Predicting brainwaves: The influence of auditory closed-loop
cueing during slow oscillation up-states on vocabulary memory
Master thesis Eva Anna Maria van Poppel
Supervisor: Lucia M. Talamini, PhD Second assessor: Thomas Meindertsma, MSc Programme Brain & Cognition, Specialisation Cognitive Neuroscience, Department of Psychology, University of Amsterdam
Student number: 6030262 Date: 29-8-2016
Index
Abstract ... 3
Keywords & Abbreviations ... 3
Introduction ... 4 Methods ... 7 1. Ethics statement ... 7 2. Subjects ... 7 3. Stimuli ... 8 4. Procedure ... 8
5. Vocabulary learning task... 9
6. Data acquisition ... 10
7. Brainwave predicting algorithm ... 11
7.1 Phase prediction criteria ... 11
8. Reactivation of Vocabulary ... 12
9. Vocabulary memory recall task ... 13
Data analysis ... 13
10. Offline analysis of algorithm performance ... 13
11. Pre-processing EEG data ... 14
12. Sleep scoring ... 14
13. Analysis of Power Changes ... 14
14. Statistics ... 14
Results ... 15
1. Phase-predicting algorithm parameter study ... 15
2. Speed ... 17
3. Offline algorithm performance ... 18
4. Online algorithm performance ... 19
5. Behavioural results ... 20
6. Event-Related Potentials (ERPs) ... 23
7. Topoplot ... 25
8. Time-Frequency Analysis (TFA) ... 25
9. Time spent per sleep stage and number of reactivations ... 27
10. Amount of cues compared to memory performance correlations ... 28
Discussion ... 30
Acknowledgments ... 34
References ... 34
Supplementary Figure 1. ... 38
Abstract 3
Predicting brainwaves: The influence of auditory closed-loop
cueing during slow oscillation up-states on vocabulary memory
Eva Anna Maria van Poppel
Abstract
Objectives: Slow Wave Sleep (SWS), also known as deep sleep, is important for the consolidation of memory traces. Slow oscillations (SOs) are the hallmark of SWS and are characterized by an up-state where neurons synchronically depolarise and have an increased excitability, followed by the down-state where neurons are hyperpolarised and silenced. The up-state seems to play an important role in memory consolidation. In this study, we tried to reactivate and enhance vocabulary memory traces by presenting auditory cues during the up-state of an ongoing slow oscillation.
Methods: We developed a brainwave predicting algorithm which can predict the next upcoming up or down-state in the ongoing brain signal. For the vocabulary memory task, participants learned words in a foreign language and then slept a whole night. During the night, half of the foreign words where cued auditory in a predicted SO up-state. In the morning, vocabulary memory was compared between cued and uncued words.
Results: Results show that the new oscillatory phase targeting algorithm is faster and more accurate than any previously reported methods and performs well in both offline and real-time recordings. An enhancement of the cueing for the vocabulary memory is found for subjects reaching a sufficient pre-sleep level of encoding. When up-state cued items are compared to down-state cued items, up-state cueing seems to enhance vocabulary memory whereas down-state cueing seems to diminish vocabulary memory. High-density EEG recordings revealed that cueing during a slow oscillation interrupts the ongoing SO pattern and induces a strong negativity in ERPs. At the neural level, we also found an increase of fast spindle (12–15 Hz), low beta (15-20 Hz) and theta band activity (5–10 Hz) associated with the cues. This might reflect the processing of language and recollection.
Conclusion: The newly developed oscillatory phase targeting method shows superior performance compared to previous validated methods and has a broad applicability. Targeted memory reactivation (TMR) enhances vocabulary memory for items cued in a slow oscillation up-state when a pre-sleep level of encoding is reached. This is accompanied by several neuronal processes reflecting the reactivation and possible strengthening of memory traces. When TMR in a slow oscillation down-state is applied, the reactivations seem to deteriorate the vocabulary memory. Further research on this matter is needed, since it can be of importance in the treatment of PTSD patients. Keywords & Abbreviations
Slow Wave Sleep (SWS), Slow Oscillation (SO), up-state, down-state,
phase prediction, Targeted Memory Reactivation (TMR), Electro-encephalogram (EEG), Event-Related Potential (ERP), Time-frequency analysis (TFA),
Introduction
According to the active system consolidation hypothesis, memory traces reactivate spontaneously during slow oscillations (SOs), which are the hallmark of slow wave sleep (SWS) (Diekelmann & Born, 2010; O’Neill, Pleydell-Bouverie, Dupret & Csicsvari, 2010; Stickgold & Walker, 2013). During targeted memory reactivation (TMR), a specific memory trace is reactivated deliberately by presenting learning-related cues during sleep. This TMR selectively enhances memory
consolidation (Oudiette & Paller, 2013; Batterink, Creery & Paller, 2016). It is possible to reactivate and enhance declarative memory traces during SWS by presenting a previously learned contextual detail, like an odor (Rasch, Büchel, Gais & Born, 2007; Rihm, Diekelmann, Born & Rasch, 2014) or sound (Rudoy, Voss, Westerberg & Paller, 2009; van Dongen et al., 2012). This presenting of
contextual details during non-REM sleep is associated with the occurrence of sleep spindles (Cox, Hofman, de Boer & Talamini, 2014a). Sleep spindles seem to play a functional role in both
declarative memory consolidation (Schabus et al., 2004) and motor memory consolidation (Lustenberger et al., 2016).
Besides cueing with a previously learned associated contextual detail, it is also possible to cue with a part of the learned stimulus itself, for example when participants learned to play a melody. The TMR with the learned melody improved the procedural memory (Anthony, Gobel, O’Hare, Reber & Paller, 2012; Schönauer, Geisler & Gais, 2013). Targeting memory traces with the learned stimulus itself is also applicable to declarative memory traces. In a vocabulary learning task, participants were cued with the learned foreign words during non-REM sleep. This enhanced the memory for the cued words compared to the uncued words (Schreiner & Rasch, 2014), while cueing previously learned words when people are awake has no effect on memory performance (Schreiner & Rasch, 2015). When the correct translation of the cued word is presented immediately after the foreign word during non-REM sleep, this seems to block the beneficial effect of TMR with single cues (Schreiner, Lehman & Rasch 2015a). The processing of words seems to elicit more activity in the theta band in the brain (Röhm, Klimesch, Haider & Doppelmayr, 2001; Bastiaansen, Van der Linden, Ter Keurs, Dijkstra & Hagoort, 2005). The cued words elicit more theta activity also during later recall compared to the uncued words (Schreiner, Göldi & Rasch, 2015b).
Slow oscillations (SOs) seem to play a functional role in memory consolidation (Marshall, Helgadottir, Mölle & Born 2006; Diekelmann & Born, 2010; Chauvette, Seigneur & Timofeev, 2012). A slow oscillation is characterized in the electro-encephalogram (EEG) as a wave with amplitude >75 µV and frequency around 1 Hz (Steriade, 2006). A SO is thought to reflect global
synchronous neuronal activity in both the neocortex and the hippocampus (Massimini, Huber, Ferrarelli, Hill & Tononi, 2004; Isomura et al., 2006; Steriade, 2006), with neuronal depolarisation and
Introduction 5 increased excitability in the up-state and widespread neuronal hyperpolarisation and strongly depressed firing in the down-state (Sejnowski & Destexhe, 2000; Steriade, 2003; Steriade & Timofeev, 2003; Buzsáki & Draguhn, 2004; Olcese, Esser & Tononi 2010; Cox, van Driel, de Boer & Talamini, 2014b). The first 180° of the SO are called the up-state, with the peak at 90°. The last 180° to 360° are called the down-state, with its trough at 270° (see figure 1 for an example).
For the memory consolidation function of SOs during sleep, the synchronisation of fast-spindle activity (12–15 Hz) to the depolarising up-state seems to be critical (Mölle & Born, 2011; Mölle, Bergmann, Marshall & Born, 2011; Cox, Hofman & Talamini, 2012).
Figure 1. Example of a slow oscillation with a 1 Hz frequency. The first 180° of the wave are
called the up-state, with the peak at 90°. The last 180° to 360° are called the down-state, with its trough at 270°.
