The EEG recording took place in a dimly lit and sound-proof room. Participants were video-monitored by the lab assistants from an adjacent room to ensure they complied to the instructions and that they did not show behavioral indications of drowsiness or sleep onset during the recording. Participants were seated at approximately 80 cm distance from the computer screen. Their chair was equipped with response buttons at both arms. The EEG session started with preparation and placement of electrodes (lasting around 30 min) and continued with the eyes-open baseline recording and two experimental tasks, which took around 2 h. The order of the experimental tasks was counterbalanced across participants. Following the second experimental task, an additional eyes-open baseline recording was performed to explore reliability and stability of EEG measures within resting state recordings, which falls out of the scope of the current experiment. The current analysis is performed on the data from the initial baseline recording and the main experimental task, i.e., the letter-speech sound binding task (see section “Experimental Task Performance”). The additional experimental task that was part of the recording session, i.e., an audiovisual-binding task, was not used in the present analysis as it is intended for event-related analyses.

The EEG was recorded DC (low-pass: 5th order sync digital filter) with a 2048 Hz sample rate. We used a 64-channel Biosemi ActiveTwo system (Biosemi, Amsterdam, Netherlands). The Biosemi system uses two additional electrodes [Common Mode Sense (CMS) and Driven Right Leg (DRL)] located to the left and right of POz, respectively, which replace the conventional ground electrode. All electrode offsets relative to CMS/DRL were brought within 20 μV in accordance with the manufacturer guidelines. The 64 electrodes were distributed across the scalp according to the extended 10–20 International system (see electrode locations in Supplementary Figure A1) and applied using an elastic electrode cap (Electro-cap International Inc.). Ten external Flat-Type Active electrodes were used. Four were used to record vertical and horizontal electro-oculogram (EOG). They were placed below both eyes aligned with the pupils approximately 3 cm outside both outer canthi of the eyes. Two electrodes were placed at mastoids and two were attached to the earlobes to be used as offline reference signals. Finally, two electrodes were used to record the electrocardiogram (ECG) and were placed at the sternum and between the lower two ribs. The ECG data were not used in the current study. Baseline and experimental task.

During the baseline recording subjects were required to look at the center of the screen during 4 min after making a button-press indicating the start of the period. A gray background was used to minimize glare on the screen and a gray fixation circle with shadowing was placed at the center of the screen to assist participants to fixate their eyes while preventing eye fatigue.

The letter-speech sound binding task is a probabilistic learning task in which subjects learned new visual-sound associations via feedback. We used the current format in a previous study examining differences in overt feedback processing between dyslexics and typical readers (Fraga González et al., 2019). In the trials, participants had to learn whether the letter-like unfamiliar symbol was matched with the simultaneously presented speech sound by pressing Yes or No and receiving feedback after their response. However, feedback was only response-related in half of the trials (consistent trials) while in the other half the feedback was random (inconsistent trials; see below in this section). The visual stimuli consisted of 16 symbols from the Georgian alphabet and the auditory stimuli were 16 Dutch phonemes. The complete list of visual symbols and phonemes used in the task is presented in Supplementary Appendix A. The phonemes were spoken by A native Dutch male speaker. There were three groups of phonemes with different durations; one group of four phonemes had a mean (SD) duration of 172.66 (22.28) ms and another group of four phonemes had a mean (SD) duration of 380.50 (19.47) ms. The third group consisted of eight phonemes with a mean duration of 451.97 (27.69) ms. The visual stimuli were presented using an ASUS VG236H (resolution 1,920 × 1,080) 60 Hz monitor with a Dell Optiplex 760 dual-core 3.0 GHz computer and an ATI HD 6570, 2Gb graphic card. The symbols were presented using ‘‘Arial Unicode MS’’ font (lower case, bold font and font size 60). The software used to present the stimuli was Presentation (Version 18.2²). The sound stimuli were presented through padded earphones.

