Memory reactivation in slow wave sleep enhances relational learning.

doi:10.21203/rs.3.rs-2982884/v1

Download PDF

Article

Memory reactivation in slow wave sleep enhances relational learning.

https://doi.org/10.21203/rs.3.rs-2982884/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 08 Mar, 2024

Read the published version in Communications Biology →

Version 1

posted

You are reading this latest preprint version

Sleep has been shown to boost the integration of memories, and thus to facilitate relational learning. This benefit is thought to rely upon memory reactivation during non-REM sleep. We set out to test this possibility by explicitly cueing such reactivation using a technique called targeted memory reactivation (TMR), in which sounds are paired with learned material in wake and then softly played back to the participant in sleep, triggering reactivation of the associated memories. Specifically, we tested whether TMR during slow wave sleep leads to enhancements in inferential thinking in a transitive inference task. Because the Up-phase of the slow oscillation is more responsive to external cues than the Down-phase, we also asked whether stimulation at this specific phase is more beneficial for such integration. Our data show that Up-phase TMR boosts the ability to make inferences, but only for the most distant inferential leaps. Such stimulation was also associated with detectable memory reinstatement, whereas Down-phase stimulation produced no such trace and led to below-chance performance. These findings demonstrate that cueing memory reactivation at the right time points in sleep can provide a direct benefit to difficult relational learning problems.

Biological sciences/Neuroscience/Learning and memory/Consolidation

Biological sciences/Neuroscience/Learning and memory/Long-term memory

Memory reactivation in sleep is thought to be important for integrative thinking. We examined this by explicitly cueing reactivation of a transitive inference task during slow wave sleep using Targeted Memory Reactivation (TMR). Because TMR at different phases of the slow oscillation has different impacts, we cued different hierarchies at Up and Down phases. Up-phase TMR enhanced inferential performance and was associated with classifiable memory reactivation. Conversely, Down-phase TMR lead to a short-term impairment in inferential thinking and no detectable reactivation. These findings provide the first evidence that TMR can boost transitive inference and is thus important for integration and reasoning.

Relational memory is the ability to integrate multiple sources of knowledge, infer indirect associations between stimuli, and make decisions when presented with novel situations (Dymond and Llewellyn 2019; Lerner and Gluck 2019). One example of such integration is transitive inference (TI), or the deduction of the rankings of non-adjacent members of a linear hierarchy which has been presented via exposure to adjacent pairs. In simpler words, knowing A > B and B > C can allow deduction that A > C in an A > B > C hierarchy. Despite being studied for many decades in humans (Bryant and Trabasso 1971) and numerous other species (Camarena et al. 2018; Grosenick, Clement, and Fernald 2007; Lazareva et al. 2020) the mechanisms for TI remain elusive (Holyoak and Lu 2021; Morgan 2017). Furthermore, the dependence of TI on sleep has only been investigated relatively recently by a seminal study (Ellenbogen et al. 2007), and its replication (Werchan and Gómez 2013), which demonstrated that sleep is beneficial to this task, but only for the most distant inference pairs. While its role in TI remains unclear, sleep is thought to facilitate the abstraction of gist from recent experiences (Durrant, Cairney, and Lewis 2013; Lutz et al. 2017), and the integration of these with prior knowledge (Hennies et al. 2016; Tamminen et al. 2010). Prior learning is spontaneously reactivated during sleep, and this is important for such memory consolidation (Rasch and Born 2013). Importantly, memory reactivation can be directly cued by re-administering sensory stimuli that have been paired with learned information using a technique known as targeted memory reactivation (TMR), see (Hu et al. 2020) for a meta-analysis.

Although some prior work has linked cognitive performance enhancement with rapid eye movement sleep (REM), e.g. (Nolan 2010; Cai et al. 2009; Walker et al. 2002), most studies have focused on non-rapid eye movement sleep (NREM) and particularly on slow wave sleep (SWS), see (Born, Rasch, and Gais 2006; Schouten et al. 2017; Oudiette and Paller 2013; Rasch and Born 2013) for reviews. The main rhythms of SWS are slow oscillations (SO) and sleep-spindles. SOs are low-frequency oscillations at 0.5-4Hz that reflect alternation between hyper-polarized neuronal down phases and depolarized up-phases (Steriade, Nuñez, and Amzica 1993). SOs drive transient sleep-spindles at 9-16Hz (Steriade 2006) which are strongly linked to reactivation (Rasch and Born 2013). An elegant study (H.-V. V. Ngo and Staresina 2022) recently showed that cueing in the SO up-phase elicited more spindles and less forgetting than cueing in the down-phase, strongly suggesting that up-phase cueing is more likely to facilitate memory consolidation.

Here, we set out to investigate how reactivation of a TI hierarchy during SWS, and particularly during SO up and down phases, influences the ability to make inferences. To this end, we adapted the experimental set-up from (Ellenbogen et al. 2007) to include three hierarchies so we could apply closed-loop TMR (CL-TMR) in a within-subject design. Thus, one hierarchy was stimulated during the up-phase (Up condition), another during the down-phase (Down condition), and the remaining hierarchy was not stimulated (Control condition). Because recent work has shown that TMR effects can continue to unfold over time (Rakowska et al. 2021; Groch et al. 2017), our participants performed a third behavioral test two weeks after the manipulation. In keeping with other studies (Ellenbogen et al. 2007; Werchan and Gómez 2013), we expected cued reactivation in sleep to benefit only the inference pairs. We also predicted a benefit in the Up condition but not in the Down condition, and we expected the benefits to last, or even increase, over time.

Our data show that TMR in Up is associated with classifiable memory reinstatement and better inferential reasoning, and that this is strongest in the most distant inference pairs. On the other hand, stimulation in Down produced no evidence of reactivation and led to a temporary inhibition of inferential reasoning that recovered after two weeks.

Participants

Thirty adults (10 males, mean age 27 ± 3.72) participated in the overnight experiment. All of them with no-self reported history of neurological, sleep or motor disorders. All participants completed a screen questionnaire before selection, provided written informed consent, and were reimbursed for their time. The experiment was approved by the School of Psychology Ethics Committee at Cardiff University. Participants agreed to abstain from caffeine and alcohol during the study and for 24-hours before. From the thirty participants that completed the task, 10 were eliminated either because of technical problems (n = 3) or because they did not have enough stable SWS (n = 7) to perform the stimulations (we required 12 rounds, these participants were mostly in light sleep and N2 stage). From the 20 participants left, 3 could not finish Session 3 due to the pandemic.