It is possible to enhance declarative memory when you boost the amount and amplitude of slow oscillations, either by applying a transcranial slow oscillation potential (Marshall et al., 2006; Massimini et al., 2007) or by cueing with sounds in the up-state of an ongoing slow oscillation (Tononi, Riedner, Hulse, Ferrarelli & Sarasso 2010; Ngo, Martinetz, Born & Mölle, 2013; Ngo, Miedema, Faude, Martinetz & Mölle, 2015).
To apply cues directly into an ongoing SO up-state, closed-loop stimulation can be used. The first known study which used closed-loop stimulation to target SO up-states in the human sleep EEG, is Ngo et al. (2013). Their approach was to analyse the incoming signal from a selected frontal electrode (Fpz) and filter in the range of interest (0.25 – 4 Hz). Every time this signal passes a predefined threshold (-80 µV), an auditory stimulation was triggered. However, a fixed time interval was used for the stimulation (Ngo et al., 2013). Since brainwaves are adaptive, you need to predict at which time point the up-state will occur, to be able to present a stimulus
exactly at the peak of the up-state. Therefore, Cox, Korjoukov, de Boer & Talamini (2014d) created a brainwave-predicting algorithm. This algorithm performs a Fast Fourier transform (FFT) over the last 10 seconds of ongoing signal from the selected electrode (Fpz). When the FFT bin in the slow oscillation range of interest (0,6 – 1,2 Hz) crosses a predefined threshold, a sine wave in the most dominant frequency is fitted over the signal and, given a sufficient fit, extrapolated into the future to reveal the next up- and down-state. This enables presenting a stimulus right at the beginning of an up or down-state.
Another approach to predict the next up-state is using a phase-locked loop (PLL). This generates an oscillation in the same frequency and phase of the input signal, enabling stimulating in the target phase (Santostasi et al., 2016). However, these previous described methods are not as fast and accurate as desirable. Since slow oscillation phases are associated with different
physiological states and neural activity, stimulation at different phases can induce different effects. To enhance memory traces during sleep, it is important to have a reliable phase targeting method. Therefore, we developed a new non-linear up- and down-state predicting algorithm that is more accurate, faster and more convenient to implement than any of these described methods.
In this study, we applied TMR directly in a predicted SO up-state. We examined the possibility of boosting vocabulary memory by cueing previously learned words directly in an up-state of an ongoing slow oscillation. We expect that cueing directly in an up-up-state will enhance vocabulary memory (Ngo et al., 2013; 2015), since the up-state is associated with neuronal
depolarisation and increased neuronal excitability (Steriade, 2003; Massimini et al., 2004; Isomura et al., 2006; Olcese et al., 2010; Timofeev, 2011; Cox et al., 2014b).
To compare the vocabulary memory and induced neuronal activity by up-state TMR, we also applied TMR in another group of people who received cues in the down-state only. For the down-state, the TMR effects are yet unknown. However, since the down-state is associated with global synchronous neural hyperpolarisation and quiescence in both the neocortex and
hippocampus (Sejnowski & Destexhe, 2000; Steriade, 2003; Steriade & Timofeev, 2003; Buzsáki & Draguhn, 2004), we expect less memory enhancement for stimuli that are cued in the down-state compared to stimuli that are cued in the up-state of an ongoing slow oscillation.
In order to enable reliable cueing directly in an up-state peak or down-state trough in an ongoing slow oscillation, we developed a new phase prediction algorithm. We then used this algorithm to present previously learned foreign words directly to the slow oscillation up or down-state during non-REM sleep. We evaluated brain-wide responses to stimuli in both up- and down-states in the event-related potential (ERP) and time-frequency domain (TFA). Finally, we
Methods 7 tested whether the cueing in an exclusive up or down-state enhances declarative memory
compared to uncued items.
Methods
1. Ethics statement
Procedures were approved by the ethics committee of the University of Amsterdam, Psychology department. All participants were provided with written informed consent and a debriefing.
2. Subjects
A total of 40 young, healthy university students (age range 18-25; 25 female) with Dutch as first language and without any Danish language skills participated in the study. Subjects reported that they never had a neurological, psychiatric or sleeping disorder, didn’t use alcohol in the 24 hours prior to the experiment, didn’t use any narcotic or stimulating drugs or medicines in the three days prior to the experiment and limited their caffeine intake at the day of the experiment to a maximum of three cups which weren’t consumed after 6 PM. Subjects were required to wear an actiwatch on the day and night prior to the experiment, to wake up at 8 AM the latest at the day of the experiment and weren’t allowed to take naps at the day of the experiment. Subjects who didn’t comply with any or more of these conditions were excluded from all analyses. Subjects were rewarded with credits to fulfil course requirements or with a monetary reward.
A total of seven subjects had to be excluded, due to technical problems (N=4) or because they couldn’t sleep (N=3). The remaining 33 subjects were randomly subdivided in three
experimental groups:
(Group 1) Who received only up cues during the night, (N=24, 13 female, mean age = 20.9 ± 1.2 years),
(Group 2) Who received only down cues during the night, (N=3, 2 female, mean age = 22.0 ± 1.0 years) and
(Group 3) Who didn’t receive any cues during the night and served as a control group, (N=6, 6 female, mean age = 20.8 ± 2.2 years).
3. Stimuli
Stimuli consisted of in total 150 non-emotional Danish nouns. The Danish nouns were pronounced by a native male Danish speaker with a neutral voice and recorded with a stereo microphone. To replicate the vocabulary study of Schreiner & Rasch (2014) as much as possible, the Danish language was chosen since the language distance of Danish and Dutch is comparable to German and Dutch (Petroni & Serva, 2008; Isphording & Otten, 2011; Isphording, 2013) (word pairs are listed in the Supplementary Table 1).
The length of all stimuli was exactly 500 ms, to fit in either a slow oscillation up or down-state. This was achieved by speeding the words up without changing the voice pitch using the Audacity software (http://audacityteam.org). Stimuli were saved as a stereo ‘wave’ sound file (signed 16-bit PCM format) with a sound quality of 44100 Hz.
4. Procedure
Subjects arrived at the sleep laboratory at 8 PM. After filling out written informed consent forms, subjects were prepared for polysomnographic registration (see Data acquisition). In all
experimental groups, the learning phase started at ∼22 PM (see figure 2 for an overview of the procedure). Subjects were asked to fill out the Stanford Sleepiness Scale (SSS) and then started with the vocabulary learning task (120 Dutch–Danish word pairs, for a detailed description see Vocabulary learning task). At ~23 PM, lights were turned off and subjects were given the opportunity to sleep a full night. When the first SWS period was detected (~0 AM), the
brainwave-predicting algorithm was turned on and cued a selection of the prior learned Danish words auditory for three hours. Cueing took place in the up-state (group 1) or down-state (group 2) of an ongoing slow oscillation. In group 3, up-state predictions were made, without playing any sound, to serve as a control group (see Brainwave-predicting algorithm for a detailed description of the algorithm, see Reactivation of vocabulary for a detailed description of the word selection). Sleep was continuously monitored by the experimenter and the stimulation was interrupted whenever polysomnographic signs of REM sleep, arousal or awakenings occurred. Three hours after the first cue, the algorithm was turned off and subjects were able to sleep the rest of the night without cueing. The following morning at 8 AM they were woken up from light or REM sleep stages and got a light breakfast to recover from sleep inertia. At ~8.30 AM, subjects were asked to fill out a questionnaire regarding general sleep characteristics and the SSS. Then recall of the vocabulary was tested (see Vocabulary memory recall task).
Methods 9
Group 1 Learning Sleep (cueing in up-state) Retrieval
Group 2 Learning Sleep (cueing in down-state) Retrieval
Group 3 Learning Sleep (no cueing) Retrieval
22.00 23.00 ~0.00 ~3.00 8.00 8.30 h
Figure 2. Experimental procedure. Learning phase started at 22 PM in all groups. At 23 PM,
lights were turned off. When the first SWS cycle appeared, cueing took place in the up-state (group 1) or down-state (group 2) of an ongoing slow oscillation. In group 3 up-state
predictions were made, without playing any sound. At 8 AM subjects were awakened and recall was tested at 8.30 AM in all experimental groups.