A schematic of the trial structure is presented in Figure 1. On each trial, a visual symbol and a phoneme were presented simultaneously. The trials were terminated by the response. The symbols were presented in white on a black background at the center of the computer screen. Participants had to decide whether the symbol and phoneme presented belong with each other by pressing the buttons located at the right and left arms of the chair. The mapping of YES and NO responses to the right and left hand was consistent across blocks for each participant but was counterbalanced across participants. Green and red stickers were placed on the buttons to indicate whether they were YES or NO buttons, respectively. The button-press was followed by blank screen with 1,000 ms duration. The blank screen was followed by feedback “GOED” (correct; presented in white upper case “Times New Roman” font with size 48), “FOUT” (incorrect; presented in red font), or “TE LANGZAAM” (too slow; presented in upper case “Times New Roman” font with size 48). After the feedback screen, a fixation cross was presented during the inter-trial intervals (ITI) with equiprobable durations of 500, 750, or 1,000 ms.

There were 4 blocks of 200 trials. For each block, two visual-sound pairs were consistently matched, and feedback depended upon the response of the participant. The two other visual-sound mappings were inconsistent and followed by random feedback (50% positive and 50% negative feedback). This feedback probability manipulation was included to analyze differential feedback-responses for informative (consistent trials) vs. uninformative (inconsistent trials) responses in a previous study (Fraga González et al., 2019). Note that the current analysis of task performance only uses consistent trials and the EEG analysis is based on a segment during performance that includes both type of trials. Each trial block contained 100 consistent and 100 inconsistent mapping trials presented in random order (50 replications of each individual trial). The duration of a trial block was approximately 14 min. The task began with a practice block of 30 consistent mapping trials. The average reaction time (RT) on correct responses during practice + 500 ms was used to determine the response window. The feedback “too slow” was provided when responses were executed after this window. Participants were told that they should infer the visual-sound associations from the feedback provided to them and that each trial block contained a new set of associations. In addition, they were told that some associations would be more difficult to learn than others.

The whole experimental session took around 3 h and 15 min including the initial behavioral measurements and the montage of electrodes. There were short rests between blocks and between tasks and resting-baselines depending on the needs of the participant. The participants were debriefed at the end of the experiment and received a monetary reward for their services.

The graph analysis followed similar pipeline steps as in our previous study (Fraga González et al., 2016). The sequence of steps of this pipeline are shown in Figure 2. The continuous EEG data were imported in EEGLAB v.12.5.4b, a Matlab-based open toolbox (Delorme and Makeig, 2004). The averaged earlobes were used as off-line reference when importing the data. In the baseline analysis a segment with a duration of 4 min was selected, time-locked to the button press indicating the start of the eyes-open resting-state recording. In the task analysis we took the initial 4 min from the beginning of the task, after the practice period. The data were high-pass filtered at 0.5 Hz using a zero-phase FIR filter and channels containing excessive artifacts were removed from the data to be interpolated later on in the pipeline (see below in this paragraph). The data were then segmented into 60 epochs with a duration of 4 s each. The epochs were visually inspected and those containing artifacts such as head or electrode cable movement and jaw clinching were removed. Subsequently, we performed an Independent Component Analysis (ICA) decomposition (Makeig et al., 1996) in order to remove blinks, eye-movements and other stereotyped artifacts from the data. We used the “runica” algorithm available in EEGlab for ICA decomposition (Lee et al., 1999) and the automatic algorithm ADJUST to identify independent components associated with artifacts (Mognon et al., 2011). The algorithm uses artifact-specific spatial and temporal features to detect artifactual components and has been previously validated (Mognon et al., 2011). After removing the independent components selected by the algorithm, data for typical readers and dyslexics were reconstructed based on a mean (SD) of 52.67 (7.82) and 49.37 (14.02) components in the task and 52.23 (4.58) and 51.29 (5.90) components in the baseline, respectively. Afterward, the data from previously removed channels were interpolated using a spherical spline interpolation method (Perrin et al., 1989). Finally, for each condition (baseline and task) a total of 30 epochs, each with a duration of 4 s, were selected per participant,³ down-sampled to 1024 Hz and exported to ASCII files for the subsequent EEG analyses.