Materials

The behavioral tasks were presented in a quiet room, participants were comfortable seated in front of the computer and stimuli were presented using Matlab©, Psychtoolbox (Brainard 1997) and Cogent 2000 (www.vislab.ucl.ac.uk). Three types of visual stimuli were presented to the participants: female faces (Lundqvist, Flykt, and Ohman 1998), outdoor scenes (taken from the internet) and unusual objects (Horst and Hout 2016), see Fig. 1B. Each stimulus was easily distinguishable from the others within and between categories. All items were presented in grey scale and matched for luminance. Each image was associated with an exclusive sound, semantically congruent with the image as closely as possible (e.g. bike image with a bike bell sound).

Sounds were taken from the internet and truncated into two different lengths, 2s and 200ms, and pitch normalized. We used the longer sounds in behavioral training, facilitating the sound-image encoding, and the shorter version for the rest of the behavioral tasks and TMR cueing. The sounds were played through noise-cancelling headphones (Sony MDR-ZX110NA) during behavioral tasks and through speakers (Dell A225) during sleep. The order of presentation of each stimulus category was counterbalanced across participants and the order of stimuli within each category was completely randomized for each subject. Hence, the experimenter was completely blind to which stimuli form the hierarchy and its order within each condition (Up, Down, Control), and which type of stimuli (faces, objects or scenes) was selected for each condition, so as not to influence the results. Each of the three hierarchies (Fig. 1B,C) comprised 6 images, each one with an associated (highly discriminable) sound. We prepared a set of 12 images and 12 sounds per hierarchy, that is a total of 36 images and sounds. At the beginning of the experiment, for each one of the three hierarchies, 6 of these images with their corresponding sounds were selected to be learned and the remaining 6 sounds were used as controls to be played during the TMR stimulation. Before participants started the first task, the 6 images which would be used during the experiment and the other 6 which would be used as control were randomly selected for each category (the experimenter was blind).

Procedure

Participants arrived at the laboratory around 8 pm and changed into their sleepwear. They reported alertness by completing the KSS (Åkerstedt and Gillberg 1990) and SSS (Hoddes, Dement, and Zarcone 1972) questionnaires. Afterwards, they were fitted for PSG recording and performed the initial training and the immediate test explained in “Experimental tasks” section and Fig. 1A. Participants were ready for bed around 11pm. During the night, the previously learned tones were played softly during SWS. From the three stimulus categories, one was kept as a control (Control) and was not played during the night, allowing us to compare it against the other two which were cued during the Up and Down phases of the SOs respectively. After 7–8 hours of sleep, participants were woken at the agreed time and allowed > 20 min to overcome sleep inertia. During this time, they could go to the toilet, eat, and drink before completing the sleep quality, KSS and SSS questionnaires. Participants then completed the Late test and another Sound-Image association test (“Experimental tasks” section and Fig. 1A). Afterwards, the electrodes were removed, and participants could shower or go home. Finally, participants came back to the laboratory two weeks later (± 2days) to complete the second Late test and Sound-Image association test, identical to the previous one but without EEG recordings, to test the robustness of sleep-TMR mediated benefits.

Experimental tasks by session

The experiment composed three sessions: evening (Session 1), next morning (Session 2) and a follow up session two-weeks after (Session 3). Each of the sessions was divided into different tasks as below:

a) Session 1: Sound-image association learning task: For each of the three categories, participants were shown each of the six items forming the category one by one. At the same time the associated sound (2s length) was played. Each sound-image pair was shown 4 times. The order of items within a category was randomized and the order of the categories themselves was counterbalanced across participants.

Sound-Image association test: Immediately after training, all participants performed a recall session to determine retention level. Three images were presented on the screen while a sound was played. Participants were asked to select as quickly and accurately as possible the image corresponding to the sound using the keyboard arrow keys. When they responded a rectangle surrounded the correct image (green if the participant’s selection was correct or red if it was wrong). Image screen position was randomized on every trial. The three images presented on the screen were pseudo-randomly selected, with the restriction that at least one of the two images was a ‘lure’ of the same category as the right answer. The sounds were cut down to only 200ms long. Participants performed two blocks with three repetitions of each sound per block. At the end of each block, accuracy was presented.

Premise pair learning task: Following previous related experiments (Ellenbogen et al. 2007; Werchan and Gómez 2013), all participants learned five relational premise pairs for each of the three categories. If each category formed a 6-item hierarchy, schematically represented as A > B > C > D > E > F (see Fig. 1C), the premise pairs would be: A > B, B > C, C > D, D > E, E > F. Where the notation "A > B" indicates "select A over B". The pairs were presented one at a time, with images stacked vertically (Fig. 1E). Subjects were instructed to select the item "hiding" a smiley emoticon from the two presented, at first by trial and error, but after practice and feedback they learned which item was correct. If they selected the correct item, it was replaced by a smiley emoticon. This is in line with (Werchan and Gómez 2013), where a smiling-emoticon was used as reinforcement. If they selected the wrong image it was replaced by an angry emoticon. After the feedback, participants received a second reinforcement as the pair was presented again but this time horizontally instead of vertically, and in the correct order (e.g. A-B) from left to right, with the corresponding sounds also played in the correct order. Pairs were organised into blocks of 10 trials for each of the hierarchies. This meant a total of 30 trials per block. Each block presented each of the five pairs of each hierarchy twice, counterbalancing the up-down positions (e.g. A above B and B above A, with A being the correct selection in both cases). The three hierarchies were not mixed within a block. For example, first all pairs for the “scenes” category were presented, then pairs in the “faces” category, and finally the “object” pairs. This order was counterbalanced across participants. Within each category, pairs were ordered pseudo-randomly to avoid explicitly revealing the hierarchy. Hence, a displayed pair cannot contain an item that was in the previous pair (e.g., A > B will never be followed by B > C). Furthermore, the order of the items within the hierarchy was randomly selected for each participant at the start of training, remaining unknown to the experimenter. At the end of each block, the overall performance for that block was shown on the screen to keep participants engaged with the task. All subjects underwent a minimum set of three blocks of training. After the third block, and every block thereafter, only performance of the "middle pairs", meaning B-C, C-D and D-E, was saved to calculate the exit criteria (Werchan and Gómez 2013). If the averaged performance of these pairs for two of the last three blocks was over 66% for one of the hierarchies, the participant stopped receiving feedback for that hierarchy. However, all the premise pairs of this category still appeared on the screen to ensure the same number of trials/appearances for each hierarchy. This continued until the participant reached criteria for all the three hierarchies or a maximum of 10 blocks. In contrast to (Ellenbogen et al. 2007) and (Werchan and Gómez 2013), where the exit criterion was set to 75% accuracy for the middle premise pairs, we used a criterion of 66% to avoid ceiling effects and increase the chances of overnight improvement. On the other hand, we added a more restrictive criterion of 2 blocks out of 3 meeting the threshold, to be sure that the criterion was not achieved by chance. Similarly, to the above-mentioned studies, we only counted the middle premise pairs to evaluate the exit threshold because they are the necessary items for building the inferences.