5. Vocabulary learning task
The vocabulary learning task consisted of 120 Danish words and their Dutch translation (word pairs are listed in the Supplementary Table 1). Danish words were presented aurally via
loudspeakers (70 dB sound pressure level). Fixation cross and the Dutch translation were
presented in the middle of a 22” LCD screen with 120 Hz refresh rate and with Verdana font and 45 font size. After 40 trials, there was a one minute break.
In the first learning round, the translation of 40 Danish words was learned in a passive way. A fixation cross (500 ms) was followed by the Dutch translation (2000 ms) and subsequently by the Danish word (500 ms). The intertrial interval was 1000 ms.
In the second learning round, the translation of the same 40 Danish words in a randomised order was learned in an active way. The fixation cross (500 ms) was followed by the Danish word (500 ms) and a 1000 ms interval. Thereafter, the subject had 20 s to type the Dutch translation. This was always followed by the presentation of the correct Dutch translation (2000 ms). Active and passive learning rounds alternated until the translations of all 120 Danish words were learned.
In the last learning round, the recall of all learned 120 words was tested without any feedback on the correct Dutch translation. The procedure was identical to the second learning round, except that the presentation of the correct Dutch translation was replaced by a confidence interval, where people had to indicate their reliance about the given translation on a scale of 0-4. 0 indicated “I took a guess”, 1 = “I am very unsure”, 2 = “unsure”, 3 = “sure” and 4 = “very sure”. Recall performance on this round was taken as a pre-sleep learning performance baseline (see figure 3 for a schematic overview of the vocabulary learning task).
Figure 3. 40 words were learned in each passive and active learning round, until all 120
word pairs were learned. Recall was tested in the last round and served as pre-sleep performance.
5.1 Pilot data of the vocabulary learning task
To make sure the vocabulary learning task had an ideal medium task difficulty without any danger of ceiling or floor effects we piloted this task (N=24, 14 female, mean age 20.0 ± 2.0 years, all university students). Pilot data showed that on average 60.29 ± 12.3 out of 120 word pairs were recalled correctly in the third recall round (recall performance 50.2%, range 40 – 80).
To test the difficulty of words, subjects without any knowledge of the Danish language were asked to guess the Dutch translation of the aurally presented Danish words (N=10, 8 female, mean age 19.5 ± 1.5 years, all university students). The results of this test are listed in supplementary table 1.
6. Data acquisition
EEG data was acquired using a 64-channel WaveGuard™ original cap (ANT, Enschede, The Netherlands) and two additional fixed mastoid electrodes. Horizontal and vertical
electrooculography (EOG) and chin electromyography (EMG) were recorded with bipolar electrodes. All impedance levels were kept below 10 kΩ. Signals were sampled at 512 Hz using a 72-channel Refa DC amplifier (TMSi, Oldenzaal, The Netherlands) and stored on a separate recording computer using a procedure developed in-home based on Polybench recording software (TMSi, Oldenzaal, The Netherlands). This developed software is able to split and send the signal from the dedicated electrode (Fpz) and record all channels simultaneously. The signal of electrode Fpz was re-referenced to the average of mastoids (Fpz-𝑋̅M) and sent over a Local Area Network (LAN) with IPv4/TCP protocol to the dedicated algorithm computer, using data blocks of 5 ms. We used a cross-over network cable (category 6) with a length of one meter to create the LAN. Passive 1 till 40 Active 1 till 40 Passive 41 till 80 Active 41 till 80 Passive 81 till 120 Active 81 till 120 Test 120 words
Methods 11
7. Brainwave predicting algorithm
The algorithm predicting up- and down-states in an ongoing EEG signal ran in the EventIDE software (http://okazolab.com) on a dedicated computer. It received its real-time input over the LAN from the IPv4/TCP protocol. The algorithm used the most recent second of data (512 samples) and fitted a sine in the frequency of interest (0.6 – 1.2 Hz) using the nonlinear solver function. When all criteria were met (see 7.1 Algorithm parameters), the algorithm was allowed to make a prediction and cue a sound. The algorithm was then paused for 3500 ms.
7.1 Phase prediction criteria
The algorithm is only allowed to make a prediction and cue a sound when all criteria are met. The following criteria can be taken into account:
Phase fitting threshold defines the minimal fit accuracy for the phase fitting error. We set this threshold to 0.3 to find the perfect balance between the amount of predictions per minute and an acceptable phase error.
Crunch time defines the period of future time in which the prediction of the target phase is allowed. Since brainwaves naturally vary over time, the maximum crunch time should not be too high. The minimum crunch time should be the total loop delay (hardware lag + analysis time) at least. We set the crunch time to 24 – 34 ms.
Amplitude threshold defines the minimum amplitude (in µV) the signal should cross. We set this threshold to 75 µV for offline tests, but when we found out it didn’t contribute significantly to the phase error we turned it off during the reactivation study.
Central frequency power threshold. This function uses the last 10 seconds of data samples, applies a Butterworth band-pass (filter order 1) with central frequency 0.9 Hz and a band-pass width of 1 Hz. It then applies a power normalisation over the range of interest (0.6 – 1.2 Hz) by performing a Fast Fourier Transform (FFT). When the power of the bin in the range of interest crosses the predefined threshold, the criteria is met. It is also possible to apply a power
normalization over the full FFT range instead of the range of interest to detect SWS.
We set this threshold to 0.4 for offline tests, but when we found out it didn’t contribute significantly to the phase error we turned it off during the reactivation study.
Self-centred stimulus property ensures centring of the middle of the stimulus to the target phase. That is, the middle of each stimulus will be aligned with the target phase. The self-centred stimulus property takes into account the stimulus duration and predicted frequency to be able to
start the stimulus half of the stimulus duration before the target phase.
For the up-state, the middle of the stimulus (at 250 ms) was centred to the 90⁰ peak, whereas for the down-state, the middle of the stimulus (at 250 ms) was centred to the 270⁰ trough.
8. Reactivation of Vocabulary
In the reactivation phase during the 3 hours after first SWS onset, Danish words were presented aurally without the Dutch translation. The presentation occurred via loudspeakers placed
approximately 50 cm from the subject’s head (50 dB sound pressure level). In case sound
presentation led to arousals or signs of waking up, the sound pressure level was lowered by about 5 dB (Cox et al., 2014d). Exposure to the Danish words occurred in both sleep stages 2 and SWS, when the algorithm predicted an up-state (for group 1) or down-state (for group 2) in the ongoing EEG signal.
Of the 120 learned words, 60 words were cued and 60 were not cued during the night. In group 1, words got cued in a slow oscillation up-state and in group 2, words got cued in a slow
oscillation down-state. The 60 cued words consisted of 30 words that subjects remembered during the pre-sleep learning test (30 correct) and 30 words that subjects did not remember during the pre-sleep learning test (30 incorrect) (see figure 4 for an overview). The words were individually and randomly assigned to each group for each subject (Schreiner & Rasch, 2014).
Figure 4. Schematic overview of the cued word distribution. 60 words got cued in the
up-state (group 1) or down-up-state (group 2) and 60 words were left uncued. The cued words consisted of 30 correct recalled words after the learning phase and 30 incorrect words in both groups.
Methods 13
9. Vocabulary memory recall task
After a full night of sleep in all experimental groups, the retrieval of the Danish words was tested again. Theretofore, the 120 learned Danish words were mixed with 30 new Danish words
(Schreiner & Rasch, 2014). After a 500 ms fixation cross, the Danish word was presented aurally for 500 ms (70 dB sound pressure level). Then a 1000 ms interval occurred after which the subject had 20 s to type whether they had learned the word in the pre-sleep test (“old”) or heard the word for the first time (“new”) as a recognition test. Irrespective of their response subjects had 20 s to type a Dutch translation. Participants were told to always type something, when necessary a guess. Afterwards, they had to fill in their confidence about the given translation in a
confidence interval of 0 to 4, where 0 indicates “I took a guess”, 1 = “I am very unsure”, 2 = “unsure”, 3 = “sure” and 4 = “very sure”. After 40 trials, there was a one minute break.
The performance on the recall task was taken as a post-sleep behavioural performance and vocabulary memory performance was calculated using the following formula:
Proportion correct = Total amount correct post-sleep / total amount correct pre-sleep * 100%
To see whether TMR during up- or down-states leads to more remembering of forgetting at the individual item level, we categorized each word. A “Gain” means the word was not recalled correctly in the pre-sleep test, but was correctly retrieved in the post-sleep test, a “Loss” means the word was recalled correctly in the pre-sleep test, but not anymore correctly retrieved in the sleep test, a “HitHit” meaning the word was recalled correctly in both the pre- and post-sleep tests and finally the “MissMiss” category meaning the word was not recalled correctly in both pre- and post-sleep tests (Schreiner & Rasch, 2014).