The ASCII files were imported in Brainwave v0.9.152.4.1 (developed by C.S.; freely available at http://home.kpn.nl/stam7883/brainwave.html) where data were re-referenced to the average of all scalp channels and filtered for each frequency band (see section “Spectral Power”) before performing subsequent analyses.

We calculated spectral power in each epoch using Fast Fourier Transformation (FFT) with a frequency resolution of 1 / 4 s = 0.25 Hz. The power spectra were averaged across segments and all the groups of electrodes described in section ‘‘EEG Preprocessing.’’ The ‘‘area under the curve’’ values were calculated for the following frequency bands: delta (0.5--4 Hz), theta (4--8 Hz), alpha (8--13 Hz),⁴ and beta (13–30 Hz). Relative power was computed as the ratio of the power of the corresponding band and the total power.

We used the Phase Lag Index (PLI) to calculate functional connectivity between all pairs of electrodes for each frequency band and epoch. In contrasts to other connectivity measures like phase coherence, the PLI reduces the effect of volume conduction by ignoring zero and π phase differences (Stam et al., 2007). It captures the asymmetry of the distribution of instantaneous phase differences, which are determined using the Hilbert transformation (Stam et al., 2007). A symmetric distribution centered around zero may indicate spurious connectivity and a flat distribution indicates a lack of connectivity. A deviancy from a symmetric distribution indicates dependency between sources. The PLI is obtained from time series of phase differences Δϕ (t_k), k = 1…N by means of:

Here “sign” is the signum function. The PLI ranges between 0 and 1. A value of 0 means no coupling or coupling with a phase difference centered around 0 (mod π). A value of 1 indicates perfect phase locking at a value of Δϕ different from 0 (mod π). Thus, PLI values closer to 1 indicate stronger nonzero phase locking. In the current analysis we use the term mean total PLI when referring to the average of the PLI between all pairs of electrodes.

For our network analysis, we calculated a Minimum Spanning Tree (MST) for each connectivity matrix (see Figure 2). We used the MST as it allows for direct group or condition comparisons minimizing the bias caused by differences in connectivity strength (e.g., Stam et al., 2014). The MST is a unique sub-graph based on a weighted matrix that connects all nodes of the network without circles or loops. Importantly, the MST always contains m = N−1 links, where N is the number of nodes. The MST was constructed by applying Kruskal’s algorithm (Kruskal, 1956) which iteratively selects the links with the lowest distance (i.e., lowest weights) and adds the link to the tree only if no loops are created. The result is a graph without cycles or loops in which all nodes are connected. In our MST computation, we define a link weight as 1-PLI. Thus, the MST represents the sub-network with maximum connectivity.

There are a various MST metrics that are used to describe the topological properties of the tree (Stam et al., 2014). We examined the following metrics: degree, leaf fraction, diameter, eccentricity, betweenness centrality (BC), tree hierarchy (Th), degree correlation (R), kappa and mean. The degree of a node refers to its number of links, and the leaf fraction represents the number of nodes (N) on the tree with degree = 1. The leaf number has a lower bound of 2 and an upper bound of N− 1. It presents an upper bound to the diameter of the MST, which is the largest distance between any two nodes of the tree. The upper limit of the diameter is d = m− L + 2, where m refers to the number of links on the tree. This formula implies that the largest possible diameter will decrease with the increasing leaf number. Eccentricity of a node is defined as the longest distance between that node and any other node and is low if this node is central in the tree. The BC of a given node u is the number of shortest paths between any pair of nodes i and j that are running through u, divided by the total number of paths between i and j. The BC value ranges between 0 and 1 and relates to the importance of a node within the network. The nodes with the highest BC have the highest load, i.e., the highest number of shortest paths between any two nodes run through these high BC nodes. For example, a central node with a BC of 1 could be easily overloaded. Degree, eccentricity and BC are different measures for relative nodal importance and may indicate the critical nodes in a tree. The measure of tree hierarchy T_h reflects a balance between efficient communication and prevention of overload of hub nodes, reflected, respectively, by small diameter and a maximal BC. This balance is proposed to be important for optimal network performance (Boersma et al., 2013) and is defined as:

Where L is leaf fraction and m the number of links. Further, the degree correlation R is an index of whether the degree of a node is correlated with the degree of its neighboring edges to which it is connected. The R is quantified by computing the Pearson correlation coefficient of the degrees of pairs of connected nodes (Newman, 2003). If R > 0 the graph is considered assortative, and if R < 0 disassortative. Kappa is the width of the degree distribution and relates to spread of information across the tree (Stam et al., 2014). High kappa indicates the presence of high-degree nodes, which facilitate synchronization of the tree but also increase the network’s vulnerability if a hub is damaged (Otte et al., 2015). Finally, we computed the MST mean, that is the mean of the PLI weights of the tree.

