Weakly encoded memories due to acute sleep restriction can be rescued after one night of recovery sleep

Sleep is thought to play a complementary role in human memory processing: sleep loss impairs the formation of new memories during the following awake period and, conversely, normal sleep promotes the strengthening of the already encoded memories. However, whether sleep can strengthen deteriorated memories caused by insufficient sleep remains unknown. Here, we showed that sleep restriction in a group of participants caused a reduction in the stability of EEG activity patterns across multiple encoding of the same event during awake, compared with a group of participants that got a full night’s sleep. The decrease of neural stability patterns in the sleep-restricted group was associated with higher slow oscillation-spindle coupling during a subsequent night of normal sleep duration, thereby suggesting the instantiation of restorative neural mechanisms adaptively supporting cognition and memory. Importantly, upon awaking, the two groups of participants showed equivalent retrieval accuracy supported by subtle differences in the reinstatement of encoding-related activity: it was longer lasting in sleep-restricted individuals than in controls. In addition, sustained reinstatement over time was associated with increased coupling between spindles and slow oscillations. Taken together, these results suggest that the strength of prior encoding might be an important moderator of memory consolidation during sleep. Supporting this view, spindles nesting in the slow oscillation increased the probability of correct recognition only for weakly encoded memories. Current results demonstrate the benefit that a full night’s sleep can induce to impaired memory traces caused by an inadequate amount of sleep.

Before applying the automatic algorithm, the EEG signal was band-pass filtered with a FIR between 13 and 16 Hz. Then, the root mean square (RMS) of the filtered signal was calculated at each sample point using a moving window of 0.2 s. The resulting RMS signal was smoothed with a moving average of 0.2 s. The threshold for SP detection in the RMS signal was set to 1.5 standard deviations of the filtered signal within the time interval established by the manual markers, as defined by the mean across EEG electrodes. A SP was detected when the RMS signal remained above the threshold for 0.5-3 s, and the beginning and end of the SP were marked at the threshold crossing points. For every detected SP, the peak and trough were defined as the maximum and minimum of the filtered signal (between the beginning and end of the SP) and the deepest trough was designated as the "SP peak" that represented the respective SP in time, i.e., the time point taken for referencing event correlation histograms. SP density, amplitude, duration, intensity (duration/amplitude) and dominant frequency was calculated for each EEG derivation.

Methods S2. Statistical analysis of STPS
For the STPS analysis, statistics were computed for every time point, and the time points whose statistical values were above a threshold (p = 0.05) were selected and clustered into connected sets on the basis of temporal and spatial adjacency. The cluster-level statistics were calculated by taking the sum of the statistical values within a cluster (e.g., t-or z-statistic). After applying a primary threshold to voxels, p-values (α = 0.05) were assessed by means of 10,000 permutations using the Monte Carlo method implemented in the Fieldtrip toolbox. Next, the maximum cluster statistics over all six regions were chosen to construct a distribution of the cluster-level statistics under the null hypothesis. The nonparametric statistical test was obtained by calculating the proportion of randomized test statistics that exceeded the observed cluster-level statistics.

Methods S3. Estimation of effect size and confidence intervals
To estimate the standardized effect size for the mean difference between groups, we first computed Cohen's d (Cohen, 1992) by dividing the mean difference by the pooled standard deviation and next applied the Hedges correction (Hedges' d) (Cumming, 2012) for small samples. Although Cohen classified standardized effect sizes as small (d = 0.2), medium (d = 0.5), and large (d ≥ 0.8), these values are arbitrary. In order to facilitate their interpretation, we converted these standardized effect sizes into a percentage by calculating the common language (CL) (McGraw & Wong, 1992). It expresses the probability that an individual from one group (or treatment) has a higher value on one measurement than a person from the other group (or as compared to another condition). To estimate the standardized effect size between two mean dependent samples, we first computed Cohen's d rm , and next applied the Hedges correction (Hedges' g m ) (Lakens, 2013). For correlation analyses, we report the correlation coefficient r as a standardized measure of effect size. In those cases in which a mixed ANOVA was applied, the partial eta squared (η 2 p ) was computed as a measure of the effect size for the within-subjects factor. This index expresses the sum of squares of the effect in relation to the sum of squares of the effect plus the sum of squares of the error associated with the effect (Keppel, 1991).