Immediate test: After criterion was met, participants enjoyed a 5-minute break before proceeding to the immediate test which assessed initial retention of the learned pairs. A similar protocol was used for testing and training with the exception that feedback and sound cues were removed in testing. Subjects were informed that they must select the right item based on previous learning. Participants performed four blocks, with 10 trials per hierarchy. Between blocks, participants solved arithmetic problems to clear short-term memory (von Hecker, Klauer, and Aßfalg 2019). Furthermore, pairs from the different hierarchies were randomly interleaved, always with the restriction of not showing the hierarchy explicitly.

b) Session 2: Late test: After filling the KSS and SSS questionnaires participants performed a similar test as before but this time, they were presented with previously learned premise pairs, new inference pairs, and one ‘anchor pair’ such that a total of 9 pairs were seen instead of just the 5 previously learned. The first new pairs were 3 inference pairs: B > D, B > E and C > E (see Fig. 1D). These pairs are named inference pairs because if you know that B > C and C > D, then you can infer that B > D. The inference pairs can be divided into 1st and 2nd degree of separation. That refers to the number of items between the ‘pair’ items, for instance between B and D is only one item (C), hence it has first degree separation, as does C-E. On the other hand, there are two items between B and E: C and D. Hence the B-E pair has a second degree of separation and is therefore the most distant pair within a 6-item hierarchy. Additionally, we also added a 4th pair, the anchor pair (A > F) as a control since inference is not needed to obtain this relationship. This is due to the fact that A is always correct and F is always incorrect (von Hecker, Klauer, and Aßfalg 2019). Participants were instructed that they might see novel combinations and if that was the case, they should try to make their best guess. At the end of each trial, they were confidence from − 2 (guessing) to + 2 (certain) using the up and down arrows. Following a similar protocol to Session 1, participants performed four blocks with math exercises between them.

Sound-Image association test: After a 5-minute break, subjects performed a new sound-image association test with the same structure as the Session 1’s test but without feedback.

c) Session 3: This used the same tasks in the same order as in Session 2. However, this time participants’ brain activity was not recorded. Finally, the participants performed an Awareness questionnaire (see Fig. <link rid="fig5">1</link>–1).

Closed-loop TMR protocol

The two categories that used for TMR and the control were counterbalanced across all participants (see SM3). The control category was not cued during the night. From the other two, one was assigned to Up and the other to Down. Stimulation started after participants entered stable SWS and was halted for arousals or any other sleep stage. Participants were exposed to an extra control-hierarchy of sounds for each of the two TMR conditions. These control-hierarchy-sounds, also 200ms duration, were completely novel to the participants and were included to allow us to distinguish the TMR effect from a normal brain ERP. Each one of the hierarchies, was composed of 6 items and played in order: A, B..., F. The order of both experimental and control hierarchies and of Up and Down cueing was randomized and counterbalanced across blocks. Each block comprised four hierarchies: Experimental Up, Control Up, experimental Down, control Down (see Fig. 1F). The minimum number of blocks to include a participant in the analysis was 12, this means 288 cues were presented during the night. Online detection of SOs was based on the detection of the negative half-wave peaks of oscillations. The electrode used as reference for the on-line detection was F3 as frontal regions are a predominant SOs area (Massimini et al. 2007), band-pass filtered in the slow-wave range (0.5-4Hz). When the amplitude of the signal passed a threshold of -80uV the auditory stimulus was delivered after a fixed delay of 500ms (H.-V. V. V. Ngo et al. 2013; H. V. V. H.-V. V. Ngo et al. 2015). Inter-trials intervals were set to a minimum of 4 seconds, that is after every sound played there was a minimum pause of 4s. The SO detection, auditory stimulation and presentation of the trigger to the EEG recording was via a custom-made Matlab-based toolbox (https://github.com/mnavarretem).

EEG recordings

Sleep was recorded using standard polysomnography including EEG, electromyographic (EMG) and electrooculography (EOG). EEG was recorded using a 64-channel LiveAmp amplifier (Brain Vision©). Electrode impedance was kept below 10KΩ and sampling rate was 500Hz and referenced to Cpz electrode. In addition to the online identification of sleep stages, polysomnographic recordings were scored offline by 3 independent raters according to the ASSM criteria (Berry et al. 2015), all of them were blind to the periods when the sounds were reactivated.

EEG analysis

Pre-processing and analysis were all performed with Fieldtrip (Oostenveld et al. 2011) and custom Matlab functions. Data were low-pass and high-pass filtered (30Hz and 0.5Hz respectively). Eye and muscle related artefacts were removed using independent component analysis (ICA). Bad channels were interpolated (spline interpolation) and data was re-referenced to linked mastoids. We calculated ERPs by segmenting the cleaned signal into 4 second segments, from − 1s before stimulus onset to 3s afterwards. A final visual inspection of the dataset was performed, and any residual artefact was manually removed. To calculate the differences between stimulation conditions we averaged across all trials, but for the classification analysis we kept the trial information. To study the time frequency evoked TMR response, we calculated the power spectrum of the signal locked to the TMR cue onset using Morlet wavelets from 4 to 20Hz with 0.5Hz resolution and a time window from − 1 to 2.4s in 50ms steps at the subject level. The width of the wavelet was set to at least 4 cycles per time window, adaptatively to the frequency of interest. Resulting TFRs were then expressed as the relative change of baseline from − 1s to 0ms pre stimulus onset.