Data analysis
10. Offline analysis of algorithm performance
To check the performance of the algorithm offline, we used the same signal as the algorithm got as input, which was channel Fpz re-referenced to the average of mastoids (Fpz-𝑋̅M). We then band-pass filtered this signal with a Butterworth filter order 1 with central frequency 0.9 Hz and a band-pass width of 1 Hz (meaning low bound = 0.4 Hz and high bound = 1.4 Hz). We then used the Hilbert transform to obtain the analytic signal phase. Due to this Hilbert transform, the analytic signal is phase-shifted with respect to the original signal, so to correct for this it was shifted back by adding 0.5 π. We then obtained the phase in degrees for the start of the sound
and the middle of the sound. For the up-state cueing, the target phase at the middle of the sound was 90⁰, whereas the target phase for the down-state cueing was 270⁰.
11. Pre-processing EEG data
Offline EEG analysis was performed using custom made Matlab scripts combined with several freely available toolboxes. For the pre-processing of the EEG data, functions of the EEGlab toolbox (http://sccn.ucsd.edu/eeglab) were used. EEG data was re-referenced to the average of mastoids, high-pass filtered (0.1 Hz) and notch filtered (48 – 52 Hz). Epochs of 3500 ms were made beginning 1000 ms before and lasting 2500 ms after sound onset. The 1000 ms interval preceding sound onset served as baseline, to ensure baseline correction for approximately a whole slow oscillation in both the up- and down-state conditions (Batterink et al., 2016). All trials were then visually inspected, trials containing artefacts were rejected and in trials containing noisy channels, these were interpolated.
12. Sleep scoring
Sleep stages were scored using Galaxy software (PHIi, Amsterdam, The Netherlands) with an epoch size of 30 s according to standard criteria (Rechtshaffen & Kales, 1968). Sleep scoring was
performed by three independent scorers.
13. Analysis of Power Changes
For time-frequency analysis, a family of complex Morlet wavelets was used to decompose all epoched time series into time-frequency representations per channel, using the default settings in the Fieldtrip toolbox (Oostenveld, Fries, Maris & Schoffelen 2011). Power estimates were decibel normalized according to dB power = 10*log10(power/baseline), where for each channel and frequency, the baseline was the average from -200 to 0 ms over all epochs. For the
time-frequency plots, data from -500 ms to 1500 ms on a linear scale and 5 to 80 Hz on a logarithmic scale were used.
14. Statistics
The behavioural data was analysed using R Studio software (R Core Team, 2015) performing non-parametric unpaired t-tests for all data.
Results 15 For statistical analysis on time-channel data (ERPs), the average EEG amplitude measured over the interval from -1000 ms before and lasting until 2500 ms after sound onset was compared between conditions, using permutation-based statistics. To correct for multiple comparisons, a false discovery rate (FDR) of P < 0.05 was used (Benjamini & Yekuteli, 2001). ERP statistics were performed with the EEGlab toolbox (http://sccn.ucsd.edu/eeglab).
For the time-frequency-channel data (TFA), EEG statistics were performed on an interval from -500 ms before and lasting until 1500 ms after sound onset. The Fieldtrip (Oostenveld et al., 2011) Monte Carlo / permutation-based statistics were used. To correct for multiple
comparisons, a cluster correction was used with the cluster-alpha parameter set to α<0.05 (Maris & Oostenveld, 2007).
To calculate statistical values regarding the signal phase, the circular statistics (Berens, 2009) toolbox was used.
The significance level was set to P ≤ 0.05 for all tests.
Results
1. Phase-predicting algorithm parameter study
To test which phase-predicting algorithm criteria contribute significantly to the phase error, we performed a regression analysis between the phase error and the phase-predicting criteria. For offline tests, EEG stretches containing SWS cycles of eight different subjects were made.
Regression analysis showed the phase fitting error threshold, total fitting error threshold (minimal fit accuracy) and crunch time (allowed time the prediction can be made in) contributed
significantly to the restraint of the phase prediction error (see Table 1).
To test to which values the criteria contributing to the phase error the most we should set for desirable results on accuracy in real-time phase predictions, we performed a second regression model on the the phase fitting error threshold (minimal fit accuracy) and crunch time (allowed time the prediction can be made in). Results indicated that the phase fitting error should be set at maximum 0.2 together with a maximum crunch time of 200 ms to have the most accurate phase-prediction results (see Figure 5).
Regression model on the phase error
Model statistics Value Comments
Adjusted RІ 0.031522 Rsq - the variance (0-1)
explained by the model
F(5,7109) 47.30928 Statistical power
p <0.00001 Highly significant
Std.Error of Estimate 50.17789 degrees
Figure 5. Regression model on the phase error for the phase fitting error threshold and the
crunch time algorithm criteria. Results indicated that the phase fitting error should be set at maximum 0.2 together with a maximum crunch time of 200 ms to have the most accurate phase-prediction results.
Results 17
Table 1. Regression analysis of the phase-predicting algorithm criteria.
Parameters statistics b* Std.Err. b Std.Err. t(7109) p-value
Intercept 48.85 4.37 11.18 0.000001***
Central Frequency Power 0.02 0.012 7.53 4.91 1.53 0.13
Phase Fitting Error 0.08 0.02 32.55 6.74 4.82 0.000001***
Total Fitting Error 0.04 0.02 44.32 18.98 2.34 0.02*
Signal Amplitude -0.01 0.01 -0.01 0.01 -0.80 0.42
Crunch Time 0.15 0.01 0.02 0.002 13.01 0.000001***
Regression analysis between the phase error and the phase-predicting showed the phase fitting error threshold, total fitting error threshold (minimal fit accuracy) and crunch time (allowed time the prediction can be made in) contributed significantly to the restraint of the phase prediction error. *p<0.05, ***p<0.001.
2. Speed
Figure 6. Measured hardware lag of the closed-loop (amplifier – recording computer –
To measure the total delay of the closed-loop set-up, we sent triggers to the amplifier, split them with the Polybench software on the recording computer and measured the incoming time at the dedicated algorithm computer. This closed-loop hardware lag was 17.65 ± 5.95 ms and was consistent over measurements. The analysis time of the algorithm predictions was 6.33 ± 2.54 ms. Summed up this gives a total closed-loop delay of 24.00 ± 8.49 ms (see Figure 6).
3. Offline algorithm performance
Figure 7. Offline algorithm performance with target phase 0⁰. SWS data stretches of 8
subjects were used to make phase predictions with target phase for sound onset at 0⁰. The mean phase error was 8.22⁰ ± 28.5⁰. 99% of the predictions fell in the right state. Used criteria to make these predictions are listed in the graph.
Fitting range 75-100% (1 sec or 512 samples) Phase fitting threshold 0.14
Amplitude threshold 75 µV Power threshold 0.4
Crunch time 10 - 30 ms
N 364
Cues in right state 99%
Results 19 To test the performance of the phase-predicting algorithm offline, SWS data stretches of 8 subjects were used to make phase predictions with target phase for sound onset at 0⁰. The mean phase error was 8.22⁰ ± 28.5⁰. In total 364 predictions were made, of which 99% fell in the right state. As criteria for these offline predictions a phase fitting threshold of 0.14 was used,
accompanied by an amplitude threshold of 75 µV, central frequency power threshold of 0.4 and a crunch time between 10 and 30 ms.
4. Online algorithm performance
(A) Actual phase at sound onset (B) Actual phase at sound centre phase locked to 90⁰
Figure 8. Online algorithm performance during the up-state (90⁰) targeting. The figure
(A) represents the actual phase of the start of all given cues during the night in the up-state (group 1, N=24). The mean phase error was 5.71⁰ with 50.74⁰ standard deviation over all cues. Figure (B) represents the sound centre of the cue, which was phase locked at 90⁰ at 250 ms after sound onset. In total 4188 cues fell in the up-state bin from 0⁰ to 180⁰, whereas 1011 cues hit the wrong phase and ended up in the down-state bin from 180⁰ to 360⁰, resulting in an up-state hitting rate of 80.6%. The mean phase error was 9.41⁰ with 57.18⁰ standard deviation over all cues.