Results S1. Effects of sleep restriction on behavioral performance during training
The NSD group (1.85 ± 0.55) and ASR group (1.57 ± 0.51) showed similar scores on the Epworth Sleepiness Scale during the training session (t (25) = 1.34,p = 0.19,d [CI 0.95 ] = 0.51 [-0.09 0.66], CL = 0.37). Although this result was unexpected, the arousal level associated with the task instructions and the subjective nature of the sleepiness measure may partially account for the lack of group differences (Belenky et al., 2003;Leproult et al., 2003).

Results S2. Contribution of encoding STPS to recognition memory
We evaluated whether the stability of EEG activity patterns during training, revealed by the strength of content-specific STPS across repetitions of face-face pairs, contributed to recognition memory. In particular, we compared the encoding STPS for subsequently remembered and forgotten events associated with the 1 st vs. 2 nd , 2 nd vs. 3 rd , and 3 rd vs. 4 th repetition.
The first contrast yielded greater STPS between the 1 st and 2 nd repetition of subsequently remembered paired-associates compared to subsequently forgotten pairedassociates over central areas between 340 and 460 ms (t (25) = 5.79, p cluster-corrected = 0.038, The second contrast (i.e., STPS between the 2 nd and 3 rd repetition) also revealed one early cluster associated with successful recognition over frontocentral areas between The stability of EEG activity patterns in response to the 4 th repetition with respect to the previous presentation indicates that too much early reactivation of neural activity associated with previously encoded stimuli might be detrimental to long-term memory while the later-in-time reactivation might be beneficial; likely because a different kind of information is being reactivated at each moment. Consistent with this interpretation, Zhang et al. (2018) found that subsequent memory was only accounted for by stimulusspecific activity occurring between 500 and 1200 ms, but not by the neural activity occurring earlier in time. As predicted by the levels of processing framework (Craik & Lockhart, 1972), it is likely that differences observed in the late time interval reflect the contribution of deep semantic processing to more elaborate, longer-lasting, and stronger memory traces than the ones that would be produced by shallow levels of processing in earlier time intervals.
Additionally, we found that although the encoding STPS associated with the difference due to memory in the early and late time intervals were tightly related to each other across all participants (F 1,31 = 30.2, p = 0.000005; for more details see Results S2 in supplementary material), only the STPS associated with remembered events in the late time window (Fig. S2D) was positively correlated across all participants with the index of recognition accuracy d' over frontal and right parietooccipital areas (r (25) [CI 0.95 (Fig. S3A). The false alarm rate was the main determining factor of this association (Fig. S3B). Accordingly, the STPS associated with the difference due to memory was positively correlated with the false alarm rate (but not with the hit rate) within the same spatiotemporal window. The cluster did not survive FWE correction but the size effect was significant (r (25) [CI 0.95 ] = -0.43 [-0.92 -0.12], p uncorrected = 0.14, CL = 1.67). The same analyses were performed for the 1 st vs. 2 nd and for the 2 nd vs. 3 rd repetition, but the stability of EEG activity patterns across repeated study only contributed to improve recognition accuracy after the 4 th presentation of face-face pairs. Table S4 shows the polysomnography-derived sleep parameters in the two groups, and Table S5 the parameters defining SOs and fast SPs. No group differences were found for any parameter related to macrostructure of sleep after applying Bonferroni correction. The lack of group differences in N3% is consistent with results of a previous study in which the percentage of SWS recovered to baseline levels after four consecutive nights of 3 h of sleep restriction (Wu et al., 2010). The parameters defining SOs and fast SPs neither survived Bonferroni correction, with the exception of the duration of the depolarizing SO up-state. Differences in the up-state duration are particularly interesting because previous studies have related the length of the up-state with memory improvement (Heib et al., 2013). The authors of this study concluded that the longer duration of the depolarizing component of SOs might provide a longer time window for replaying and transferring the recently encoded memories to longer-lasting storage. On the contrary, recognition accuracy was not related to this parameter in our study; this was the case when all participants were combined (r (25) [CI 0.95  Next, we tested group differences in the SO-SP coupling. As previously reported (Mölle et al., 2002(Mölle et al., , 2011Fogel & Smith, 2011;Staresina et al., 2015), event correlation histograms revealed a well-defined temporal relationship between fast SPs and SOs in both groups. Fig. S4A shows the temporal grouping of both frontocentral and centroparietal fast SPs by SOs in each group. With reference to the negative half-wave peak of the SOs, fast SP counts were suppressed around the down-state and increased during the up-state in the two groups. Although the ASR group showed a higher temporal coordination of fast SPs by SO up-states as compared to the NSD group (Fig.   S4A), permutation testing did not reveal significant differences between groups, likely motivated by the small sample size. Accordingly, the effect sizes reached statistical significance for most time points in the SO up-state interval but not in the down-state period ( Fig. S4B-C), regardless of whether SPs were detected over frontocentral (effect sizes ranged from 0.35 to 0.65) or centroparietal electrodes (effect sizes ranged from 0.38 to 0.88).