Statistics

Statistical assessment of EEG data was based on nonparametric cluster permutation tests with the following parameters: 2,000 permutations, two-tailed, cluster threshold of p < 0.05, and a final threshold of p < 0.05 using Fieldtrip toolbox (Oostenveld et al. 2011). The time-frequency statistical analysis was restricted to the post-cue interval (0 to 2.4ms) to avoid the natural differences of the Up and Down phases of the SOs before cueing onset. To examine the accuracy of the TMR protocol (circular statistics) the R package Circular was used (Jammalamadaka, S. Rao and SenGupta 2001). Behavioral analysis was performed using robust statistical methods from the R package WRS2 (Mair and Wilcox 2020) to avoid any possible issues with normality and homoscedasticity assumptions. Repeated measures analysis of variance (RM-ANOVA) or simple 1-way ANOVA analysis was performed accordingly, always keeping the trial information and adding individual differences into the analysis (subjects ID’s). One sample t-test (Students or Wilcoxon signed-rank) tested for difference over chance level (50%) of each group, Condition and Session of interest. Significance of Pearson correlations between classification performance and behavior used a bootstrap method implemented in R, “boot” package (Buckland, Davison, and Hinkley 1998).

Classification

Classification of single-trial data was performed using MVPA-light (Treder 2020) for each participant and each time point (− 1 to 3s) using the sleep-ERP values (filtered between 4 to 20Hz) of the 60 EEG channels as feature. Performance of two classifiers was compared using a linear discriminant analysis (LDA) and a support vector machine with linear kernel (SVM). We used a 5-fold cross-validation method with 2 repetitions and principal component analysis (PCA) to reduce dimensionality (n = 20). The data within each fold was z-scored to avoid bias. Additionally, we used two different metrics to evaluate performance of each classifier: traditional accuracy (ACC), defined as % correct predictions, and area under the curve (AUC), or trade-off between the true positive and false positive rates. Once classifiers were calculated for each participant, we performed a between-subject cluster analysis (Treder 2020) to determine at what time points the Experimental and Control sounds were statistically different for each condition. All code for the analysis of this study is available at https://github.com/Contrerana.

Prior to sleep, participants performed a transitive inference task (figure 1A) with three hierarchies (figure 1B) of 6-items each (figure 1C). Each of the items was associated to a separate semantically related sound (e.g., traffic noise with a city landscape). Participants learned the sound-image associations until they were over 90% accurate in a retrieval task. They then learned the transitive inference task by repeatedly viewing adjacent (premise) pairs (e.g. A-B), for each hierarchy, and indicating which hid the smiley face, with feedback (figure 1E). After a 5-minute delay, the same task was performed without feedback to assess learning level (immediate test). Next morning, participants performed this test again, but inference pairs were now included, B-D, C-E and B-E, (delayed test, figure 1D). Sounds associated with two out of the three hierarchies were presented during SWS, with one hierarchy presented in Up and another in Down using CL-TMR. To control for other EEG underlying mechanisms induced by the cueing, control sounds were also played during the night for both states, Up and Down.

Behavioral results

Premise pairs

Training Performance: Immediate test performance on premise pairs was above chance (50%) in all three conditions (Down: 74.7±0.02%, Control: 79.2±0.02%, Up: 79.6±0.02%) but not at ceiling (see table 2-2). A 1-way ANOVA showed no significant differences between these conditions at baseline (F=2.69, p=0.07). Further post-hoc analysis (2-tailed t-test) showed no differences: Down vs Control (p=0.095), Down vs Up (p=0.095), Up vs Control (p=0.93). Hence, participants had equal premise pair knowledge for all conditions in the immediate test. Performance on the sound-image association was over 90% for all the stimuli (figure 2-1) and the number of times each stimulus was used for each condition remained relatively constant, table 2-1. Hence, we can rule out any bias due to stimuli or cue-memory associations.

To examine how in the premise pair performance evolved over time, and whether there were any impacts of TMR on this, we performed a repeated measures (RM)-ANOVA with Session (3 levels) and Condition (3 levels) as within subject factors, and with accuracy as the dependent variable (see Figure 2-2). We found main effects of Condition (F=3.5, p=0.029) and Session (F=32.39, p<0.0001), but no interaction (p=0.72). Further analysis for Condition revealed no significant differences between conditions in any of the three sessions (smallest p=0.103). On the other hand, there were clear Session effects: both between Session 1 (pre-sleep) and Session 3 (two-weeks later), and between Session 2 (next morning) and Session 3, figure 2-2 (all p<0.001). Thus, in keeping with normal declarative forgetting, there was a marked drop in premise pair performance over two weeks irrespective of condition, see table 2-2 for details. There was also no overnight improvement in premise pair accuracy, but this is in line with previous TI literature (Ellenbogen et al. 2007; Werchan and Gómez 2013). This result might be surprising given that associative memories are often strengthened by sleep (Rasch and Born 2013), but is also in line with the idea that sleep facilitates more weakly encoded memories (Schapiro et al. 2017), though this could also depend on other factors such as the type of task or the strength of the cue-memory associations (Denis et al. 2021).

Inference pairs

Inference pairs were introduced in the Late test performed during Session 2 and repeated in the 2-week follow up (Session 3). To examine the effect of TMR on inference pair performance both the next day and two weeks later, and how this differed for close (1st degree) and distant (2nd degree) inferences, we performed a RM-ANOVA with the factors Session (2 levels), Condition (3 levels) and Degree of separation (2 levels) (see Figure 1D). This revealed main effects of Degree (F=5.91, p=0.016) and Condition (F=13.97, p=0.002) but not Session (F=1.91, p=0.17). This ANOVA also revealed two interactions: between Degree and Condition (F=9.89, p=0.008), and between Session and Condition (F=7.86, p=0.021), but no interaction between Session and Degree (p=0.601). Finally, the interaction between all three factors was not significant (F=0.65, p=0.722). We examine the significant results in further detail below.

Interaction between Degree and Condition: To directly investigate the interaction between Degree and Condition, our post-hoc t-tests collapsed across Session. This showed that the Up condition differed from both Control (p=0.015) and Down (p=0.017) at the 2^nd Degree only. There was also a difference between 1st and 2nd degree in the Up condition (p=0.004), but there were no differences between the Control and Down conditions for either Degree of separation (lowest p=0.79). All probabilities were corrected for multiple comparisons, see figure 2, 2-3 and table 2-3. These results clearly demonstrate that Up-phase TMR can provide a benefit to transitive inference, but in keeping with (Ellenbogen et al. 2007), this is only significant for the most distant items.