For the algorithm performance in real-time ongoing EEG signal, both the phase at sound onset and sound centre were calculated. For the up-state (90⁰) targeting, 4281 predictions were made in total. The mean error at sound onset was 5.71⁰ ± 50.74⁰ for all cues. The mean error at sound centre was 9.41⁰ ± 57.18⁰ over all cues. In total 4188 out of 4281 predictions fell in the up-state bin from 0⁰ to 180⁰, whereas 1011 cues hit the wrong phase and ended up in the down-state bin from 180⁰ to 360⁰, resulting in an up-state hitting rate of 80.6% (see Figure 8).
5. Behavioural results
Up vs. uncued (group 1) Down vs. uncued (group 2)
Figure 9. Up-state cued performance compared to down-state cued performance. Memory
for up-state cued items (103.91% ± 2.45%) was significantly improved when compared to down-state cued items between groups (94.44% ± 1.11%). Proportion correct was calculated as a percentage of the total items correct post-sleep / pre-sleep. Values represent mean ±SEM. ** p < 0.01.
Memory for up-state cued items (103.91% ± 2.45%) was significantly improved when compared to down-state cued items between groups (94.44% ± 1.11%, p < 0.01). Proportion correct was calculated as a percentage of the total items correct post-sleep divided by the total items correct pre-sleep. For memory performance of cued items compared to uncued items within groups, no significant effects were found. The uncued memory performance (99.00% ± 1.93%) in the up-state group was compared to the memory performance of the up-up-state cued items (p>0.05). In the down-state group, the uncued memory performance (99.38% ± 1.98%) was compared to the memory performance for the down-state cued items (p>0.05).
In the down-state cued group, uncued items (56.00 ± 0.58) were recognised better during
92% 94% 96% 98% 100% 102% 104% 106% 108% Up cued Uncued
Proportion correct
92% 94% 96% 98% 100% 102% 104% 106% 108%Down cued Uncued
Proportion correct
Results 21 persistent when we corrected for the amount of items in each condition, uncued item recognition (93.33% ± 0.96%) compared to down-state cued item recognition (82.78% ± 2.78%, p<0.05). For recognition between up-state cued items (86.39% ± 1.50%) compared to uncued items (84.77% ± 2.15%), no significant differences were found (p>0.05) (see Table 2A for all up-state cued results).
We found a difference at trend level between down-state cued (28.33 ± 0.33) and uncued (35.33 ± 2.19) items during the retrieval test (p<0.1) (see Table 2B for all down-state cued results). We also found a trend on the amount of word gains in the up-state cued group when we compared up-state cued (4.38 ± 0.53) to uncued items (3.04 ± 0.39, p=0.05). A word is
considered a “gain” when subjects did not recall the word correctly in the pre-sleep vocabulary test, but did recall correctly after the sleep cueing (see Figure 10). This gain effect persists when we correct for the amount of possible word gains by using the formula:
% Change gains = absolute value of gains / pre-test amount incorrect * 100%
The percentage change of gains in the up-state cued condition (15.61% ± 1.76%) was higher compared to the unced condition (11.13% ± 1.42%, p<0.1) at the trend level, indicating up-state cueing might improve the gain of words on an individual item level.
Figure 10. Amount of gained items in the up-state cued group, N=24, p= 0.05. Values
represent mean ±SEM.
0 1 2 3 4 5 6
Up cued GAIN Uncued GAIN
Table 2. Overview of memory performance
(A) Up-state cued (group 1)
Cued Uncued p-value
Learning 29.92 ± 0.08 30.5 ± 2.12 0.79 Retrieval 31.08 ± 0.73 29.88 ± 1.98 0.53 Change +1.17 ± 0.73 -0.63 ± 0.48 0.10 % Change 103.91 ± 2.45 99.00 ± 1.93 0.20 Gains 4.38 ± 0.53 3.04 ± 0.39 0.05 † % Change gains 15.61 ± 1.76 11.13 ± 1.42 0.09 † Losses 3.17 ± 0.40 3.71 ± 0.37 0.32 % Change losses 10.57 ± 1.32 13.21 ± 1.40 0.17 HitHit 26.75 ± 0.38 26.92 ± 2.09 0.93 MissMiss 25.71 ± 0.53 26.13 ± 2.09 0.97 Recognition Hits 51.63 ± 0.84 50.63 ± 1.27 0.28 Recognition % 86.39 ± 1.50 84.77 ± 2.15 0.30 (B) Down-state cued (group 2)
Cued Uncued p-value
Learning 30 35.67 ± 2.85 0.18 Retrieval 28.33 ± 0.33 35.33 ± 2.19 0.08 † Change -1.67 ± 0.33 -0.33 ± 0.67 0.27 % Change 94.44 ± 1.11 99.38 ± 1.98 0.21 Gains 2.33 ± 0.33 3.00 ± 0.58 0.18 % Change gains 7.78 ± 1.11 12.24 ± 1.60 0.11 Losses 4.00 ± 0.58 3.33 ± 0.33 0.42 % Change losses 13.33 ± 1.92 9.38 ± 0.75 0.10 Recognition Hits 49.67 ± 1.67 56.00 ± 0.58 0.05 * Recognition % 82.78 ± 2.78 93.33 ± 0.96 0.05 * (C) Uncued (group 3)
Cued Uncued p-value
Learning 57.67 ± 2.86
Retrieval 58.00 ± 3.76 0.86
Change +0.33 ± 1.74
% Change 100.36 ± 3.05
Data are means ± SEM. Gains and losses are shown in both absolute values and values corrected for the amount of possible gains (absolute value of gains / pre-test amount
incorrect) or possible amount of losses (absolute value of losses / pre-test amount correct). † = p<0.1, * = p≤0.05.
Results 23 When we corrected for the amount of words subjects encoded during the pre-sleep vocabulary learning task, we did find a cueing effect for cueing in the up-state. For subjects passing the threshold of ≥55 words out of 120 correct during the pre-sleep task, 104.9% ± 3.1% of the cued words were recalled correctly during the post-sleep task, compared to 96.8% ± 1.4% of the uncued words (p<0.05). We set the pre-sleep test performance as a 100% baseline for each subject, meaning that on average, subjects actually gained word knowledge in the up cued condition, whereas they lost word knowledge in the uncued condition. This effect is only found when subjects reached a pre-sleep encoding threshold on the learning task, suggesting they need to have a certain amount of encoding or motivation for cueing to have an effect (see Figure 11 for this comparison).
Figure 11. Up cued vs. uncued with succeeded learning phase (≥55 correct)
N=17, *p<0.05. Values represent mean ±SEM.
6. Event-Related Potentials (ERPs)
To compare the induced effects of cueing at the neuronal level, we tested a control group of people who didn’t receive any cues, but with up-state predictions in the EEG data. The ERP analyses shown here are on the selected channel Fz (for a whole brain analysis, see
Supplementary figure 1). When real sound onsets in the SO up-state are compared to up-state predictions (Figure 12A), the ERP reveals that the algorithm uses a whole slow oscillation with amplitude >65 µV for the fitting in the baseline in both groups. There is a high similarity of ERPs in the baseline period between real up-state sounds and control up-state predictions,
0,92 0,94 0,96 0,98 1 1,02 1,04 1,06 1,08 Up cued Uncued
Proportion correct
*Figure 12. Grand-average ERPs for channel Fz. Grey shading indicates the time windows of
significant differences at FDR p<0.05 level. (A) Up-state sound onsets compared to control (sham) up-state predictions. (B) Down-state sound onsets compared to control up-state predictions. (C) Up-state sound onsets compared to down-state sound onsets.
suggesting these brainwaves are physiologically comparable. On average, both real and control cues start at the beginning of an up-state slow oscillation (sound onset at 0 ms). However, the sound seems to induce an extra slow oscillation, with a big negative amplitude (-65µV) around 500 ms to 738 ms after stimulus onset (p<0.01). This enhancement of the slow oscillation
rhythm by cueing was also seen in previous studies (Marshall et al., 2006; Massimini et al., 2007; Ngo et al., 2013).
