The data in this study have been used in two previous studies (Pietto et al., 2016; Parra et al., 2017). It consists of people with familial MCI and their controls and people with sporadic MCI and their controls, detailed separately below. In both cases patients were evaluated with the Mini-Mental State Examination (MMSE) with results previously described (Pietto et al., 2016). The tasks were performed in an electrically shielded room with dim lighting. Participants sat comfortably at a desk facing the task display screen. The subjects were checked to ensure that none had a history of psychiatric or neurological diseases.

The familial MCI data consisted of 10 patients diagnosed with MCI (age 44.4 ± 3.2, years of education 7.3 ± 4.1) and 10 healthy controls (age 44.3 ± 5.6, years of education 6.8 ± 2.9) from Antioquia, Colombia. Each patient carried the mutation E280A of the presenilin-1 gene which leads to familial AD in 100% of carriers. The data consist of sixty-channel EEG activity recorded with a 64 channel EEG cap using SynAmps 2.5 in Neuroscan at 500 Hz and bandpass filtered from 1 to 100 Hz with impedances below 10 KΩ. Four ocular channels were discarded after being used to factor out oculomotor artifacts. The patients had not yet developed clinical symptoms warranting a diagnosis of dementia. Analysis of power-frequency spectrum showed that this data had been subject to a low pass filter with cut-off at 20 Hz.

The sporadic MCI data consisted of 13 patients diagnosed with MCI (age 73.1 ± 9.0, years of education 14.1 ± 4.4) and 19 healthy controls (age 67.2 ± 10.14, years of education 16.5 ± 2.0) recruited from the Institute of Cognitive Neurology (INECO), Buenos Aires, Argentina. Criteria implemented for diagnosis derived from Petersen (2004) and Winblad et al. (2004). Nine of the patients were at particularly high risk from AD conversion having been classified as single or multi-domain amnestic MCI while three classified as non-amnestic MCI multi-domain (Mitchell et al., 2009). The data consist of EEG activity recorded with a Biosemi Active 128-channel Two system at 512 Hz and bandpass filtered from 1 to 100 Hz. This was then downsampled to 256 Hz.

For both datasets we use data from the encoding period during the performance of shape only and shape-color binding tests since deficits at this stage seem to be responsible for the VSTM binding problems found in AD (Parra et al., 2017). This consists of 1.2 s of continuous activity with 0.2 s pre-stimulus. Signals were re-referenced to an average reference before proceeding, following (Chella et al., 2016). Further oculomotor artifacts were removed using visual inspection and independent component analysis and epochs with other artifacts exceeding ±100 μV were discarded as detailed in Pietto et al. (2016). We seek to uncover underlying physiological substrates of the impaired binding function. In this way incorrect responses are not informative so only the trials where the subject responded correctly are included (Pietto et al., 2016).

From a neuropsychological perspective, the two groups of patients presented with similar backgrounds, Table 1. Both groups showed similar level of global cognitive impairment as denoted by the MMSE with instrumental abilities (IADL) denoting very mild but similar level of impairment. Memory and executive functions were affected in both groups as denoted by the recall of the Rey Figure and Fluency Tests. Attention was preserved in both groups as informed by the TMT-A (see Pietto et al., 2016 for details on performance). Taken together these data suggest that both groups were in very similar stages of multiple-domain amnesic MCI (maMCI) (Albert et al., 2011).

All participants provided written informed consent in agreement with the Helsinki declaration and the studies were approved by the Ethics Committees of the University of Antioquia and INECO.