To examine how the effects of Condition on 2^nd degree items change over time we performed separate ANOVAs for 2^nd degree items at Sessions 2 and 3. At Session 2 this showed better performance for Up vs. Control (p=0.027) and Down (p=0.006), but not between Down and Control (p=0.880), all corrected for multiple comparisons, figure 2C and table 2-3. At Session 3, two-weeks later, comparisons of Up vs. Control and Up vs. Down were no longer significant (p=0.087 and p=0.251 respectively). The benefit of the Up stimulation therefore fades over the two-week consolidation period (Figure 2D).

In summary, stimulation of the SO up state was associated with significantly better performance on the most distant (2nd Degree) inference pairs compared to Control stimulation in the next-day data, figure 2B lower left panel. Interestingly, stimulation of the Down phase appeared to impair inferences when compared to Up, irrespective of degree of separation (figure 2, left panel). However, this picture changed after two-weeks, when performance on inference pairs stimulated in the Down phase improved to above chance levels and did not differ from performance on pairs in the Control and Up conditions (figure 2A, right panel). Performance on Up and Control conditions did not change over this retention interval.

Interaction between Session and Condition

Post-hoc analysis of the interaction between Session and Condition revealed a difference between Up and Down conditions (p<0.001, corrected for multiple comparisons) at Session 2, supporting the idea that the Up and Down states are associated with distinct forms of neural processing (H.-V. V. Ngo and Staresina 2022). However, comparison of Up and Down with Control showed no differences after correction for multiple comparisons (p=0.46 and p=0.19 respectively). No differences were found for Session 3. This analysis also revealed an unexpected difference between Sessions for the Down condition (p=0.038), with a significant improvement of performance from under chance level (50%) at Session 2 (46.4±2.8%) to over chance level two-weeks later (58.6±2.9%). Control and Up conditions remained over chance level across both sessions, figure 3A and tables 3-7 for full report.

Global effect of TMR: Finally, because the 1st Degree items also showed a trend towards benefit from TMR, we collapsed across both Session and Degree for an exploratory analysis to test our a-priori hypothesis that Up state cueing would benefit all inferences. This showed better performance after Up than Control across all stimuli (p=0.03). Performance was also better for Up vs Down stimulation (p=0.002), but not for Control vs Down (p=0.687), all corrected for multiple comparisons.

EEG results

To distinguish the effect of TMR upon EEG responses from the effect of a playing a sound, we used both Experimental sounds, which were associated with the previously learned hierarchy and Control sounds which were associated with an unlearned hierarchy as indicated in figure 1F.

Efficiency of the closed-loop algorithm: To check that our online algorithm correctly differentiated between the Up and Down conditions, we calculated event-related potentials (ERPs) separately for each Condition (Up/Down) and Cue type (Experimental/Control sounds) (figure 3B). In addition to the expected ERP differences between Up and Down before cue onset, there was a third significant cluster for Control but not Experimental sounds. Furthermore, after the sound offset, both cues and type of sounds elicited a second SO cycle that made them statistically indistinguishable in line with previous literature (Göldi et al. 2019; H.-V. V. Ngo and Staresina 2022). To corroborate the accuracy of our closed-loop algorithm we calculated the phase of the cortical SO at stimulus onset for each trial and participant (figure 3A and figure 3-1). For Experimental sounds, the average values at channel F3 were 358.20° (standard deviation (SD):0.58) and 205.61° (SD:0.44) for the Up and Down conditions respectively. Similar values were obtained for the Control sounds: 358.79° (SD:0.59) and 208.37° (SD:0.50) respectively. Circular statistics corroborate a significant difference between Up and Down conditions for both Experimental and Control (p<0.001) but no differences between Experimental and Control in either Down or Up condition (p>0.1).

Time-frequency analysis at the stimulation electrode (F3) compared Experimental vs. Control sounds for frequencies between 4 and 20Hz for Up and Down cueing. For Up, there was a significant power decrease in both, SP (cluster p=0.008, 2.05s to 2.1s) and theta bands (cluster p=0.008, 1.6s to 2.3s) towards the end of the trial, figure 3C. For Down, the opposite was observed, with a power decrease apparent (p=0.012, t=1.7s to 1.8s and p=0.010, t=0.7s to 2.35s). There also was a significant interaction between cueing Condition and type of sound for both SP band (p=0.077, t=1.7s to 1.8s) and theta (p=0.007, t=1.6s to 2.3s). These temporal late differences are in line with previous work (Göldi et al. 2019; Cairney et al. 2018). We also compared Up and Down directly for the Experimental sounds to assess the beneficial effect of the Up condition. There is a positive cluster (Experimental > Control) around the sound onset (p=0.008) for SP band lasting until around 1.25s. This is similar to other CL-TMR literature (H.-V. V. Ngo and Staresina 2022). On the other hand, for theta band there is a decrease in power (Experimental < Control) with a significant cluster (p=0.037) ranging from 0.65s to 2.4s.

Detection of reactivation: To determine if memory-related neural activity was reactivated during the night, we trained two machine learning algorithms to differentiate between Experimental and Control sounds for each condition (Up/Down). After calculating the classifiably of data for each participant, we performed cluster statistics at the group level for each condition (Treder 2020). Only the Up condition presented a significant cluster of above-chance classification. This cluster (p=0.037) ranges from 1204 to 1298ms after stimulus onset for the SVM classifier with area under the curve (AUC) as performance metric (figure 4A). Similar values were obtained for the rest of the tested combinations, tables 4-1 to 44 & figure 4-1. There were no significant clusters for the Down condition (lowest p=0.074). The absence of significant classification here could indicate that the brain is not able to activate the corresponding memory trace above chance when the cues are presented in the down phase. Hence, no behavioral benefit is obtained the following morning. However, this does not explain the performance improvement observed two weeks later.

