Scientific Papers

Aberrant oscillatory activity in neurofibromatosis type 1: an EEG study of resting state and working memory | Journal of Neurodevelopmental Disorders


The current study is an extension of the analysis presented in Pobric et al. [16]. Specifically, we conducted oscillatory analyses on the same participants, using the same EEG resting state and n-back data, and used some (see the ‘Procedure’ section) of the same behavioural measures described in Pobric et al. [16].

Participants

Thirty-two participants completed this study. Participants were adolescents with NF1 (n = 16) and age- and sex-matched controls (n = 16). These data were collected as part of a pilot intervention study involving the use of transcranial direct current stimulation (tDCS) [NCT03310996].Footnote 1 Post hoc analysis of Ribeiro et al.’s [17] resting state theta effect in NF1 using G*Power [42] suggests that n = 17 participants per group would result in 79% power to detect an effect size of d = 0.89. Participants were required to meet each of the eligibility criteria in Table 1.Footnote 2 We selected a pragmatic sample of children with NF1. Previous literature suggests that social communication difficulties (not meeting the criteria for ASD) may be very common affecting up to 60% of children with NF1 [3]. Given how common comorbid neurodevelopmental conditions such as ASD, ADHD, and developmental coordination disorder are in NF1, we decided to include a sample that would be representative of the children seen in the clinic. Parents/guardians gave oral and written consent, and adolescents assent (where developmentally appropriate), prior to participation.

Table 1 Eligibility criteria

The NF1 sample was recruited through the Manchester Centre for Genomic Medicine, Neurofibromatosis charities, social media platforms, and newsletters and were adolescents who satisfied the National Institute of Health’s (1988) diagnostic criteria for NF1 [43]. The control sample (CON) was age- and sex-matched at group level and recruited via institutional newsletter advertisements and contacting local schools. Demographic information of the sample is reported in Table 2. There were no significant differences between groups in age (t(30) = 0.540, p = 0.593) or sex (χ2 = 0.00, p = 1.00).Footnote 3 In a previous paper reporting on this sample [16], the NF1 group demonstrated poorer performance relative to typically developing controls on various cognitive measures, including measures of working memory (i.e. digit-span forward/backwards) and attention (Sky search attention, TEACh).

Table 2 Participant demographics

Procedure

This study received ethical approval from the Greater Manchester West Research Ethics Committee (17/NW/0364) and was conducted in accordance with the Declaration of Helsinki. During the study visit, participants and their parents/guardians were first familiarised with the EEG equipment and study procedures. Subsequently, a battery of behavioural and cognitive assessments was administered, including parent-rated and cognitive measures that tapped into: overall adaptive function, inattention, hyperactivity, communication, daily living skills, socialisation, short-term memory, working memory, sustained attention, and attentional switching, followed by EEG. This paper focuses on parent-reportedFootnote 4 Adaptive Behaviour Composite (ABC) scores on the Vineland Adaptive Behaviour Scale (VABS-III) [44], performance on an adaptive auditory n-back task [16], and performance on a non-adaptive visual n-back task [16] performed during EEG (see Pobric et al. [16] for details of the other tasks performed that are not reported here).

VABS-II measures daily living skills, socialisation, and communication, with ABC scores reflecting standardised age equivalent overall adaptive functioning [44]. Performance on the adaptive auditory and non-adaptive visual n-back tasks measure working memory. Each trial of the n-back task began with a fixation cross ( +) presented in the centre of the screen (adaptive auditory n-back: 2500 ms; non-adaptive visual n-back: 2000 ms, + / − random jitter up to 100 ms in 17 ms steps). This was followed by a single uppercase English consonant (C, G, H, K, P, Q, T, or W) presented aurally (auditory n-back: 1000 ms) or visually in the centre of the screen (non-adaptive visual n-back: 500 ms). Participants were instructed to respond as quickly and accurately as possible with a mouse-click whenever the current stimulus was the same as the one presented ‘n’ steps back in the sequence. No responses were required for non-targets. The auditory n-back was adaptive, such that after each block of 20 trials, the difficulty level of the next block was adjusted based on the participant’s performance to ensure participants were always training at the top of their ability (see Pobric et al. [16] for further task details). In contrast, the visual n-back performed during EEG recording was not designed to push participants’ ability to their limit. Instead, it was developed to provide a sufficient number of trials to permit the investigation of electrophysiological differences between the groups during working memory performance. In this non-adaptive task, four fixed-order blocks were presented: 1-back, 2-back, 2-back, and 1-back, with self-paced breaks in between to reduce fatigue. In each block, there were 100 trials, 25 of which were target trials (i.e. the same letter as ‘n’ screens back). As existing studies report a load-related increase in power during working memory maintenance [45], two load levels (‘n’ = 1 and 2) were included to permit investigation of load-dependent effects on EEG measures.