The binding function of visual short-term memory (VSTM) is singled out by contrasting tasks for the recognition of colored shapes, which requires binding of shape and color in memory retention (binding), and the recognition of single shapes which only requires the retention of constituent features. In the change detection task assessment of VSTM for shape alone, the arrays consist of three different black shapes and in the binding task the arrays consist of three different shapes each with a different color. Each task trial consists of an encoding period (500 ms), during which a study array is displayed on screen, followed by an unfilled short delay (900 ms) and test period with a test array. During the test period, participants are prompted to respond whether or not the objects in the two arrays are identical. The positions of the objects are randomized between arrays to avoid use of location as a memory cue. Both shapes and colors are chosen randomly for each trial from a set of eight shapes and a set of eight colors. A randomly chosen fifty percent of the trials have the same objects in both arrays. In the other 50 percent, two shapes seen during the encoding periods are replaced with two new shapes selected from the set, whereas in the binding task two colored shapes of the test display swap the colors they had during the encoding period. All participants start with a brief practice session before undergoing one hundred trials per task. Binding and shape tasks are delivered in a counterbalanced order across participants. Figure 1 shows an example trial for the two conditions of the VSTM binding task.

Both groups showed similar level of performance across task conditions i.e., Shape Only vs. Shape-Color Binding (MCI Z = 1.54 and their Controls Z = 1.42; MCI-FAD Z = 1.63 and their controls Z = 1.17; see Pietto et al., 2016). While MCI patients showed poorer performance than controls on both conditions of the STM binding test, MCI-FAD patients showed poorer performance than controls only on the Shape-Color binding condition. The source of such a discrepancy has been recently addressed (Parra et al., 2019). Assessing older samples (i.e., MCI) with large set sizes (3 items) might reduce the discrepancy classically reported between the two conditions of the STM binding test. The authors argued that this does not undermine the specificity of the Shape-Color binding condition for AD but reflects the influence that memory load exerts on patients with more advanced cognitive impairments (i.e., maMCI) (Petersen, 2004)—see also Parra et al. (2010a,b). This is reinforced by the observation that the two groups presented with a very similar profile of Shape-Color Binding impairment (Mann–Whitney U: 63, Z = −0.09, p = 0.93, d = 0.02; see Pietto et al., 2016).

The Phase-Lag Index (PLI) was computed to assess the phase-dependent functional connectivity of the EEG channels (Stam et al., 2007b). This measures the strength and consistency of pairwise lead/lag relationships of electrode activity of the brain's electromagnetic pulses. Such phase-based measures are particularly useful for assessing interregional dependencies from EEG due to their immunity to the volume conduction effect. Alpha and Beta have been frequently found to show deviations in connectivity of subjects with dementia (Tijms et al., 2013). Thus, the PLI is computed for each trial and for each signal pair after being band-passed in Alpha (8–13 Hz) and lower Beta (13–20 Hz), using an order 70 FIR filter. Note, only lower Beta could be compared across datasets due to the previously mentioned different low-pass filters implemented. These connectivity computations are then averaged over trials for each task and for each subject to remove inter-trial variability and so better bring out the specific task function. The resulting averages constitute adjacency matrices of weighted networks, one for each subject-task-frequency band triple.

Before studying the network hierarchies, the weighted PLI connectivity networks are binarised using the Cluster-Span Threshold (CST). This threshold is based on the clustering coefficient, fixing the network at the balance point of integrative and segregative properties (Smith et al., 2015). It coincides with where hierarchical information is dense, providing a sensitive and powerful binarisation of EEG PLI connectivity (Smith et al., 2017a). The hierarchy of a network is defined based on the node degrees, i.e., number of edges adjacent to each node. Nodes with more adjacent edges are higher in the node hierarchy, being more central to the network topology.

We study two indices of network hierarchies described in Smith and Escudero (2017). An illustration of what these measure in a network is shown in Figure 2. The degree variance, V, measures the spread of the hierarchy and thus is indicative of the large range of the general strength of network nodes (Smith and Escudero, 2020) and is an important indicator of the dominance of hub nodes. Here we use the recently proposed normalized version (Mones et al., 2012):

Where k is the variable denoting the degrees of the network, n is the number of nodes, m the number of links, and d the link density.

