Group sample sizes ranged widely across studies (HC n = 7–135, M_HC = 28.8, SD = 23.4; MCI n = 7–154, M_MCI = 33.0, SD = 26.7; AD n = 6–318, M_AD = 40.0, SD = 47.8). There was a marginal difference across sample types amongst the studies in the review, F(2,283) = 2.95, p = 0.054, η 2 = 0.02. Post-hoc contrasts were checked despite the marginal omnibus effect, given the lack of specific hypothesis and exploratory nature of the comparison; sample sizes were comparable between AD and MCI (7.0, p = 0.20) and MCI and HC (4.2, p = 0.40), with AD samples overall larger than HC samples (11.3, p = 0.02).

Mean age was reported in 105 out of 124 studies (i.e., 85%). Seven additional studies reported only the age range or minimum age of the samples. Across studies that reported mean age, there was a significant overall difference amongst sample types, F(2,232) = 9.97, p < 0.001, η 2 = 0.09. Post-hoc contrasts showed the effect was attributable to younger HC groups (range: 57.4–80.1 years, M_HC = 69.3, SD = 4.8) than MCI groups (range: 61.0–85.5 years, M_MCI = 71.1, SD = 5.1; −1.80, p = 0.02) and AD groups (range: 64.0–83.9 years, M_AD = 72.4, SD = 4.0; −3.12, p < 0.001). Average age was not significantly different between AD and MCI groups (1.30, p = 0.11).

Only 61 out of 124 studies (i.e., 49%) reported average years of education by group. Four additional studies reported education as a binary category (e.g., 5–11 years or > 11 years), and two studies reported only a minimum education level. Across studies reporting mean education, there was a significant difference between sample types, F(2,134) = 5.23, p = 0.007, η 2 = 0.07. Post-hoc contrasts showed that this difference was due to significantly fewer years of education in the AD groups (range: 3.31–14.14 years, M_AD = 9.6, SD = 2.48) when compared with MCI groups (range: 5.86–15.39, M_MCI = 11.0, SD = 2.49; −1.35, p = 0.04) or HC groups (range: 3.36–16.50, M_HC = 11.2, SD = 2.87; −1.87, p = 0.005). HC did not significantly differ from MCI (0.53, p = 0.37).

The distribution of female and male participants was reported in 103 out of 124 studies (i.e., 83%). Across these studies, sex distribution varied widely, with percent female ranging from 25–74% in HC (M_HC = 54.68%, SD = 12.04), 24–76% in MCI (M_MCI = 52.39, SD = 12.74), and 20–86% in AD (M_AD = 57.54%, SD = 13.18). A Kruskal-Wallis non-parametric ANOVA showed no overall a significant difference in distribution of females, H(2) = 2.37, p = 0.31.

83 out of 124 studies (i.e., 67%) reported average MMSE scores (possible range: 0–30) for their samples (Folstein et al., 1983). Of the 41 papers that lacked MMSE data (see Supplementary Tables S1, S2), four used a different metric, 17 reported MMSE in a manner that could not be accurately tabulated (e.g., ranges, cutoffs, medians, unclear group values), and 13 papers provided no specific cognitive information or only reported that participants “met criteria” for the diagnostic category. For the 83 studies reporting MMSE, there was a notable range of MMSE scores particularly in AD groups (HC: 25.73–30.00, MCI: 20.3–28.41, AD: 9.40–26.30). Moreover, there was a significant overall group difference in MMSE, as expected (F(2,181) = 320.15, p < 0.001, η 2 = 0.78). Post-hoc contrasts showed that MMSE was significantly lower in AD than in MCI (−5.91, p < 0.001, M_AD = 19.94, SD = 3.07, M_MCI = 25.85, SD = 1.66) and in HC (−8.67, p < 0.001, M_HC = 28.61, SD = 0.93), and MCI was lower than HC (−2.76, p < 0.001). Notably, MMSE within HC groups was comparable regardless of study sample composition differences (i.e., studies that compared HC with AD, HC with MCI, or all three groups, F(2,79) = 0.30, p = 0.74, η 2 = 0.008).