Results S4. Contribution of STPS E-R to memory performance
As illustrated in Fig. S5A, the STPS E-R over the first 150 ms over right frontal and bilateral parietooccipital areas was mainly associated with forgotten events (t (25) = -4.86, p cluster-corrected = 0.02, d rm [CI 0.95  in the two groups, early reinstatement (i.e., first 500 ms) associated with recognition failure and late reinstatement (i.e., 400-600 ms) associated with successful recognition was mainly evident in the ASR group (Fig. S5B-C). Table S1. Artifact-free trials included in the analyses. STPS 3 rd vs. 4 th (training) 11.5 ± 4.1 11.9 ± 3.9 STPS E-R 11.8 ± 4.6 11.6 ± 3.9

Contrast
Results are expressed as mean ± SD (standard deviation). Results are expressed as mean ± SD (standard deviation). FA = False alarms in two or more consecutive trials; iCV = intra-individual coefficient of variation; CL = common language Table S3. Post-training sleep parameters in the NSD and ASR group. Results are expressed as mean ± SD (standard deviation). TST = total sleep time; SOL = sleep onset latency; SE = sleep efficiency. *These data show a significant deviation from normality, so the median, the Mann-Whitney U statistic, and the median difference are reported. Results are expressed as mean ± SD (standard deviation). * p Bonferroni-corrected < 0.008 Figure S1. EEG montage. Scalp EEG electrodes were grouped into six regions. To obtain more stable spatial patterns, the electrodes in the border of two regions were included in both regions. Reprinted with permission from publication (Lu et al., 2015).   The same as in B for the SO up-state. Note that most effect sizes were significant only in the upstate interval. Figure S5. Contribution of encoding-retrieval STPS to subsequent memory. Withinsubjects STPS, expressed as averaged z-values, between the EEG activity patterns elicited by the and 4 th repetition of paired associates at encoding and the same pairs presented at retrieval for subsequently remembered (REM) and forgotten (FOR) paired-associates in the recognition task across all participants (A) for the NSD group (B) and for the ASR group (C). The x-axis represents time, and the y-axis the spatial locations shown in Fig. 2A. The statistics of contrasting STPS between remembered and forgotten paired-associates is shown in the right panel. The red and blue squares refer to significant clusters resulting from comparing the two memory conditions. The NSD group showed greater STPS E-R for remembered than for forgotten