The FFT power spectra per condition and group are presented in Figure 3. As expected, there were significant differences between the task and baseline recordings in theta [F(1, 53) = 41.83, p < 0.001, η² = 0.44], alpha [F(1, 53) = 109.88, p < 0.001, η² = 0.68] and beta relative power, F(1, 53) = 32.10, p < 0.001, η² = 0.38. Relative power was significantly larger in the baseline compared to the task (see Figure 4). There was no evidence for significant interactions or main effect of group in these analyses, ps < 0.258.

The main analysis on PLI is presented on Table 3 (see Supplementary Table A3 for all tests that were performed). There was a significant main effect of condition in the alpha band indicating larger PLI in the baseline compared to task [F(1, 53) = 29.02, p < 0.001, η² = 0.35], but no interactions or main effect of group in that band, ps > 0.119. A significant effect in the same direction was detected in the beta band [F(1, 53) = 24.64, p < 0.001, η² = 0.32], together with a trend for lower values over both conditions in dyslexics compared to typical readers, F(1, 53) = 3.1, p = 0.084, η² = 0.06. In the theta band there was no main effect of group or condition (ps > 0.151) but, there was a significant interaction between condition and group [F(1, 53) = 4.45, p = 0.040, η² = 0.08], indicating lower PLI in dyslexics vs. typical readers during the task but not in the baseline. The task vs. baseline in dyslexics but not in typical readers (see Figure 5 and Table 3).

The condition and group interactions were followed by group comparisons in task and baseline data separately (see Table 4 and Supplementary Table A4). In the task, PLI theta was significantly lower in dyslexics compared to typical readers, F(1, 53) = 7.63 p = 0.008, η² = 0.13 (see left panel in Figure 5). The mean (SD) total PLI theta was 0.167 (0.005) and 0.170 (0.005) for dyslexics and typical readers, respectively. The mean total PLI beta was lower in dyslexics compared to typical readers, F(1, 53) = 5.88, p = 0.019, η² = 0.10. The mean (SD) total PLI beta was 0.090 (0.005) and 0.093 (0.006) for dyslexics and typical readers, respectively. The analysis of the baseline data showed no evidence for significant group differences in PLI, although there was trend for stronger alpha connectivity in dyslexics vs. typical readers at p = 0.091, all other ps > 0.388.

The results of the main ANOVA on MST metrics revealed significant group differences across conditions (see Table 3 and Supplementary Table A3). Dyslexics showed lower theta degree correlation, i.e., lower network integration, over both task and baseline recordings, F(1, 54) = 6.36, p < 0.015. In addition, there were significant main effects of condition for all MST metrics except for betweenness centrality in theta and MST mean in beta. The largest effect sizes for the change across conditions were found on degree (alpha) leaf fraction (theta, alpha and beta), kappa (alpha), tree hierarchy (theta and alpha) and degree correlation (theta and alpha) with partial eta-squared > 40. The direction of these differences suggests a less integrated network configuration in task compared to the pre-task baseline. There were significant interactions between condition and dyslexia for theta tree hierarchy, alpha kappa and beta MST mean and theta MST mean (see Table 3). The follow-up analyses on these interactions are presented in Table 4 (and Supplementary Table A4). These analyses showed a trend for lower the tree hierarchy in dyslexics compared to typical readers during task [F(1, 53) = 3.92, p = 0.053] but not in the baseline p = 0.544. For alpha kappa, dyslexics showed a trend for larger kappa than typical readers in baseline, p = 0.060, that was absent in the task 0.948 (see Table 4).