Cognitive and Electrophysiological Correlates of Working Memory Impairments in Neurofibromatosis Type 1

Visuospatial and auditory adaptive n-back tasks were used to assess WM (Conway et al. 2005; Kane & Engle, 2002) in the lab before the EEG session. Based on our pilot results, we used a simplified version of a task designed for testing WM in healthy adults described in previous studies (Jaeggi et al. 2009). The simplified visuospatial stimuli consisted of pictures of animals appearing at one of 4 different loci (instead of 9 in the original task) spaced equally and symmetrically around a constantly present white fixation cross in the centre of a black screen (monitor-to-eyes distance 56 cm). For the auditory task, the verbal material comprised eight aurally presented English consonants (c, g, h, k, p, q, t, w) spoken by a female voice. For both N-back tasks, the stimuli lasted for 1000 ms with an interstimulus interval of 2500 ms (Fig. 1). Participants were instructed to respond as quickly and accurately as possible whenever the current stimulus was the same as the one presented N positions back in the sequence (N depending on the load level, that is, 1, 2, 3; see Fig. 1). The task was adaptive: if the performance was 90% or above in one particular n-back block, the level of the next block was increased by one. If the performance was 70% or below in one particular n-back block, the level of the next round was decreased by one. Otherwise, the level for the next block stayed the same. After two blocks with unchanged n-back level, the experiment was terminated. Within each block there were 20 critical screens plus the additional screens at the start of the sequence needed to create the particular n-back level. (i.e. the first two screens cannot contain a target in a 2-back task). Six out of the 20 screens contained a target and the other 14 screens were non-targets. No responses were required for non-targets. Performance was assessed via two dependent variables– mean n-back (the average n-back level reached) and response times (RT) for target presses.

The Corsi Blocks Task was used as an index of visual short-term memory (Lezak, 1995). Originally developed by Corsi (1972), this task consists of nine cubes mounted on a board (Corsi, 1972). The examiner taps a sequence of blocks, which the participant has to repeat subsequently in the correct sequential order. By increasing the length of the sequences, the capacity of the visuospatial short-term memory can be measured. Participants were asked to complete a computerised version of the task using PEBL software (Mueller & Piper, 2014). Flash time was set at 500 ms, with an Inter-Stimulus-Interval of 1000 ms. When two sequences were incorrectly repeated, the task was terminated. The test measured the mean length of the largest two correctly remembered sequences. The understanding of instructions and tasks was verified with 3 practice trials which were three blocks long.

This is a standardized assessment measure for attention in children and adolescents. Four subtests of the TEA-Ch (Manly et al. 2001) were used including the Sky Search, Score, Creature Counting, and ‘Sky Search Dual Task’ to measure focussed attention, sustained attention, attentional switching, and dual task attention respectively. Raw scores for each sub-test were converted into age and gender corrected scaled scores based on standardised tables provided in the TEA-Ch manual.

Working memory capacity was assessed Digit span forward/backward subtest of the Wechsler Intelligence Scale for Children – Fourth edition (WISC-IV) (Grizzle 2011). Age-scaled scores were used for the analyses.

Parents completed the Vineland which is a measure to assess child adaptive behaviour in communication, socialization and daily living skills domain. Standardized age equivalent overall functioning was computed and is expressed as Adaptive Behaviour Composite score (Hill et al. 2017).

The task consisted of a fixed order (non-adaptive) of four blocks: 1-back, 2-back, 2-back, 1-back. Stimuli were visually presented single letters. Each experimental block consisted of 100 trials, 25 of which were targets (same letter as n trials back). Targets were roughly evenly distributed across each block and only rarely occurred consecutively (6 instances in 2-back blocks). Stimuli that would be considered a target in the other n-back condition (i.e., 2-back target in 1-back block, or 1-back target in 2-back block) were never presented. On each trial, a fixation cross (‘ + ’) was presented in the centre of the screen for 2000 ms (+ / − random jitter of up to 100 ms in 17 ms steps), followed by a single uppercase letter in the centre of the screen for 500 ms. Mouse-click responses were allowed for 2000 ms following stimulus onset (the cursor was not visible on screen). No feedback was given during experimental trials. Stimuli were presented in a light grey font on a black background.