Physiologically, this measure could then inform about the expansion and/or strengthening of network structures and hierarchies. Complexity on the other hand, arises from the structure of the interactions between units (i.e., modules, nodes, or networks themselves) (Smith et al., 2019). The hierarchical complexity, R, is based on the diversity of connectivity patterns throughout the degree hierarchy. It is measured by the variability of neighborhood degree sequences for nodes of identical centrality (BioSemi Headcaps, 2017; Smith and Escudero, 2017). Thus, hierarchical complexity (R) could be understood physiologically as the level of network operational organization.

Differences of network index values for binding and shape are computed. These differences are contrasted between patients and controls using Wilcoxon rank sum tests with statistical significance noted at the standard α = 0.05 level. The false discovery rate procedure is implemented over reported p-values with q = 0.05. Effect sizes using Cohen's d are reported for significant differences. Wilcoxon signed rank tests were also implemented between datasets for the healthy controls and the MCI subjects for both shape and binding tests in order to assess whether any task activity could be discerned to be different due to aging.

We computed V and R for all subjects in both tasks in Alpha and lower Beta (Beta1) in both datasets. Differences were computed between binding and shape values for patients and controls separately. These values were then compared between patients and controls using Wilcoxon rank sum tests. Table 2 shows the results for each data set separately and also when healthy controls and MCI subjects in datasets are combined. A significant difference is noted in V in Beta1 for the familial MCI data. In 9 of the 10 patients, the degree variance of the binding condition is larger than in the shape condition, Figure 3, bottom left. Indeed, the difference in binding and shape conditions is generally greater than the difference found in healthy controls with an effect size of 1.1627. Although the trend appeared similar in the sporadic case, this was not found to be statistically significant. However, a significant difference between sporadic MCI and control was noted in R in Alpha with an effect size of 1.0124, which was not replicated in familial MCI and control. In 12 of the 13 patients, the hierarchical complexity of the binding condition was larger than in the shape condition, Figure 3, top right. The controls are roughly balanced between higher shape and higher binding values but larger values tended toward higher shape.

Both the greater degree variance in MCI in Beta1 and greater hierarchical complexity in MCI in Alpha held as effects when combining the datasets. The p-values were roughly of the same magnitudes as in the individual dataset differences found, although the effect sizes were comparatively reduced.

The network indices V and R are based on variances of degrees which can be expected to be higher for larger network sizes. Thus, to assess healthy aging of VSTM binding we downsampled the sporadic 128-channel dataset to the same size (sixty) as the familial dataset. These datasets use different layout systems for electrodes, but approximate mapping between these layouts is known (Parra et al., 2014). Following this, we reprocessed the sporadic dataset according to the 60-channel format and proceeded with comparisons. For V and R, task contrast values for each participant were attained by taking the difference of binding and shape only values. Wilcoxon rank sum tests were then conducted for older versus middle-aged adults. The results are shown in Table 3. As hypothesized, we report no significant differences in hierarchical characteristics of phase-based functional connectivity due to aging.

Our results indicate that cases of MCI due to familial AD present with greater hierarchical spread (V) seemingly accounting for VSTM binding deficits. These results fit well the current understanding of network reorganization in carriers of AD mutation (e.g., E280A-PSEN1) prior to the dementia onset. For example, using a measure of information sharing (i.e., symbolic mutual information), Parra et al. (2017) recently reported that increased brain connectivity characterizes cases in the early stages of familial MCI while decreased connectivity was a feature of more advanced stages. Patients in the early stages not only showed increased connectivity but over-recruitment, and such changes correctly characterized 90% of the sample. The literature reporting increased connectivity and over-recruitment of task-related networks and DMN as an early feature of AD is growing rapidly (Parra et al., 2013; Gardini et al., 2015; Quiroz et al., 2015; Serra et al., 2016). This functional reorganization appears to be an early manifestation of brains undergoing neurodegeneration. This is particularly relevant in this sample of middle-age mutation carriers as they do not present with the comorbidities and risk factors that are normally associated to age. Hence, this evidence more genuinely indicates the presence of AD pathology and its neurobiological consequences.