Comparison of ERPs between the down-state cues and the up-state control predictions (A) Up vs. control
group
(B) Down vs. control group
(C) Up vs down group
Results 25 (Figure 12B) show that for the down-state group, a positive up-state is used by the algorithm as baseline. The beginning of a down-state at sound onset (0 ms) is expected. However, sound onset seems to interrupt the ongoing SO pattern. At sound onset, the ERP changes into a positive direction for ~200 ms and than continues the down-state pattern with a large negative amplitude (-80µV). The down-state cued ERP remains more negative than the control ERP.
The ERP between the up-state cued items and the down-state cued items (Figure 12C) demonstrated up-state sounds were presented on an upward going wave, while the down-state sounds were presented when the beginning of a down-state was expected. However, down-state cueing seems to interrupt the ongoing down-state SO pattern and in both up- and down-cued groups the ERP reveals a large negative wave in almost the same time window (400 – 700 ms after sound onset).
7. Topoplot
Figure 13. Topoplot. Scalp maps representing the topographical distribution for grand
average event-related potentials (ERP) of up- and down-state targeted stimuli. The time window is between 800 and 1100 ms after stimulus onset. Red dots indicate significant electrodes at p < 0.05, false discovery rate (FDR). Results indicate a pronounced frontal distribution.
8. Time-Frequency Analysis (TFA)
In order to demonstrate the process of cueing on a neural level, we analysed oscillatory responses to vocabulary cues during the slow oscillation up-state compared to a slow oscillation control prediction. We analysed power changes in the whole brain responding frequency range (5 – 100 Hz) in the time window -200 ms before stimulus onset and lasting 1500 ms after. After correcting
for multiple comparisons using cluster correction (Maris & Oostenveld, 2007), we revealed enhanced activity in the fast spindle (12 – 15 Hz) and low beta (15 – 20 Hz) range in the time window 750 to 1500 ms after stimulus onset. Moreover, an enhanced theta activity (5 - 10 Hz) at 300 to 750 ms after stimulus onset was found, which may reflect the processing of language and recollection (Röhm et al. 2001; Bastiaansen et al., 2005; Schreiner et al., 2015b).
Figure 14. Time-frequency analysis. Oscillatory power during cueing in the up-state
group minus control up-state predictions. An enhancement in theta activity (5 -10 Hz) at 300 – 750 ms after the cue is observed, together with fast spindle (12 -15 Hz) & low beta (15-20 Hz) activity at 750 – 1500 ms after the cue. The red line indicates a positive cluster, the blue line indicates a negative cluster (Figure made by Leander de Kraker).
Results 27 9. Time spent per sleep stage and number of reactivations
Table 3. Sleep and reactivation parameters
Up cued (group 1) Control (group 4) p-value Duration (min) S1 32.56 ± 3.61 14.60 ± 4.92 0.02 * S2 265.75 ± 9.29 219.30 ± 38.86 0.30 S3 42.75 ± 3.03 37.90 ± 9.38 0.64 S4 14.56 ± 3.57 31.85 ± 9.88 0.16 REM 102.23 ± 7.08 105.75 ± 18.00 0.86 Wake 52.27 ± 9.51 32.25 ± 20.11 0.40 Movement 6.77 ± 1.96 6.80 ± 1.37 0.98 Total 516.94 ± 3.93 448.45 ± 53.40 0.27 SWS (S3 + S4) 57.31 ± 4.53 69.75 ± 15.16 0.47 Sleep latency 28.79 ± 4.52 24.80 ± 8.60 0.69 WASO 23.48 ± 6.78 13.90 ± 10.83 0.48 Duration (%) S1 6.3 ± 0.7 3.56 ± 0.81 0.03 * S2 51.4 ± 1.7 51.28 ± 3.75 0.98 S3 8.3 ± 0.6 7.92 ± 2.00 0.87 S4 2.8 ± 0.7 6.41 ± 2.00 0.15 REM 19.7 ± 1.3 21.80 ± 3.88 0.63 Wake 10.16 ± 1.9 7.56 ± 3.57 0.54 Movement 1.3 ± 0.1 1.47 ± 0.23 0.55 Total 100 100 1.00 SWS (S3 + S4) 11.1 ± 0.9 14.34 ± 3.26 0.39 Sleep latency 5.6 ± 0.9 4.86 ± 1.64 0.72 WASO 4.61 ± 1.34 2.69 ± 2.08 0.46 Sleep efficiency 90.20 ± 1.95 92.54 ± 3.58 0.58 Number of reactivations S2 48.61 ± 11.88 SWS (S3 + S4) 134.91 ± 22.31 S1 0.30 ± 0.13 Movement 0.61 ± 0.16 REM 0.26 ± 0.22 Wake 0.74 ± 0.20 Total 185.48 ± 29.39 Cues / min 3.43 ± 0.26
10. Amount of cues compared to memory performance correlations
There seems to be no correlation between the amount of cues during the night and the cued vocabulary memory performance (r = - 0.15, t= - 0.70, df= 21, p=0.49). When we distinct the amount and percentage of cues given in SWS on memory performance, a slightly negative correlation is found, yet not significant (r = - 0.30, t= -1.43, df=21, p=0.17). When we consider the amount and percentage of cues given in sleep stage 2 (S2) compared to the memory
performance, a slightly positive correlation is found, however not significant either (r = 0.24, t= 1.16, df=21, p=0.26).
Figure 15. Correlation between the total amount of cues and the cued memory performance. Pearson’s product-moment correlation r = - 0.15, t= - 0.70, df= 21, p=0.49.
Results 29
Figure 16. Correlation between the percentage of cues in SWS and the cued memory performance. Pearson’s product-moment correlation r = - 0.30, t= -1.43, df=21, p=0.17.
Figure 17. Correlation between the percentage of cues in sleep stage 2 (S2) and the cued memory performance. Pearson’s product-moment correlation r = 0.24, t= 1.16, df=21,
Discussion
Our findings show that cueing prior learned foreign vocabulary during a SO up-state enhances vocabulary memory whereas cueing prior learned foreign vocabulary during a SO down-state seems to diminish vocabulary memory. Moreover, down-state cued items are recognised worse compared to uncued items. On the neural level, cueing during a SO is associated with
interruption of the ongoing SO pattern and a strong negative SO-wave in ERPs. We also found an increase of fast spindle (12–15 Hz), low beta (15-20 Hz) and theta band activity (5–10 Hz) associated with the cues. This might reflect the processing of language and recollection (Röhm et al. 2001; Bastiaansen et al., 2005; Schreiner et al., 2015b). With regard to the newly developed phase-predicting algorithm, our findings show that it is faster and more accurate than previously reported methods (Ngo et al. 2013; Cox et al., 2014d; Santostasi et al. 2016).
Cox et al. (2014d) online
Santostasi et al. (2016) offline
van Poppel et al. (2016) offline
van Poppel et al. (2016) online Phase error 22.25 ± 70.7° 12.51 ± 28.85° 8.22 ± 28.5° 5.71 ± 50.74°
SO state Up and down Only up Up and down Up and down
Cues in right state ? 79% 99% 81%
Data time TCP/IP - 20 ms 5 ms 5 ms
Total delay ? 70 ± 5 ms 6.33 ± 2.54 ms 24 ± 8.49 ms
Amount cues /
min ? ? 2.6 3.43 ± 1.16
Subjects 12 5 8 24
Table 4. Overview of currently available up- and down-state predicting algorithms.
Data represent mean ± standard deviation.
Our new phase-predicting algorithm is more accurate in both offline and online tests compared to existing phase-predicting methods. For offline tests, EEG stretches containing SWS cycles of eight different subjects were taken as input and the algorithm predicted upcoming up-states. Our mean phase error was 8.22 ± 28.5° and 99% of the cues ended up in the up-state compared to a mean phase error of 12.51 ± 28.85° and 81% of the cues in the up-state in a previously reported comparable method (Santostasi et al. 2016). In real-time ongoing data stretches, our mean phase error was 5.71 ± 50.74°, compared to the mean phase error of 22.25 ± 70.7° in a previously reported comparable method (Cox et al., 2014d). When predicting upcoming phases in real-time
Discussion 31 data, the accuracy seems to be highly dependent on the noisiness of the signal, while the closed-loop total delay stays robust and convenient. Also, one has to find a perfect balance on the amount of predictions per minute and the yielded phase error, since those are highly dependent on each other. In conclusion, our phase-predicting algorithm is faster and more accurate than any of the previously reported methods (Ngo et al. 2013; Cox et al., 2014d; Santostasi et al. 2016) (see table 3 for an overview of previously reported methods).