EEG acquisition

EEG data were recorded using an ActiveTwo system (BioSemi, Amsterdam, Netherlands) with 64 EEG channels in standard 10–10 system locations plus HEOG, VEOG, and mastoids, with a sampling rate of 512 Hz. During the recording, participants were asked to remain still, in a comfortable/relaxed position, and to minimise eye movements and blinking where possible. Recording started with 2.5 min of eyes open and 2.5 min of eyes closed resting state, in which participants were asked to simply relax and not think of anything in particular. This was followed by recording during the visual n-back task.

EEG analysis

MATLAB (2019a) and SPM12 (version 7771) were used to conduct data analyses. Custom functions [46, 47] calling several functions from EEGLAB (version 13.6.5b) and FieldTrip (embedded in SPM release) were used.

Common pre-processing

Continuous EEG data were re-referenced to averaged mastoids, high-pass filtered (0.1 Hz), downsampled (256 Hz), low-pass filtered (resting state: 200 Hz; task-related: 120 Hz), and notch-filtered (48–52 Hz), before epoching (resting state: arbitrary 1900 ms (baseline correction: 0–1900 ms, i.e. mean-centring); task-related: 0–1900 ms relative to stimulus onset). The eyes open and eyes closed data were then concatenated (i.e. combined into the same file).

Independent component analysis (ICA) was used to project blink and eye-movement signals out of the data.Footnote 5 Channels containing noise unrelated to blinks (characterised by large positive deflections) or eye movements (characterised by square-wave deflections) were temporarily omitted (channel TP7 was persistently bad and omitted from ICA for all participants). Thirty-two components were extracted from EEG channel data only. ICA components with uniquely high temporal correlations with VEOG and HEOG, and/or uniquely high spatial correlations with the blink topography, were identified using custom code [46] and following the procedure described in Pobric et al. [16]. The resulting weight matrix (less the artefact components) was applied to the epoched data using SPM12’s ‘montage’ function.

Baseline correction was then re-applied on the ICA-cleaned data. Epochs were rejected as noisy if they contained signal that exceeded a threshold (resting state: 200 μV; task-related: 120 μV; higher threshold for resting state data due to higher alpha power during eyes closed resting state). A channel was declared ‘bad’ if the threshold was exceeded on > 20% of trials, and epoch rejection was re-run ignoring any bad channels. To reconstruct these noisy channels, a channel-weight interpolation matrix was created using FieldTrip’s ‘channelrepair’ function and applied to the epoched data using SPM12’s ‘montage’ function. EEG data were then re-referenced to the common average reference. The mean number of components removed, channels interpolated, and trials remaining can be seen in Table 3.

Table 3 Number of ICA components removed, channels interpolated, and trials remaining

Resting state analysis

To be eligible for inclusion, participants were required to have a minimum of 15 valid epochs remaining in each condition (open/closed) after artefact-contaminated trials were removed. Two participants were excluded from the NF1 group, one due to having fewer than 15 valid trials in the eyes open condition and the other due to low-quality data (i.e. all channels automatically marked ‘bad’ during the artefact detection routine). The sample used for the resting state analyses therefore comprised 30 participants (n = 16 CON; n = 14 NF1).

Task-related analysis

Trials with incorrect responses were excluded. Subsequent analyses were conducted on target and non-target trial data without distinguishing between these conditions, since n-back performance requires maintenance of information during both trial types, particularly in the late post-ERP time window (described below). For the same reason, the number of non-target trials was not decimated to match the number of target trials (which was done in the P300 analysis presented by Pobric et al. [16]).

To be eligible for inclusion participants were required to have a minimum of 15 valid epochs remaining in each load (1-/2-back) after artefact-contaminated trials and incorrect trials were removed. One participant from the NF1 group was excluded owing to having fewer than 15 valid trials in the 2-back load level (this was a different participant from the two resting state exclusions). The sample size used for the task-related analyses therefore comprised 31 participants (n = 16 CON; n = 15 NF1).

Spectral power

For estimation of task-related power, the time window of interest was 900–1900 ms post-stimulus onset (i.e. during the fixation cross of the next trial). This time window was chosen as existing studies investigating WM-related oscillatory activity typically use the maintenance period of the working memory task as the time window of interest as increased oscillatory activity is observed during this period [22, 26, 36, 48]. We followed the previous literature’s definition of the maintenance period as the time following a response to the stimuli, determined using the average (or median) response time on the given task [36]. In the current study, the average response time over 1-/2-back blocks was 627 ± 124 ms (median: 601 ms). However, to ensure that the majority of participants had responded, 900 ms was chosen as the beginning of the time window. The time window ended at 1900 ms to provide a sufficient number of samples for power estimation. For consistency, the same time window (in the arbitrary epoch), and therefore number of samples, was used for the estimation of resting state power.