EEG data were recorded with a 64-electrode ActiveTwo system (BioSemi, Amsterdam, Netherlands; 64 EEG channels plus HEOG, VEOG, and mastoids, all sampled at 512 Hz). Data analyses were conducted using Matlab (r2012a) and SPM12 (version 7487; https://​www.​fil.​ion.​ucl.​ac.​uk/​spm/​) (Litvak, 2011) with custom functions (https://​www.​github.​com/​jason-taylor) calling several functions from EEGLAB (version 13.6.5b; https://​sccn.​ucsd.​edu/​wiki/​EEGLAB) (Delorme & Makeig, 2004) and FieldTrip (http://​www.​fieldtriptoolbox​.​org/​) (Oostenveld et al. 2011). A common pre-processing pipeline was applied to all data: Continuous EEG data were re-referenced to averaged mastoids, down-sampled to 200 Hz, high-pass (0.1 Hz), low-pass (120 Hz), and notch (48-52 Hz) filtered, before epoching (-600 to 1400 ms relative to stimulus onset). Independent component analysis (ICA) was used to identify blink- and eye movement-related artefacts. 32 ICA components were extracted from the 64 EEG channels (only) using EEGLAB’s ‘runica’ function (with temporal extension option). A temporal correlation was computed between each component’s time-course and VEOG and HEOG channel data (all signals filtered between 1 and 20 Hz before correlations were computed). Spatial correlations were computed between each component’s channel weights and the topography of that participant’s average blink (blink events detected automatically using SPM12′s ‘eyeblink’ artefact routine; epochs extracted from -100 to 300 ms around blink events; epochs averaged; blink topography defined as time-window average from 50-250 ms on all EEG channels). For both temporal and spatial correlations, z-scores were computed for each IC’s correlation by subtracting the average of all correlations and dividing by the standard deviation of all correlations. The resulting z-scores then index not simply the magnitude of the correlation but how unusual it is relative to other IC-artefact correlations. Components with high (absolute) z-scores (> 2) were identified as ‘suspects’ and their time-courses and channel-weight topographies visually inspected. A channel-weight projection matrix was created to remove components that were confirmed to be related to eye-related artefacts (Controls: M = 1.73 ± 0.46, range 1–2 components removed; NF1: M = 1.88 ± 0.50, range 1–3 components removed). To reconstruct any noisy channels, a channel-weight interpolation matrix was created using FieldTrip’s ‘channelrepair’ function (Controls: 1.40 ± 0.51, range 1–2 bad channels interpolated; NF1: 2.19 ± 0.83, range 1–4 bad channels interpolated; channel TP7 was persistently bad and interpolated for all participants). These two weight matrices were then applied to the epoched EEG data using SPM12′s ‘montage’ function.

ICA-cleaned N-back trial data were cropped to -100 to 900 ms relative to stimulus presentation and baseline corrected (-100 to 0 ms), and a 30-Hz low-pass filter was applied. Epochs containing (absolute) values greater than 120uV on EEG channels were rejected. Incorrect trials were also rejected; P300 analyses were conducted on correct-trial data only. Because correct non-target trials outnumbered correct target trials, trial numbers were balanced by selecting non-target trials that were temporally proximal to target trials and rejecting the rest. This resulted in an average of 39.2 ± 10 valid trials per condition (range 17–49) for Controls and 38.5 ± 9 (range 17–48) valid trials remaining (the number of trials did not differ significantly between groups, t < 1, p > 0.8).

To identify a time-window and channel for subsequent between-group analyses, grand averages over all participants (collapsing over group) were computed for 1-back targets and non-targets, and the difference (target–non-target) computed. Consistent with the P300 literature, the peak difference occurred on channel Pz, with the two conditions diverging from about 300 ms. Visual offset ERPs were present from about 720 ms (stimulus duration was 500 ms). Therefore, a time-window of 300-700 ms on channel Pz was chosen for ‘canonical’ P300 for amplitude and latency analyses. Time-window averaged amplitude and fractional area latency (the point at which 50% of the area under the ERP in the time window was reached) were computed.

For topographic analyses, in light of latency differences found in the Pz analysis (see Results), the time window was split into early (300-500 ms) and late (500-700 ms) time windows. Topographic maps of time-window averaged amplitude are presented for both target and non-target stimuli. For statistical analysis, data were extracted from four clusters of electrodes representing left and right frontal and parietal regions (see Fig. 5). Four electrodes (F3, F4, P3, and P4; underscored in the lists below) were taken as the ‘centroids’ of these regions, and data from each was combined with that of its 6 nearest neighbours:

In the topographic ANOVAs described below, two spatial factors – FP (frontal, parietal) and LR (left, right) – were included.

Unlike the behavioural n-back tasks reported above which featured an adaptive staircase procedure, the EEG n-back task was not designed to push participants’ performance to its limits. Rather, it was designed to provide a sufficient number of trials to analyse ERP responses at two levels of memory load (1-back and 2-back) in both participant groups. Indeed, a similar number of correct trials between groups (i.e., no performance difference) is desirable here, since then ERP would not be biased by the number of trials on which they were based. Statistical analysis of task performance is therefore provided only for completeness.