When the new phase-predicting algorithm was used to cue prior learned words in the up-state of an ongoing SO, we found the up-up-state cueing enhances vocabulary memory when a sufficient pre-sleep level of encoding is reached. This is consistent with previous findings, suggesting memory traces are only consolidated during sleep when they are sufficiently strong encoded upon entering sleep (Cox, Tijdens, Meeter, Sweegers & Talamini, 2014c; Creery, Oudiette, Anthony & Paller, 2015). The effect of cueing during a slow oscillation up-state reveals an
enhancement of the slow oscillation amplitude at the neural level. In the ERP, a strongly negative trough is associated with a cue compared to normal slow oscillations. This is consistent with other up-state cueing studies (Ngo et al., 2013; 2015). As an electrophysiological oscillatory response, we find an increase in the theta activity band (5 – 10 Hz) in the time window 300 to 750 after stimulus onset. Other reactivation studies do find this theta activity increase also (Cox et al., 2014d; Schreiner & Rasch, 2014). It may reflect the processing of language (Röhm et al. 2001; Bastiaansen et al., 2005) but is also linked to the process of recollection (Düzel, Neufang & Heinze, 2005). We also find an increase in the fast spindle (12 – 15 Hz) range 750 to 1500 ms after sound onset. This fast sleep spindle activity seems to play a key role in declarative memory consolidation (Schabus et al., 2004; 2008; Mölle et al., 2011; Cox et al., 2012; Rasch & Born, 2013; Cox et al., 2014a).
We are the first to report that TMR during an SO up-state enhances vocabulary memory compared to TMR during an SO down-state, which seems to decrease vocabulary memory. Moreover, we are the first to report that items cued during an SO down-state are recognised worse the following morning compared to items that were left uncued during the night. This in line with the underlying neuronal activity which is the underlying base of a slow oscillation. The up-state is associated with global synchronous neuronal depolarisation and an increased neuronal excitability, whereas the down-state is associated with widespread neuronal hyperpolarisation and strongly depressed firing in neurons in both the neocortex and the hippocampus (Sejnowski & Destexhe, 2000; Steriade, 2003; Steriade & Timofeev, 2003; Buzsáki & Draguhn, 2004; Massimini et al., 2004; Isomura et al., 2006; Steriade, 2006; Olcese et al., 2010; Cox et al., 2014b).
The memory enhancement we found induced by cueing in a slow oscillation up-state is in line with previous reactivation studies, where contextual details like odor (Rasch et al., 2007; Rihm et al., 2014; Cox et al., 2014a), sound (Rudoy et al., 2009; van Dongen et al., 2012; Batterink et al., 2016) or
the learned stimuli itself, like melodies (Anthony et al., 2012; Schönauer et al., 2013) or a foreign language (Schreiner & Rasch, 2014) were used as cues. In addition, our study shows cueing with prior learned words in the up-state of an ongoing slow oscillation leads to both memory
enhancement and the enhancement of the slow oscillation rhythm, which is consistent with other up-state targeting studies (Tononi et al., 2010; Ngo et al., 2013; 2015).
On the individual item level, we found an indication that there are more “gains” in the up-state cued items compared to the uncued items. A “gain” is an individual word which was not remembered correctly before sleep, but is remembered correctly after sleep. This is consistent with other studies which looked into the “gains” on the individual item level, which found there are more items “gained” after nocturnal sleep compared to wakefulness (Dumay, 2016), and a TMR study which found cued items “gain” more compared to uncued items overall (Schreiner & Rasch, 2016a).
We have to report that results found in the down-state cued group should be interpreted cautiously, due to the small amount of subjects in this group (N=3). However, these preliminary results are interesting, since it’s indicating that applying cues in the SO down-state diminishes the memory and recognition for this items. Further research on this matter is advisable, since the findings can be of use in the treatment of patients suffering from posttraumatic stress disorder (PTSD). In this disorder, patients suffer from nightmares induced by traumatic events. We propose to cue the traumatic events during a SO down-state to see whether the memory for these events diminishes.
Pilot data of the vocabulary learning task show that the Dutch translation for some Danish words is easy to guess. This guessing cannot explain the effect of cueing during sleep, however, since words are randomly assigned to a cued or uncued condition and the amount of correct and incorrect words in the cued condition is fixed for all subjects. The degree of difficulty of words might play a role on the depth of encoding and words that are easily guessed are
possibly less deep encoded than other words, since more guessable items are following an expectation pattern and are associated with a more shallow encoding (Sweegers, Coleman, Van Poppel, Cox & Talamini, 2015).
When we compared SO up-state cued items to uncued items, we found no cueing effect on the group-level. This can be due to the fact that the enhancement of TMR depends on a sufficient level of encoding (Cox et al., 2014c; Creery et al., 2015). It also possible that only some individuals take advantage of TMR, whereas other people don’t show an enhancement after TMR (Rudoy et al., 2009; Creery et al., 2015). These individual differences could arise from different
Discussion 33 capability for learning a foreign language or capability for learning in general. Another possible explanation of the inconsistent group-level results, is the newly proposed working model for stabilising reactivated memory traces during sleep (Schreiner & Rasch, 2016b). This model proposes that a neuronal theta – gamma activity interaction and a sleep spindle occurring together or shortly after theta activity increase are necessary for a memory trace enhancement. They further propose the theta – gamma interaction is required for the strengthening of the memory trace and an associated sleep spindle is required for the stabilisation and integration of the reactivated memory trace during sleep. This is an interesting train of thought, since sleep spindles seem to play a functional role in both declarative memory consolidation (Schabus et al., 2004; 2008; Mölle et al., 2011; Rasch & Born, 2013) and motor memory consolidation (Lustenberger et al., 2016).
Moreover, the positive involvement of sleep spindles in memory consolidation is SWS specific (Cox et al., 2012). Further research on this matter is needed.
Future research should indicate whether the enhancing effect of TMR on memory is a longer lasting effect than the following morning. This knowledge can be obtained by using longer retention periods. Furthermore, it should be investigated whether the positive effect of TMR with words is persistent after a full night of sleep, instead of waking up immediately after the cueing period (Schreiner & Rasch, 2014; 2015). This is of importance when people want to use TMR to improve their vocabulary in real-life setting, without disrupting their normal sleep pattern. When TMR during the night improves vocabulary memory in a real-life setting, it is of importance for people who want to learn a foreign language, for example students, refugees or immigrants.
In conclusion, we developed a new oscillatory phase-predicting algorithm that is faster, more accurate and easier to apply than previously described methods. Up-state cued items show an enhancement on vocabulary memory, whereas down-state cued items seem to deteriorate vocabulary memory. Moreover, down-state items are recognised worse compared to uncued items. When a pre-sleep level of encoding is reached, the TMR during a slow oscillation up-state in Non-REM sleep enhances the vocabulary memory. This TMR is accompanied on the
electrophysiological level with an induced negative SO-like wave lasting from 500 ms to 738 ms (N550 component) in the ERP and increased theta (5 – 10 Hz), fast spindle (12 – 15 Hz) and low beta (15 – 20 Hz) activity in the TFA, suggesting these oscillations play a crucial role in the strengthening of memories during TMR in sleep.
Acknowledgments
This work was supported by the Brain & Cognition research programme of the University of Amsterdam. I would like to thank Lucia Talamini, PhD, for the never-ending trust in me performing my own research, unlimited access to the sleep lab and necessities and helpful comments on earlier versions of my master thesis. I would like to thank Ilia Korjoukov for scripting the phase predicting algorithm, Roy Cox, PhD for help on custom made Matlab scripts and Klaus Linkenkaer-Hansen, PhD for speaking in all the Danish words. For assistance in data collection, I would like to thank Milanne Buiten, Christa van der Heijden, Leander de Kraker, Sjoerd Manger, Chioma Nwatarali, Angela Rosink and Veronique Vael. I would like to thank Lysanne van Beek, Allison McDonald and Mitja Seibold for sleep-scoring the data. For helpful comments on earlier versions of my master thesis, I would like to thank Kjilleke van Baren, Seth van Heeringen, Eric van Poppel and Iris Stoltenborg.
