Replay of large-scale spatio-temporal patterns from waking during subsequent NREM sleep in human cortex

Abstract

Animal studies support the hypothesis that in slow-wave sleep, replay of waking neocortical activity under hippocampal guidance leads to memory consolidation. However, no intracranial electrophysiological evidence for replay exists in humans. We identified consistent sequences of population firing peaks across widespread cortical regions during complete waking periods. The occurrence of these "Motifs" were compared between sleeps preceding the waking period ("Sleep-Pre") when the Motifs were identified, and those following ("Sleep-Post"). In all subjects, the majority of waking Motifs (most of which were novel) had more matches in Sleep-Post than in Sleep-Pre. In rodents, hippocampal replay occurs during local sharp-wave ripples, and the associated neocortical replay tends to occur during local sleep spindles and down-to-up transitions. These waves may facilitate consolidation by sequencing cell-firing and encouraging plasticity. Similarly, we found that Motifs were coupled to neocortical spindles, down-to-up transitions, theta bursts, and hippocampal sharp-wave ripples. While Motifs occurring during cognitive task performance were more likely to have more matches in subsequent sleep, our studies provide no direct demonstration that the replay of Motifs contributes to consolidation. Nonetheless, these results confirm a core prediction of the dominant neurobiological theory of human memory consolidation.

Figure 1

Spatio-temporal cortical activity Events and Motifs. (a) Example Waking Event, comprising a sequence of HG peaks occurring within 2s. A.U.: arbitrary units. (b) Example spatio-temporal Motif of HG peaks. Events were spatially- and temporally-clustered to obtain 60–250 representative Motifs for each subject, each with >20 Events (see “Motif selection and matching index analysis” in Methods and Fig. S1). Motifs with the most representative temporal order were matched against Events found in NREM (Fig. 3a). (c) Example NREM Event whose HG peaks come from channels involved in the Motif to the left and are in the same temporal order as the same channels’ HG peaks in the Motif.

Figure 2

Waking Events and Motifs do not arise from chance, and are spatio-temporally similar. (a) The number of Events obtained from actual data is less than expected numbers from chance. Pseudo-Events were obtained from 1000 data shuffles (with the inter-peak intervals of each channel randomly permuted). For each subject, the distribution of ratios between the number of Events from shuffles and the actual number of Events is summarized. In this and all panels, the red horizontal line marks the mean, and the blue vertical line marks the standard deviation (SD). (b) The number of Motifs obtained from actual data is greater than expected numbers from chance. Pseudo-Motifs were obtained from >300 of the shuffles performed in (a) The distribution of ratios between the number of pseudo-Motifs from shuffles and the actual number of Motifs is summarized similar to (a). (c) The number of channels involved in waking Events are thus greater than expected from chance. The distribution of ratios between the mean numbers of channels from pseudo-Events in (a) and the actual means are summarized as in (a and b). (d,e) Events are more similar spatio-temporally to the Motifs they matched to than the other Motifs. (d), The distribution of spatial similarity indices calculated for Motif-to-Event matches (with red mean lines) and the distribution of indices calculated between Motifs and Events that are not matched to them (with green mean lines). The spatial index for each Motif-Event pair is defined as 2 x (number of channels in common)/(total number of channels in both the Motif and the Event). (e), The distribution of temporal similarity indices for the Motif-Event pairs from (d). The temporal index for each Motif-Event pair is defined as [1 + (m − n)/(m + n)]/2, where m is the number of Event channel pairs in the same temporal order as the same channels in the Motif, and n is the number of such channel pairs in the opposite temporal order. For both (d and e, the blue distributions are significantly different from the black (two-sample Kolmogorov-Smirnov test, p < 0.0001), and the blue distributions have significantly greater medians (Mann-Whitney U test, p < 0.0001).

Figure 3

Events similar to waking cortical spatiotemporal activity Motifs occur more often in Sleep-Post than Sleep-Pre. (a) Schematic for Motif-Event matching. HG activity Motifs in the waking period W₀ were matched to the Events found in four consecutive NREM periods: Sleep-Pre (S₋₂, S₋₁) and Sleep-Post (S₁, S₂). (b–f), Motif-Event matching results. Among waking Motifs with consistent (i.e. more Sleep-Pre than Sleep-Post matches for both S₋₂ and S₋₁, or vice versa) matches found across four nights, a higher percentage has more Sleep-Post matches than Sleep-Pre matches. The data are counts, and the null hypothesis is that the counts of Motifs showing greater occurrence in Sleep-Post than Sleep-Pre would be equal to those showing the converse, so a simple binomial test was performed on each subject. In (b) only Motifs found in W₀ that did not occur in W_-1 were used; for (b,c and d) exact matches were required; for (e) and (f) significant non-exact matches were also accepted; (b,d and f) only include waking Motifs that passed a significance test (see main text). SEEG subjects 5 and 6 did not have sufficient channels for the analyses in (b,d and f). Red lines represent chance (equal percentage of patterns with more Sleep-Post or more Sleep-Pre matches). (2-tailed binomial tests: (b) **p = 0.009, ***p < 0.0001; (c) *p = 0.0368, **p = 0.0022 for 2, p = 0.0061 for 4, ***p < 0.0001 for 3 and 6, p = 0.0007 for 5; (d) ***p < 0.0001; (e) p = 0.3806 for 4, **p = 0.0048 for 1, p = 0.0035 for 2, p = 0.0036 for 5, ***p < 0.0001; (f) ***p < 0.0001). (g) The occurrence of exact matches across four nights (S) and one waking period (W) to waking Motifs that matched to more Events in Sleep-Post than Sleep-Pre. Histogram bin size: 8 min.