To better understand the relationship between reactivation and consolidation, we performed a series of correlations between classification performance within the significant cluster (mean and peak) and the behavioral metrics described above. This revealed a negative correlation (R=-0.65, p_corrected=0.022) between mean classification (SVM+AUC combination) and behavioral performance on 2^nd degree inference pairs in Session 2, figure 4B. This fits well with the fact that only 2^nd degree pairs in the Up

condition presented a significant difference in overnight improvement when compared with the control condition for Session 2, and similarly only Up showed between condition differences in the degree of separation of inference pairs. The correlation was constant across all tested machine learning algorithms, see table SM5.1. No other correlation with any behavioral metric was significant (all p>0.05 before correction). However, the fact that this correlation is negative, e.g., the better the algorithm can classify the worse participants performed on that particular pair is perhaps surprising. Similar negative correlations have been reported, e.g. between classification performance and behavioral metrics (Abdellahi et al. 2021). One possible hypothesis is “good” performers try to fit the control sound into the previously learned hierarchy. That is, when they hear the Control sound, they may reactivate the cued hierarchy again, making it more difficult for the classifiers to differentiate between control from learned sounds.

Transitive inference is a key cognitive ability and a hallmark of deductive reasoning (Vasconcelos 2008). By showing that TMR in SWS strengthens such inferential thinking, our data support the idea that memory reactivation is important for more than the mere strengthening of memories which has been investigated by the bulk of the TMR literature (Hu et al. 2020). It is also involved in more complex integration and restructuring processes (Lewis and Durrant 2011). Furthermore, our findings show that such integration can be intentionally boosted through an external intervention, an observation which may be important for the development of cognitive enhancers targeting this kind of thinking. The fact that our manipulation had a stronger effect on the more distant 2nd degree inferences is exciting in that it suggests the power of reactivation to assist in the integration of quite distant pieces of information.

Phase of the slow oscillations and memory

The concept of an optimal phase or window for TMR stimulation stems from the fact that depolarising Up-states are more likely to activate larger groups of neurons synchronously (Vyazovskiy et al. 2009) and to drive thalamocortical spindles and sharp-wave ripples in the hippocampus (Batterink, Creery, and Paller 2016), thus facilitating hippocampal memory reactivation. This could explain why our classifiers only detect memory reactivation after Up stimulation, and why we found next-morning behavioural benefit only after stimulation in this phase. Our findings are in keeping with recent work showing that Up phase stimulation elicits classifiable reactivation and boosts memory performance, while Down phase stimulation does not (H.-V. V. Ngo and Staresina 2022) as well as work showing that Up phase TMR more effectively promotes emotional updating than stimulation at other SO phases (Xia et al. 2023). Thus, our results join a growing literature showing that Up phase TMR has a stronger impact on neural processing.

TMR time course

The question of how long sleep related impacts on memory lasts is increasingly topical (Cordi and Rasch 2021; Pereira and Lewis 2020). One study reported that a TMR related reduction of implicit bias was retained after a week (Hu et al. 2020), while another study reported that TMR benefit to a serial reaction time task peaked after ten days (Rakowska et al. 2021). Here, accuracy for inference pairs was maintained two weeks after the manipulation for both Up and Control conditions (Fig. 2). However, performance in Down increased from below chance to above chance, reaching similar overall accuracy to the other two conditions after two-weeks (Fig. 2A-right). This improvement suggests that despite an initial inhibition due to TMR in the Down phase, the neural representation was able to recover over time. It is possible that subsequent nights of sleep without any manipulation may have allowed spontaneous reactivation of the memories, and that this occurred equally for all conditions. While Up and Control conditions, which had reactivated successfully in the first night, derived no benefit from this additional reactivation, the Down condition did benefit, and the associated consolidation allowed the Down-stimulated hierarchy to essentially catch up with the other hierarchies. Additionally, the next morning test in Session 2, which included presentation of inference pairs, may also have helped to trigger subsequent reactivation from which Down condition benefited.

Summary

Our results show that the complex process of making indirect inferences can be facilitated by cued reactivation during slow wave sleep. Importantly, this was only true for the most distant inference pairs, and only Up phase cueing was effective. Importantly, these effects were not transitory, but instead lasted for the full two-week period examined here. These results provide strong support for the idea that memory reactivation in sleep is important for high-level qualitative changes in memory, such as integration and relational memory. Our findings also hold promise for the use of sleep-based interventions to drive improvement in such complex memory, and its application to real-word tasks.

Acknowledgments

We would like to thank Elena Schmidt, Ralph Andrews and Duarte Pereira for contributing to the data collection, Dominic Carr for helping with the sleep scoring and the NAPs lab in general. This work was funded by the ERC Consolidator grant SolutionSleep, 681607 to PL.

Conflict of interest

“The authors declare no competing financial interests.”