For each EEG channel and epoch (resting state: eyes open and eyes closed; task-related: 1-back and 2-back), a Fast Fourier Transform with a Hanning window and a frequency resolution of 1 Hz was used to extract frequency spectra collapsed over time (900–1900 ms). The resulting power values were then log-transformed before averaging spectra over epochs. For exploratory analysis, average log-transformed power over all EEG channels was computed in canonical frequency bands: delta: 1–3 Hz; theta: 4–7 Hz; alpha: 8–11 Hz; beta: 12–29 Hz; low-gamma: 30–47 Hz, and high-gamma: 53–100 Hz. Additionally, as the literature consistently reports increased mid-frontal theta power during working memory maintenance [22], we performed targeted analysis of task-related mid-frontal theta (4–7 Hz) power. To achieve this, log-transformed power was averaged over channels Fz, F1, and F2 to create a mid-frontal region of interest prior to statistical analysis. We measured absolute power (i.e. power in one frequency band, independent of activity in other frequency bands), as opposed to relative power (i.e. power in one frequency band divided by the amount of activity in all frequency bands) to avoid the potential confound that any abnormalities in one frequency band may affect the relative power of other frequency bands—a particular concern in neurodevelopmental disorder studies [49].

Peak alpha frequency

Differences in peak alpha frequency (PAF) between groups were investigated. PAF was defined as the frequency with the maximum power in a loose alpha range (6.5–13.5 Hz) at channel Pz. Pz was chosen as alpha power is typically high at this channel [50]. For analysis of PAF, for each EEG channel and epoch (eyes open/closed), a Fast Fourier Transform with a Hanning window and a frequency resolution of 0.25 Hz was used to extract frequency spectra collapsed over time (arbitrary 1900 ms epoch). First, each individual’s 1D spectrum was adjusted to reduce 1/f noise as this flattens the spectrum and causes the alpha peak to ‘pop out’ [51]. This was achieved by fitting a second-order polynomial to the log-transformed frequencies (omitting alpha and notch-filter frequencies), and the difference between the spectrum and this model was computed. The resulting spectrum was smoothed with a Gaussian kernel to remove spurious peaks. Next, four adjustments were made based on visual inspection of each participant’s spectrum: Two CON and one NF1 participant had maxima that fell on the ascending slope of the beta peak (eyes open: 12.25 Hz, 13.25 Hz, and 12.75 Hz; eyes closed: 13.25 Hz, 12.75 Hz), and these were adjusted to small visible alpha peaks (eyes open: 11.50 Hz, 10.25 Hz, and 10.75 Hz; eyes closed: 11.75 Hz, 11.75 Hz) that our algorithm had missed, and one NF1 participant’s maximum fell on the descending delta slope (eyes open and closed: 6.5 Hz) and was adjusted to small visible alpha peaks (eyes open: 7 Hz, eyes closed: 7.5 Hz) missed by the algorithm.

Theta phase coherence

Prior to estimating phase coherence, the task-related data were spatially filtered using the Surface Laplacian, implemented using the laplacian_perrinX function in MATLAB [52]. The Surface Laplacian reduces the influence of volume conduction, which is particularly important given the electrode-level connectivity analysis performed [53]. We investigated theta phase coherence in the frontoparietal network (Fig. 1). The mid-frontal region acted as a seed region and coherence was estimated between this region and left-parietal, mid-parietal, and right-parietal regions [54]. Each region comprised of a set of electrodes: mid-frontal (F1/Fz/F2), left-parietal (P3/P5/P7), mid-parietal (P1/Pz/P2), and right-parietal (P4/P6/P8). Coherence was estimated between each possible mid-frontal–parietal connection (i.e. 27 channel pairs), before averaging coherence over electrode sets, resulting in three coherence estimates: mid-frontal to (1) left-parietal (ML), (2) mid-parietal (MM), and (3) right-parietal (MR).

Fig. 1
figure 1

Graphical representation of the channels included in each region of interest. Grey lines indicate the 27 channel pairs that coherence was computed between. Black lines represent coherence averaged over electrode sets (ML: mid-frontal to left-parietal; MM: mid-frontal to mid-parietal; MR: mid-frontal to right-parietal)