Pearson’s correlations between age and visual n-back task performance (hits–false alarms) were run for each group, and these correlations were compared between groups (see Supplementary Table S2). None of the correlations differed between groups; therefore, age was used as a covariate in the ANCOVA. N-back performance was submitted to a 2 (group: NF1, Controls) × 2 (n-back: 1-back, 2-back) mixed ANCOVA. There was a main effect of n-back (F(1,28) = 42.009, p < 0.001, η_p² = 0.600), with better performance in the 1-back (M = 85.4 ± 13.9%) than in the 2-back blocks (M = 64.1 ± 20.9%); however, there was no main effect of group (F(1,28) = 0.234, p = 0.632, η_p² = 0.008) and no interaction between group and n-back (F(1,28) = 0.108, p = 0.745, η_p² = 0.004); 1-back: CG: M = 87.3 ± 11.1%; NF1: M = 83.6 ± 16.2%; 2-back: CG: M = 67.2 ± 20.5%; NF1: M = 61.2 ± 21.5%). The covariate age showed a significant main effect (F(1,28) = 10.745, p = 0.003, η_p² = 0.277), but it did not interact with n-back (F = 1.547, p = 0.224, η_p² = 0.052). The correlation between age and n-back performance was positive in both groups and in both n-back levels, indicating that older participants performed better.

For RT in the EEG n-back task, Pearson’s correlations again did not differ between groups (see Supplementary Table S2), so age was covaried in the 2 × 2 ANCOVA. A main effect of n-back was found (F(1,28) = 5.607, p = 0.025, η_p² = 0.167), with faster response times in the 1-back (M = 598 ± 26 ms) than in the 2-back condition (M = 661 ± 25 ms), but there was neither a main effect of group (F(1,28) = 0.396, p = 0.534, η_p² = 0.014) nor a group x n-back interaction (F(1,28) = 1.183, p = 0.286, η_p² = 0.041). The main effect of age was significant (F(1,28) = 4.365, p = 0.046, η_p² = 0.135), but age did not interact with n-back (F(1,28) = 0.048, p = 0.828, η_p² = 0.002). Correlations between age and RT were all negative, indicating that older participants responded more quickly.

Figure 4 shows ERPs from channel Pz and time-window averaged topographies for Controls and NF1 in response to 1-back and 2-back targets, as well as bar charts summarising P300 amplitude and latency in each group. The P300 was maximal on mid-parietal channels (Pz is indicated by green dots on topographies in Fig. 4) in both groups and N-back tasks. ERP time-courses appeared to differ, peaking earlier in NF1 than in Controls, but differentiating between conditions (1-back > 2-back) more in Controls than in NF1. The net result appeared to be no difference in time-window averaged amplitude between the groups, but an apparent difference in latency (NF1 < Controls).

Pearson correlations between age and P300 measures did not differ between groups (see Supplementary Table S3); therefore, age was included as a covariate in all ANCOVA analyses reported below. It is noteworthy that out of 80 age-P300 correlations, only two were significant, which suggests that age did not explain much individual variability in P300 measures overall. Age main effects were non-significant in all ANCOVAs reported below.

In the 2 × 2 Group x N-back mixed ANCOVA on P300 amplitude, a main effect of N-back was found, F(1,28) = 11.077, p = 0.002, η_p² = 0.283 – amplitude was larger for 1-back (M = 14.002 uV, SEM = 1.057) than for 2-back (M = 11.497 uV, SEM = 0.967). Neither the main effect of Group (F < 1, p > 0.8) nor the Group x N-back interaction (F(1,28) = 1.115, p = 0.300) was significant. Age showed no significant main effect or interaction (Fs < 1, ps > 0.3).

The same 2 × 2 Group x N-back ANCOVA was run on P300 fractional area latency. A main effect of Group was found, F(1,28) = 6.788, p = 0.015, η_p² = 0.195 with shorter latencies in NF1 (M = 485.16 ms, SEM = 10.83) than in Controls (M = 522.83 ms, SEM = 11.19). This group effect was modulated by a significant Group x N-back interaction, F(1,29) = 5.794, p = 0.023, η_p² = 0.171. The group difference in latencies was significant in the 2-back task (t(29) = 3.082, p = 0.004, difference M = 54.917 ms, SEM = 17.820) but not in the 1-back task (t(29) = 1.219, p = 0.233, difference M = 20.438 ms, SEM = 16.768). Additionally, whereas Controls’ P300 latency was significantly shorter in the 1-back than in the 2-back task (t(14) = 2.306, p = 0.037, difference M = 21.667 ms, SEM = 9.394), this difference was not significant for the NF1 latency (t(15) = 1.103, p = 0.288, difference M = 12.813 ms, SEM = 11.619). Age showed no significant main effect or interaction (Fs < 3, ps > 0.1).