Abdellahi, M.E.A., Anne, C.M., Koopman, M.S., Treder, Lewis, P.A.: “Targeting Targeted Memory Reactivation: Characteristics 2 of Cued Reactivation in Sleep.” (2021). https://doi.org/https://doi.org/10.1101/2021.12.09.471945
Åkerstedt, T., and Mats Gillberg: Subjective and Objective Sleepiness in the Active Individual. Int. J. Neurosci. 52(1–2), 29–37 (1990). https://doi.org/10.3109/00207459008994241
Batterink, L.J., Jessica, D., Creery, Paller, K.A.: Phase of Spontaneous Slow Oscillations during Sleep Influences Memory-Related Processing of Auditory Cues. J. Neurosci. 36(4), 1401–1409 (2016). https://doi.org/10.1523/JNEUROSCI.3175-15.2016
Berry, R.B., Brooks, R., Gamaldo, C.E., Harding, S.M., Lloyds, R.M., Marcus, C.L., Vaughn, B.V.: The AASM Manual for the Scoring of Sleep and Associated Events. Version 2. Darien: American Academy of Sleep Medicine. aasm.org. (2015)
Born, J., Björn, Rasch, Gais, S.: “Sleep to Remember.” Neuroscientist. 12 (1073–8584 (Print)): 410–24. (2006)
Brainard, D.H.: The Psychophysics Toolbox. Spat. Vis. 10(4), 433–436 (1997). https://doi.org/10.1163/156856897X00357
Bryant, P.E., Trabasso, T.: Transitive Inferences and Memory in Young Children. Nature. 232(5311), 456–458 (1971). https://doi.org/10.1038/232456a0
Buckland, S.T., Davison, A.C., Hinkley, D.V.: Bootstrap Methods and Their Application. Biometrics. 54(2), 795 (1998). https://doi.org/10.2307/3109789
Cai, D.J., Sarnoff, A., Mednick, E.M., Harrison, J.C., Kanady, Sara, C.M.: REM, Not Incubation, Improves Creativity by Priming Associative Networks. Proc. Natl. Acad. Sci. U.S.A. 106(25), 10130–10134 (2009)
Cairney, S.A., Anna á.Váli Váli Guttesen, Nicole El Marj, and, Staresina, B.P.: “Memory Consolidation Is Linked to Spindle-Mediated Information Processing during Sleep.” Current Biology, March. (2018). https://doi.org/10.1016/j.cub.2018.01.087
Camarena, H.O., García-Leal, O., Burgos, J.E., Parrado, F., Laurent, Ã.-C.: Transitive Inference Remains Despite Overtraining on Premise Pair C + D-. Front. Psychol. 9 (2018). https://doi.org/10.3389/fpsyg.2018.01791
Cordi, M.J., Björn, Rasch: How Robust Are Sleep-Mediated Memory Benefits? Curr. Opin. Neurobiol. 67(April), 1–7 (2021). https://doi.org/10.1016/j.conb.2020.06.002
Denis, D., Mylonas, D., Poskanzer, C., Bursal, V., Payne, J.D., and Robert Stickgold: Sleep Spindles Preferentially Consolidate Weakly Encoded Memories. J. Neurosci. (2021). https://doi.org/10.1523/JNEUROSCI.0818-20.2021
Durrant, S.J., Scott, A., Cairney, Lewis, P.A.: Overnight Consolidation Aids the Transfer of Statistical Knowledge from the Medial Temporal Lobe to the Striatum. Cereb. Cortex. 23(10), 2467–2478 (2013). https://doi.org/10.1093/cercor/bhs244
Dymond, S., and Samuel Llewellyn: Time, Sleep, and Stimulus Equivalence-Based Relational Memory. Psychol. Record. 69(3), 357–365 (2019). https://doi.org/10.1007/s40732-019-00343-8
Ellenbogen, J.M., Peter, T., Hu, Jessica, D., Payne, D., Titone, and Matthew P. Walker: Human Relational Memory Requires Time and Sleep. Proc. Natl. Acad. Sci. U.S.A. 104(18), 7723–7728 (2007). https://doi.org/10.1073/pnas.0700094104
Göldi, M., van Poppel, E.A.M., Rasch, B., and Thomas Schreiner: Increased Neuronal Signatures of Targeted Memory Reactivation during Slow-Wave up States. Sci. Rep. 9(1), 2715 (2019). https://doi.org/10.1038/s41598-019-39178-2
Groch, S., Preiss, A., McMakin, D.L., Björn Rasch, S., Walitza, R., Huber, and Ines Wilhelm: Targeted Reactivation during Sleep Differentially Affects Negative Memories in Socially Anxious and Healthy Children and Adolescents. J. Neurosci. (2017). https://doi.org/10.1523/JNEUROSCI.1912-16.2017
Grosenick, L., Clement, T.S., and Russell D. Fernald: Fish Can Infer Social Rank by Observation Alone. Nature. 445(7126), 429–432 (2007). https://doi.org/10.1038/nature05511
von Hecker, U., Klauer, K.C., Aßfalg, A.: A Robust Anchoring Effect in Linear Ordering. Q. J. Experimental Psychol. 72(11), 2680–2689 (2019). https://doi.org/10.1177/1747021819855234
Hennies, N., Ralph, M.A.L., Kempkes, M., Cousins, J.N., Lewis, P.A.: Sleep Spindle Density Predicts the Effect of Prior Knowledge on Memory Consolidation. J. Neurosci. (2016). https://doi.org/10.1523/JNEUROSCI.3162-15.2016
Hoddes, E., Dement, W.: and Vincent Zarcone. “The Development and Use of the Stanford Sleepiness Scale (SSS).” Psychophysiology (1972)
Holyoak, K.J., and Hongjing Lu: Emergence of Relational Reasoning. Curr. Opin. Behav. Sci. 37(February), 118–124 (2021). https://doi.org/10.1016/j.cobeha.2020.11.012
Horst, J.S., Hout, M.C.: The Novel Object and Unusual Name (NOUN) Database: A Collection of Novel Images for Use in Experimental Research. Behav. Res. Methods. 48(4), 1393–1409 (2016). https://doi.org/10.3758/s13428-015-0647-3
Hu, X., Cheng, L.Y., Chiu, M.H., Paller, K.A.: Promoting Memory Consolidation during Sleep: A Meta-Analysis of Targeted Memory Reactivation. Psychol. Bull. 146(3), 218–244 (2020). https://doi.org/10.1037/bul0000223
Jammalamadaka, S., Rao, SenGupta, A.: Topics in Circular Statistics. World Scientific (2001)
Lazareva, O.F., Gazes, R.P., Elkins, Z., and Robert Hampton: Associative Models Fail to Characterize Transitive Inference Performance in Rhesus Monkeys (Macaca Mulatta). Learn. Behav. 48(1), 135–148 (2020). https://doi.org/10.3758/s13420-020-00417-6
Lerner, I., Gluck, M.A.: Sleep and the Extraction of Hidden Regularities: A Systematic Review and the Importance of Temporal Rules. Sleep Med. Rev. 47(October), 39–50 (2019). https://doi.org/10.1016/j.smrv.2019.05.004
Lewis, P., Simon, J.D.: Overlapping Memory Replay during Sleep Builds Cognitive Schemata. Trends Cogn. Sci. 15(8), 343–351 (2011). https://doi.org/10.1016/j.tics.2011.06.004
Lundqvist, D., Flykt, A., and Arne Ohman: The Karolinska Directed Emotional Faces (KDEF). Stockholm: Department of Neurosciences Karolinska Hospital. 91–630 (1998). https://doi.org/10.1017/S0048577299971664
