had a marginally weaker effect in the ES than LS condition. Participants were tested randomly in a sad or a happy mood in the LS or ES condition, and then in the opposite mood in the other condition 2 weeks later. This protocol makes it impossible to compare within-subject induction in the same mood in the LS versus ES conditions, and it cannot be assumed that happiness level on a mood scale is the perfect opposite of a similar sadness level on the same mood scale (see e.g. Rafaeli and Revelle, 2006). Thus, whether or not subtle differences in mood induction could annihilate interference effects remains an open issue. Finally, we chose not to use a wake control condition in this study. Although a circadian confound effect on performance cannot be ruled out entirely, sleep-related mood decontextualization was found in participants who learned and were tested at the same time of the day in our previous study, thus minimizing this potential confound (Deliens et al., 2012).
To sum up, our results confirm the mood–state-dependent retrieval theory positing that participants who learn and recall material in the same mood state have better recall performance than participants who learn and recall in a different mood. Also, we cannot conclude that a specific role of REM sleep protects memories from emotional interference, as predicted in the SFSR model. Further studies are needed to test the hypothesis that the demodulation process is actually initiated during the first post-learning night, potentially with a prominent role of REM sleep which may, however, need several nights to be achieved.
The authors thank Aurélie Le Clair for help in data acquisition. G.D. is Research Fellow at the Belgian Fonds National de la Recherche Scientifique (FRS-FNRS). This study was supported by FRS-FNRS ‘Credit aux Chercheurs’ grant 184.108.40.206.F.