Theta phase was computed for the whole epoch (0–1900 ms) and then phase coherence was computed in the time window of interest, 900–1500 ms post-stimulus (the time window ended at 1500 ms to prevent inclusion of edge effects as per epoch definition). We calculated inter-site phase clustering (ISPC) [20]. ISPC over trials is a measure of the consistency of phase angles between two electrodes averaged over trials. For task-related data, ISPC-trials is an appropriate method given our analysis is hypothesis-driven (i.e. limited to the frontoparietal network) and not exploratory (more suited to weighted phase lag index) [20]. ISPC has been used previously in studies with similar methodology [54]. Phase angle time series for each channel was extracted by convolving the data with a complex Morlet wavelet (4 cycles) separately for frequencies 4 Hz, 5 Hz, 6 Hz, and 7 Hz. For each time point, the average vector length was calculated across trials to quantify ISPC trials, defined as:

$$ISPC_f=\left|n^{-1}\sum_{t=1}^ne^{i\left(\phi_{xt}-\phi_{yt}\right)}\right|$$

(1)

where n represents the number of trials and ϕx and ϕy are phase angles from channels x and y at frequency f. ISPC ranges from 0 (perfectly randomly distributed phases) to 1 (perfect phase-locking). For each channel pair, ISPC trials were calculated for each load (1-/2-back) and frequency (4–7 Hz). The result was then averaged over the time window of interest (900–1500 ms post-stimulus), then over frequencies, and finally over channel sets. This resulted in one coherence value for each frontoparietal region pair/load combination (ML, MM, MR × 1-back, 2-back) for each participant.

Statistical analysis

Statistical analyses were conducted using SPSS (Version 25). The alpha level was set to 0.05. Visual inspection of Q-Q plots showed that, for each analysis, data were normally distributed. For each analysis of variance discussed below, Box and Whisker plots were inspected for extreme outliers. Values were considered extreme outliers if they fell outside of 3rd quartile + 3*interquartile range and 1st quartile − 3*interquartile range. Where extreme outliers were identified, sensitivity analyses were run. It can be assumed that there were no extreme outliers identified where sensitivity analysis is not reported.

In each frequency band, a 2 (CON/NF1) × 2 (open/closed) analysis of variance (ANOVA) was run for the scalp-averaged resting state data. ANOVAs were run separately for each frequency band as there is a known 1/f effect, whereby the means of low frequencies are larger than those of high frequencies [55]. As frequency bands are estimated independently, and therefore each ANOVA is performed on independent data, no correction for multiple comparisons was used. Moreover, to investigate whether resting state power follows the typical reactivity pattern observed in neurotypical populations [39], in each frequency band a paired t-test investigated power differences between eyes open and eyes closed resting state in the NF1 group. A 2 (CON/NF1) × 2 (open/closed) ANOVA was also used to analyse PAF.

In the n-back task, maintenance of items in working memory spans trials. We therefore used eyes-open resting state recordings as a baseline to investigate task-specific power modulation (i.e. change from rest). To achieve this, we divided task-related power by resting state power before log-transforming the dataFootnote 6 (equivalent to: log(task power) − log(resting state power)), which is referred to as task-specific power henceforth. In each frequency band, a 2 (CON/NF1) × 2 (1-/2-back) ANOVA investigated scalp-averaged task-specific power. Consistent with the task-related power analyses, we investigated task-specific theta phase coherence by adjusting for baseline (resting state). For comparability, eyes open resting state theta phase coherence was estimated using the same method as task-related phase coherence (note the ‘trials’ in ISPC-trials are arbitrary in resting state). Prior to statistical analysis, resting state phase coherence was subtracted from task-related phase coherence. A 2 (CON/NF1) × 3 (ML/MM/MR) × 2 (1-/2-back) ANOVA using task-specific frontoparietal theta phase coherence was performed.

As existing research suggests a significant relationship between age and oscillatory activity in typically developing children that may not be present in neurodevelopmental disorders [56], Pearson’s correlations were performed to investigate associations between EEG measures and age, followed by statistical significance testing of the difference in r between groups to determine whether the relationship between age and oscillatory activity was significantly different between groups. This was achieved using Fisher’s r to z transformation, before statistically comparing the resulting z scores. Moreover, as individuals with NF1 typically exhibit worse overall functioning relative to typically developing children [6, 57], and there is suggestion that oscillatory activity might be a neural marker of cognitive function in neurodevelopmental disorders [56], Pearson’s correlations were performed to investigate associations between EEG measures and overall adaptive functioning, using Vineland ABC scores. Again, this was followed by statistical significance testing of the difference in r between groups to determine whether the relationship between overall adaptive function and oscillatory activity was significantly different between groups. Finally, to assist interpretation of the oscillatory findings, Pearson’s correlations were performed to investigate associations between EEG measures and working memory performance on the adaptive auditory n-back task (which was conducted separately to the EEG session), and the difference in r between groups compared. To correct for multiple comparisons, a 5% false discovery rate (FDR) [58] correction was applied to outcomes with p-values less than 0.05. FDR was applied to the set of EEG measures for each demographic/behavioural domain (i.e. five p-values).



Source link