Figure 4

Motif occurrence enrichment in cognitively rich periods. (a) Motifs that matched more often to Events in Sleep-Post than to Events in Sleep-Pre (Post-over-Pre Motifs) preferentially occurred during task times in waking. For each randomly chosen 5-minute segment in waking, we calculated the following ratio: the number of Events similar to Post-over-Pre Motifs over the number of Events matched to all Motifs. 2000 random waking segments were selected, half within and half outside of the time for task performance. *p < 0.001 (Wilcoxon rank-sum test). (b) Motifs whose occurrences (i.e. waking Event matches) overlapped with face-including movie frames were more likely than non-overlapping Motifs to have more matches in Sleep-Post. *Two-tailed p-value (Chi-square test) = 0.0292.

Figure 5

Putative replay frames are associated with spindle, delta, and theta activity. (a,b) Neocortical activity during sleep triggered on peaks of putative replay Events (i.e., Events matched to waking Motifs with “exact matches” criterion and no shuffle thresholds). Cortical theta, alpha, beta and gamma power increases for up to 1000 ms during putative replay events. Green mask for p > 0.05 (two-tailed bootstrap) from baseline period (−1500 ms to −1000 ms). Vertical solid lines/horizontal dash lines indicate shared X- or Y- axes across subplots, respectively. (c,d) Peri-HG-peak histograms of theta and spindle occurrences. Theta centers preferentially occur prior to Event peaks (Sub. 1–5 grand average: p < 0.001 from 600 ms before to 200 ms before Event HG peaks; FDR-corrected permutation tests), and spindles preferentially occur near Event peaks (Sub. 2–3: p < 0.001 in the 0.2 s before the HG Event peaks; FDR-corrected permutation test). Orange boxes: significant time stretches. Putative replay frames are from the first Sleep-Post period. Similar effects were found for the second period (not shown). The black horizontal dash line indicates median ripple rate, and the orange horizontal dash line indicates the 99^th percentile.

Figure 6

Putative replay frames are associated with hippocampal activity. (a,c) Examples of automatically detected hippocampal ripples (in orange squares, which covers 125 ms) and the surrounding LFP from subjects 2, 3, and 5, respectively. (d) Hippocampal LFP average triggered on 1^st cortical HG peaks of matched Events (exact matches, no shuffle thresholds). (e) Aggregate time-frequency plot with all three subjects (470 randomly selected trials each); broadband hippocampal activity increases for ~1000 ms during matched Events. Power from 1–200 Hz (linear scale), with green mask for p > 0.05 (two-tailed bootstrap) from baseline period (−1500 ms to −1000 ms). (f,g) Time-frequency analysis of automatically detected ripples from posterior hippocampus shows power in the ripple frequency range (70–100 Hz). Vertical solid lines/horizontal dash lines that go across panels indicate shared X- or Y- axes across subplots, respectively. (h,i) Ripple occurrence rates become elevated near and after Motif-Event match peaks. The black horizontal dash line indicates median ripple rate, and the orange horizontal dash line indicates the 99^th percentile. Time windows that show significance are described here as [start,end] in milliseconds, where negative/positive values refer to the temporal distance from trigger times: p = 0.039 at [−2000, 2000] and p < 0.001 at [0, 4000] for Sub. 5; p = 0.013 at [−2000, 2000], p = 0.0446 at [0, 4000], p = 0.039 at [1600, 2000] and at [2400, 2800] for Sub. 3 (FDR-corrected permutation tests). Orange boxes: significant time stretches. Due to the long time-base, only the first HG peak from a given Event would be used in this analysis. All Events are from the first Sleep-Post period. Similar effects were found for the second period (not shown).

Figure 7

Participation of widespread cortical areas in Motifs. Spheres represent ECoG electrodes, with circles (o), plusses (+), or stars (☆) on the spheres indicating that they showed a significant deviation from chance (FDR corrected p < 0.05) in their occurrence rate, in their temporal order in the motifs, or in both, respectively. Specifically, a circle indicates that the amount of high gamma peaks seen at this cortical location was significantly different (more or less) from expected value based on chance alone. Similarly, a plus indicates that the likelihood for this cortical location to have high gamma peaks in the first quarter of an Event was significantly different (higher or lower) from chance, i.e. 25%. The size of each sphere indicates HG peak occurrence rate; color indicates the likelihood of HG peaks being early versus late in an Event (red: early, blue: late). Thin red lines connect the electrode pairs that produced temporally adjacent HG peaks within Events. Yellow lines connect the ten electrode pairs that produced the largest numbers of temporally adjacent HG peaks. Plotted are HG peaks from Motifs occurring in all waking and sleep periods (i.e. from both waking and NREM (the two Sleep-Post periods) Events that matched to waking Motifs with higher occurrence in Sleep-Post than in Sleep-Pre).