Interestingly, AD-related network changes at older ages are characterized by greater hierarchical complexity (R). Hierarchical complexity is a new paradigm for brain networks which has previously been explored in EEG signals and structural MRI from healthy participants (Smith et al., 2017a, 2019) as well as on networks across broad scientific domains (Smith, 2019). It has also shown clinical relevance in structural MRI with implications for neonatal development (Blesa et al., 2021; Valdes Hernandez et al., 2021). This is the first study to apply this approach to AD. We found that such hierarchical changes were apparent in a lower frequency band (Alpha). We interpret this as the additive effects of age and AD related network changes. These findings can have some implications. Firstly, they suggest that hierarchical complexity may be an index of compensatory network coupling whereby networks operating in a particular frequency regime that becomes less efficient with aging might be compensated by networks operating at different (i.e., slower) frequencies. It is well-known that such compensation occurs at a neuroanatomical level (Heuninckx et al., 2008; Ho et al., 2012; Song et al., 2014). Based on this earlier evidence and the results presented here we feel compelled to suggest that the greater hierarchical complexity observed at a slower frequency band in age-related MCI may be informing on cross-frequency compensatory coupling. Networks operating at slower frequency bands, which remain functional in old age, may inherit the functions of decaying faster networks. Second, this evidence adds to the Scaffolding Theory of Aging and Cognition (Reuter-Lorenz and Park, 2014) as it suggests a different level of functional reorganization (see Sala-Llonch et al., 2015) characterized by cross-frequency network compensation. Finally, future research will have to investigate the extent to which these compensatory changes revealed via increased network complexity are reflecting adaptive mechanisms to cope with the effects of age, the influence of cognitive reserves, or underlying subthreshold pathology.

Early and genetically driven AD pathology operating in younger brains alters the topology of a task-related network supporting VSTM binding by increasing its hierarchical spread (strength and recruitment) while in older brains experiencing MCI with unknown genetic factors, it hampers their hierarchical complexity. We can speculate that the latter indicates the additive effect of aging on AD pathology. Homeostatically, younger brains are better equipped to cope with pathology and to reorganize functional networks. AD impacting on older brains may encountered a less favorable biological scenario. Such a scenario has already witnessed compensatory changes (Deary et al., 2009; Reuter-Lorenz and Park, 2014; van Geldorp et al., 2014; Bastin, 2017) which are normally characterized by a loss of functional and structural network efficiency and organization. Interestingly, we found that such changes did account for VSTM binding deficits in MCI in old age, but not for a decline of VSTM binding functions due to age per se (i.e., we reported no significant differences in hierarchical characteristics of phase-based functional connectivity due to aging). ERP data involving patients at risk of AD due to different disease variants and with very different ages indicated that neither of these factors modified the effect that AD exerted on VSTMB, thus suggesting that its decline can be more reliably linked to the AD pathology (Pietto et al., 2016). Nevertheless, it remains of a paramount importance to develop and refine methodologies that can tease apart the contribution of AD and that of normal aging.

This is the first report of network related activity associated to the well-known insensitivity of VSTM binding (i.e., in its conjunctive form) to normal aging (Grady, 2008; Deary et al., 2009; Logie and Maylor, 2009; Reuter-Lorenz and Park, 2014; Wakefield et al., 2014; Sala-Llonch et al., 2015; Spaan, 2016). It is also the first to study network heterogeneity and hierarchical complexity of EEG networks in a clinical setting. The evidence presented here indicates that the behavioral specificity of VSTM to AD relative to normal aging also holds at the biological level when age-related compensatory changes in brain activity are considered. We can therefore hypothesize that the analysis of hierarchical topology of EEG connectivity during VSTM binding performance can be considered a potential diagnostic biomarker for AD. Nonetheless, more studies particularly with larger sample sizes are necessary to confirm these findings.