Lutz, N.D., Diekelmann, S., Patricia Hinse-stern, Jan Born, and, Rauss, K.: “Sleep Supports the Slow Abstraction of Gist from Visual Perceptual Memories.” Nature Publishing Group, no. January: 1–9. (2017). https://doi.org/10.1038/srep42950
Mair, P., and Rand Wilcox: Robust Statistical Methods in R Using the WRS2 Package. Behav. Res. Methods. 52(2), 464–488 (2020). https://doi.org/10.3758/s13428-019-01246-w
Massimini, M., Ferrarelli, F., Esser, S.K., Riedner, B.A., Huber, R., Murphy, M., Peterson, M.J., and Giulio Tononi: Triggering Sleep Slow Waves by Transcranial Magnetic Stimulation. Proc. Natl. Acad. Sci. U.S.A. (2007). https://doi.org/10.1073/pnas.0702495104
Morgan, A.: Effect of Time and Sleep on the Transitive Inference Task. ” Harvard Extension School (2017)
Ngo, H.-V.V.: and Bernhard P. Staresina. “Shaping Overnight Consolidation via Slow-Oscillation Closed-Loop Targeted Memory Reactivation.” Proceedings of the National Academy of Sciences 119 (44): e2123428119. (2022). https://doi.org/10.1073/pnas.2123428119
Ngo, H.-V.V.V., Martinetz, T., Born, J., Mölle, M.: Auditory Closed-Loop Stimulation of the Sleep Slow Oscillation Enhances Memory. Neuron. 78(3), 545–553 (2013). https://doi.org/10.1016/j.neuron.2013.03.006
Ngo, H.V.V.H.-V.V., Miedema, A., Faude, I., Martinetz, T., Mölle, M., Born, J., Molle, M., and Jan Born: Driving Sleep Slow Oscillations by Auditory Closed-Loop Stimulation—A Self-Limiting Process. J. Neurosci. 35(17), 6630–6638 (2015). https://doi.org/10.1523/JNEUROSCI.3133-14.2015
Nolan, B.: “REM, Not Incubation, Improves Creativity by Priming Associative Networks.” Yearbook of Neurology and Neurosurgery 2010 (January): 191–92. (2010). https://doi.org/10.1016/S0513-5117(10)79183-3
Oostenveld, R., Fries, P., Eric Maris, and Jan-Mathijs Mathijs Schoffelen: FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data. Comput. Intell. Neurosci. 2011, 1–9 (2011). https://doi.org/10.1155/2011/156869
Oudiette, D., and Ken a. Paller: Upgrading the Sleeping Brain with Targeted Memory Reactivation. Trends Cogn. Sci. 17(3), 142–149 (2013). https://doi.org/10.1016/j.tics.2013.01.006
Pereira, S.I., Ribeiro, Lewis, P.A.: The Differing Roles of NREM and REM Sleep in the Slow Enhancement of Skills and Schemas. In: Current Opinion in Physiology. Elsevier Ltd (2020). https://doi.org/10.1016/j.cophys.2019.12.005
Rakowska, M., Mahmoud, E.A., Abdellahi, PAulina Bagrowska, M., Navarrete, Lewis, P.A.: Long Term Effects of Cueing Procedural Memory Reactivation during NREM Sleep. NeuroImage. 244(December), 118573 (2021). https://doi.org/10.1016/j.neuroimage.2021.118573
Rasch, B., and Jan Born: About Sleep’s Role in Memory. Physiol. Rev. 93(2), 681–766 (2013). https://doi.org/10.1152/physrev.00032.2012
Schapiro, A.C., Elizabeth, A., McDevitt, L., Chen, K.A., Norman, S.C., Mednick, Timothy, T., Rogers: Sleep Benefits Memory for Semantic Category Structure While Preserving Exemplar-Specific Information. Sci. Rep. (2017). https://doi.org/10.1038/s41598-017-12884-5
Schouten, D.I., Sofia, I.R., Pereira, M., Tops, and Fernando M. Louzada: State of the Art on Targeted Memory Reactivation: Sleep Your Way to Enhanced Cognition. Sleep Med. Rev. (2017). https://doi.org/10.1016/j.smrv.2016.04.002
Steriade, M.: Grouping of Brain Rhythms in Corticothalamic Systems. Neuroscience. 137(4), 1087–1106 (2006). https://doi.org/10.1016/j.neuroscience.2005.10.029
Steriade, M., A Nuñez, and, Amzica, F.: A Novel Slow (< 1 Hz) Oscillation of Neocortical Neurons in Vivo: Depolarizing and Hyperpolarizing Components. J. Neuroscience: Official J. Soc. Neurosci. 13(8), 3252–3265 (1993). http://www.ncbi.nlm.nih.gov/pubmed/8340806
Tamminen, J., Payne, J.D., Stickgold, R., Wamsley, E.J., Gareth Gaskell, M.: Sleep Spindle Activity Is Associated with the Integration of New Memories and Existing Knowledge. J. Neurosci. (2010). https://doi.org/10.1523/JNEUROSCI.3028-10.2010
Treder, M.S.: MVPA-Light: A Classification and Regression Toolbox for Multi-Dimensional Data. Front. NeuroSci. 14 (2020). https://doi.org/10.3389/fnins.2020.00289
Vasconcelos, M.: “Transitive Inference in Non-Human Animals: An Empirical and Theoretical Analysis.” Behavioural Processes 78 (3): 313–34. (2008). https://doi.org/10.1016/j.beproc.2008.02.017
Vyazovskiy, V.V., Olcese, U., Lazimy, Y.M., Faraguna, U., Esser, S.K., Williams, J.C.: Chiara Cirelli, and Giulio Tononi. 2009. “Cortical Firing and Sleep Homeostasis.” Neuron 63 (6): 865–78. https://doi.org/10.1016/j.neuron.2009.08.024
Walker, M.P., Conor Liston, J.A., Hobson, Stickgold, R.: Cognitive Flexibility across the Sleep–Wake Cycle: REM-Sleep Enhancement of Anagram Problem Solving. Cogn. Brain. Res. 14(3), 317–324 (2002). https://doi.org/10.1016/S0926-6410(02)00134-9
Werchan, D.M., Gómez, R.L.: Generalizing Memories over Time: Sleep and Reinforcement Facilitate Transitive Inference. Neurobiol. Learn. Mem. 100(February), 70–76 (2013). https://doi.org/10.1016/j.nlm.2012.12.006
Xia, T., Yao, Z., Guo, X., Liu, J., Chen, D., Liu, Q., Paller, K.A., and Xiaoqing Hu: Updating Memories of Unwanted Emotions during Human Sleep. Curr. Biol. 33(2), 309–320e5 (2023). https://doi.org/10.1016/j.cub.2022.12.004

There is NO Competing Interest.

nrreportingsummary.pdf
Reporting Summary
Supplementarymaterials.docx

Download PDF

Journal Publication

published 08 Mar, 2024

Read the published version in Communications Biology →

Version 1

posted

You are reading this latest preprint version

Memory reactivation in slow wave sleep enhances relational learning.

Status:

Journal Publication

Version 1

Abstract

Figures

Significance Statement

Introduction

Materials and Methods

Results

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1