Figure 5 shows topographic maps for each group (columns: CG, NF1, CG–NF1 difference) and condition (rows: target, non-target, T–NT difference), in each N-back task (upper half: 1-back; lower: 2-back) and time-window (A: early, B: late). Differences between groups are most apparent in the condition difference (T–NT, bottom row of each sub-section) and group difference (CG–NF1, 3rd column of each sub-section). Across time-windows, the Control target–non-target difference topographies remained relatively stable, with a stronger positivity in the 1-back than in the 2-back task, and with symmetrical frontal positivity in both tasks. By contrast, whilst the NF1 T–NT early 1-back topography looked relatively similar to that of Controls, with only a slight Right > Left asymmetry, the NF1 late 1-back topography showed a strongly asymmetrical (right > left) frontal distribution, and both the early and late 2-back NF1 topographies showed a stronger parietal positivity and weaker frontal positivity than was evident in the Controls. This general pattern was largely supported by the statistical analysis presented below.

To relate individual differences in P300 measures and behavioural measures of cognitive function, correlations were run with scores on the Vineland ABC (reflecting overall adaptive functioning), Sky Search (reflecting selective attention), Conners inattention and hyperactivity, and mean n-back on the Auditory N-back task (reflecting working memory capacity, measured independently of EEG). Given the topographic ANCOVA results above, P300 amplitude of the difference (Target – Non-target) was extracted for the following time-window/condition/spatial location combinations: (i) early 2-back left parietal, (ii) late 1-back left frontal, (iii) late 2-back left frontal, and (iv) late 2-back left parietal. Given the group differences in latency found above, 2-back target latency (FAL) at Pz was also included. Correlations are reported in Table 3. A false discovery rate (FDR; Benjamini & Hochberg, 1995) of 10% was applied to each behavioural measure’s set of correlations (n = 10) to correct for false positives whilst remaining sensitive to true positives.

Vineland ABC showed moderate negative correlations with Controls’ P300 amplitude (i.e., higher assessment of general cognition was related to lower P300 amplitude) in 2-back blocks in the left parietal region (early time-window: r = -0.51, p = 0.049, did not survive FDR; late time-window: r = -0.58, p = 0.022), and in 1-back blocks in the left frontal region in the late time-window (r = -0.61, p = 0.015). No ABC-P300 correlations were found for the NF1 group (all |r|< 0.3, p > 0.3). Sky Search showed moderate negative correlations with NF1s’ P300 amplitude (i.e., higher selective attention performance was related to lower P300 amplitude) in the late time-window in 2-back blocks in the left parietal (r = -0.54, p = 0.031) and left frontal (r = -0.46, p = 0.075, did not survive FDR) regions; no such correlations were found in Controls (all |r|< 0.4, p > 0.15). Sky search also negatively correlated with P300 latency (i.e., higher selective attention performance was related to shorter P300 latencies) in NF1 (r = -0.63, p = 0.009), but not in Controls (r = 0.12, p = 0.670). Mean Auditory N-back showed only one moderate (and marginal) negative correlation (i.e., higher auditory N-back performance was related to lower P300 amplitude): Controls’ left frontal amplitude in the late time-window of the 1-back task (r = -0.47, p = 0.077, did not survive FDR). No significant correlations were found for Conners inattention or hyperactivity measures.

Scatterplots of some of these relationships (see Fig. 6) appeared to show a nonlinear inverted-U relationship between P300 amplitude and behavioural measures, particularly in the NF1 group, and in the late time-window. The Matlab function ‘fitglm’ was used to fit generalised linear models to each P300-behavioural pair for each group, with and without a quadratic component, and the Akaike Information Criterion corrected (AICc) was used to test whether the quadratic model was significantly better than the linear model (p < 0.05, where p = exp((AIC_quadratic – AIC_linear)/2).

Two significant (and significantly better than linear) quadratic relationships emerged for NF1 ABC and left frontal P300 metrics: late 1-back (quadratic model: R² = 0.467, AIC = 97.74, p = 0.023; linear model: R² = 0.012, AIC = 103.82, p = 0.697; p(quadratic > linear) = 0.048) and late 2-back (quadratic model: R² = 0.660, AIC = 80.44, p = 0.002; linear model: R² = 0.010, AIC = 93.43, p = 0.911; p(quadratic > linear) = 0.002). A quadratic model also improved the R² by more than 10 percentage points for NF1s’ relationship between ABC and left parietal P300 in 1-back late time-window, but this increase was not significant according to AIC, nor was the quadratic model significant (quadratic model: R² = 0.130, AIC = 97.32, p = 0.433; linear model: R² = 0.015, AIC = 96.00, p = 0.661). In all three cases, the nonlinear relationship was an inverted-U, such that both low and high assessments of general cognitive function were related to relatively low P300 amplitude, whereas mid-range cognition was associated with relatively high P300 amplitude. No other variable combinations showed significant quadratic relationships in either group.