References
ANT, Enchede, The Netherlands, https://www.ant-neuro.com/
Antony, J. W., Gobel, E. W., O'Hare, J. K., Reber, P. J., & Paller, K. A. (2012). Cued memory reactivation during sleep influences skill learning. Nature neuroscience, 15(8), 1114-1116. Audacity software, version 2.0.3., http://audacityteam.org.
Batterink, L. J., Creery, J. D., & Paller, K. A. (2016). Phase of Spontaneous Slow Oscillations during Sleep Influences Memory-Related Processing of Auditory Cues. The Journal of Neuroscience, 36(4), 1401-1409.
Bastiaansen, M. C., Van Der Linden, M., Ter Keurs, M., Dijkstra, T., & Hagoort, P. (2005). Theta responses are involved in lexical—Semantic retrieval during language processing. Journal of cognitive neuroscience, 17(3), 530-541.
Benjamini, Y., & Yekutieli, D. (2001). The control of the false discovery rate in multiple testing under dependency. Annals of statistics, 1165-1188.
Berens P. (2009). CircStat: A MATLAB toolbox for circular statistics. Journal of Statistical Software, 31(10), 1–21.
Buzsáki, G., & Draguhn, A. (2004). Neuronal oscillations in cortical networks. Science, 304, 1926–1929.
Chauvette, S., Seigneur, J., & Timofeev, I. (2012). Sleep oscillations in the thalamocortical system induce long-term neuronal plasticity. Neuron, 75, 1105–1113.
Circular Statistics toolbox, version 2012a,
https://www.mathworks.com/matlabcentral/fileexchange/10676-circular-statistics-toolbox--directional-statistics-.
References 35 Cox, R.., Hofman, W.F. & Talamini, L.M. (2012). Involvement of spindles in memory
consolidation is slow-wave sleep specific. Learning & Memory, 19, 264-267. Cox, R., Hofman, W.F., de Boer M. & Talamini, L.M. (2014a). Local sleep spindle modulations in relation to specific memory cues. NeuroImage, 99, 103-110.
Cox, R., van Driel, J., de Boer, M. & Talamini, L.M. (2014b). Slow oscillations during sleep coordinate interregional communication in cortical networks. The Journal of Neuroscience, 34, 16890-16901.
Cox, R., Tijdens, R.R., Meeter, M.M., Sweegers, C.C.G. & Talamini, L.M. (2014c). Time, not sleep, unbinds contexts from item memory. PLoS ONE, 9, e88307.
Cox, R., Korjoukov, I., de Boer, M. & Talamini, L.M. (2014d). Sound Asleep: Processing and Retention of Slow Oscillation Phase-Targeted Stimuli. PLoS ONE, 9, e101567.
Creery JD, Oudiette D, Antony JW, Paller KA. (2015). Targeted memory reactivation during sleep depends on prior learning. Sleep, 38, 755–763.
Diekelmann, S. & Born, J. (2010). The memory function of sleep. Nature Review Neuroscience, 11, 114–126.
Dumay, N. (2016). Sleep not just protects memories against forgetting, it also makes them more accessible. Cortex, 4e11.
Düzel E, Neufang M, Heinze H-J. (2005). The oscillatory dynamics of recognition memory and its relationship to event-related responses. Cerebral Cortex, 15, 1992–2002.
EEGLAB toolbox, version 13.5.4b, https://sccn.ucsd.edu/eeglab/. EventIDE, version 2013.11.11, build 23-5-2016, http://okazolab.com. Fieldtrip toolbox, version 2016.03.22, http://fieldtrip.fcdonders.nl.
Galaxy software version 2.7.14, Personal Health Institute international, Amsterdam, The Netherlands, http://www.phi-i.com/
Isomura, Y., Sirota, A., Ozen, S., Montgomery, S., Mizuseki, K., Henze, D.A. & Buzsáki, G. (2006). Integration and segregation of activity in entorhinal-hippocampal subregions by neocortical slow oscillations. Neuron, 52, 871–882.
Isphording, I. E., & Otten, S. (2011). Linguistic distance and the language fluency of immigrants. Ruhr Economic Paper, (274).
Isphording, I. (2013). Disadvantages of Linguistic Origin: Evidence from Immigrant Literacy Scores. Discussion Paper IZA, 7360.
Lustenberger, C., Boyle, M. R., Alagapan, S., Mellin, J. M., Vaughn, B. V., & Fröhlich, F. (2016). Feedback-Controlled Transcranial Alternating Current Stimulation Reveals a
Functional Role of Sleep Spindles in Motor Memory Consolidation. Current Biology, in press. Maris, E., & Oostenveld, R. (2007). Nonparametric statistical testing of EEG-and MEG-data. Journal of neuroscience methods, 164, 177-190.
Marshall, L., Helgadottir, H., Mölle, M. & Born, J. (2006). Boosting slow oscillations during sleep potentiates memory. Nature, 444, 610–613.
Massimini, M., Huber, R., Ferrarelli, F., Hill, S. & Tononi, G. (2004). The sleep slow oscillation as a traveling wave. Journal of Neuroscience, 24, 6862–6870.
Massimini, M., Ferrarelli, F., Esser, S.K., Riedner, B.A., Huber, R., Murphy, M., Peterson, M.J., and Tononi, G. (2007). Triggering sleep slow waves by transcranial magnetic stimulation. Proceedings of the National Academy of Sciences USA, 104, 8496–8501.
Mölle, M., & Born, J. (2011). Slow oscillations orchestrating fast oscillations and memory consolidation. Progress in Brain Research, 193, 93-110.
Mölle, M., Bergmann, T.O., Marshall, L. & Born, J. (2011). Fast and slow spindles during the sleep slow oscillation: disparate coalescence and engagement in memory processing. Sleep, 34, 1411–1421.
Ngo, H.V., Martinetz, T., Born, J. & Mölle, M. (2013). Auditory closed-loop stimulation of the sleep slow oscillation enhances memory. Neuron, 78, 545-553.
Ngo, H.V., Miedema, A., Faude, I., Marinetz, T. & Mölle, M. (2015). Driving sleep slow oscillations by auditory closed-loop stimulation – A self-limiting process. The Journal of Neuroscience, 35, 6630-6638.
O’Neill J, Pleydell-Bouverie B, Dupret D, Csicsvari J. (2010). Play it again: reactivation of waking experience and memory. Trends in Neuroscience, 33, 220–229.
Olcese, U., Esser, S.K. & Tononi, G. (2010). Sleep and synaptic renormalization: a computational study. Journal of Neurophysiology, 104, 3476 –3493.
Ong, J. L., Lo, J. C., Chee, N. I., Santostasi, G., Paller, K. A., Zee, P. C., & Chee, M. W. (2016). Effects of phase-locked acoustic stimulation during a nap on EEG spectra and declarative memory consolidation. Sleep Medicine, 20, 88-97.
Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J. M. (2011). FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational intelligence and neuroscience, 2011, e156869.
Oudiette D. & Paller K.A. (2013). Upgrading the sleeping brain with targeted memory reactivation. Trends in Cognitive Science, 17, 142–149.
Petroni, F. & Serva, M. (2008). Language distance and tree reconstruction. Journal of Statistical Mechanics: Theory and Experiment, 8, 8-12.
R Core Team (2015). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/.
Rasch, B., Büchel, C., Gais, S. & Born, J. (2007). Odor cues during slow-wave sleep prompt declarative memory consolidation. Science, 315, 1426-1429.
Rasch, B., & Born, J. (2013). About sleep's role in memory. Physiological reviews, 93(2), 681-766.
Rechtschaffen, A., & Kales, A. (1968). A manual of standardized terminology, techniques and scoring system for sleep stages of human subjects. Los Angeles: Brain Information Service/Brain Research Institute, University of California.
Rihm, J.S., Diekelmann, S., Born, J. & Rasch, B. (2014). Reactivating memories during sleep by odors: Odor specificity and associated changes in sleep oscillations. Journal of Cognitive Neuroscience, 26, 1806-1818.
Röhm, D., Klimesch, W., Haider, H., & Doppelmayr, M. (2001). The role of theta and alpha oscillations for language comprehension in the human electroencephalogram. Neuroscience letters, 310(2), 137-140.