The papers in this review represented most of the various metrics available for studying EEG connectivity, with some studies applying more than one approach. The primary approaches were coherence (n = 41), phase-locked (n = 35), and graph theory (n = 44), with multiple other less frequently applied approaches also represented (n = 34 studies). Importantly, these metrics analyze fundamentally different aspects of neural connectivity, which can contribute to the appearance of variability across studies (Cohen, 2014). Moreover, even within methods, there were computational differences and methodological choices that might contribute to differences in study findings (see Supplementary Table S3). Four methodological factors deserve particular attention. First, only 29% of studies (i.e., 36/124) analyzed connectivity in source space. Instead, the majority of studies did analyses at the electrode level. Sensor-level approaches are limited by the effects of volume conduction, such that multiple EEG sensors record signals from the same underlying brain region, which can result in connectivity results that reflect activity from shared brain regions (Chiarion et al., 2023; Mahjoory et al., 2017; Michel and Brunet, 2019; Schoffelen and Gross, 2009; Van de Steen et al., 2019).

There was substantial variability across studies in the duration of the data segments used to compute connectivity. For connectivity metrics in the frequency domain, frequency resolution is determined by the number of samples included in the data segments. Durations ranged widely, from 400 milliseconds at 256 Hz (~1,563 samples, ~2.5 Hz resolution; Bagattini et al., 2022) to 40 s at 500 Hz (80,000 samples, ~.025 Hz resolution; Kabbara et al., 2018; but see also Vyšata et al., 2015). Generally, it is recommended to use a time window that includes three cycles at the lowest frequency of interest or at least one second before and after the time window of interest to ensure sufficient frequency resolution (Cohen, 2014). Thus, some of the reviewed samples used too few time points, likely resulting in smearing, which is a distortion (or “blurring”) of the EEG signal that makes it difficult to interpret a signal’s actual frequency (Buzzell et al., 2022; Cohen, 2014).

The studies reviewed primarily analyzed EEG connectivity during the resting state. Indeed, only 19% (i.e., 24/124) of studies analyzed EEG during active task engagement. Eleven of 24 (46%) task-related studies used coherence methods. The tasks used typically tapped memory-related functions (n = 16; e.g., digit span, n-back, mental arithmetic) or basic sensory-attentional processing (n = 7; e.g., visual, auditory, or olfactory oddball paradigms).

Most of the reviewed studies analyzed connectivity using multiple frequency bands. The most common frequency bands were alpha (n = 94), theta (n = 87), beta (n = 84), and delta (n = 72). Gamma activity was less frequently included (n = 46), which is likely attributable to greater concerns of EEG artifact within high frequencies (Muthukumaraswamy, 2013). An additional 23 studies analyzed “broadband” EEG signals, using a range that included activity across multiple frequency bands (e.g., 1-100 Hz). Within alpha and beta bands, about half of the studies divided analyses into lower (alpha-1, beta-1) and upper (alpha-2, beta-2) bands. This approach may be advantageous given the relatively wide frequency range within alpha (typically ~8-12 Hz) and beta (~12-30 Hz), relative to delta (~2-4 Hz) and theta (~4-8 Hz), and some evidence of differential patterns by band subdivision (Bazanova and Vernon, 2014; Klimesch, 1999).

Coherence is a non-directed metric that assesses the relationship between the power spectra of two signals. Coherence metrics are primarily based on the consistency of phase differences between the two sensors or brain regions, which are sensitive to both the phase and amplitude of the signals (Bastos and Schoffelen, 2016; Cao et al., 2022; Srinivasan et al., 2007).

The phase-locked metrics reported in these studies primarily included phase lag index, phase-locking value, and phase synchronization, with variations including weighted, debiased, and lagged measures. These are non-directed frequency domain metrics that assess the phase synchrony between two signals. Unlike coherence, these measures are sensitive specifically to the phase information, rather than both phase and amplitude (Bastos and Schoffelen, 2016; Cao et al., 2022; Chiarion et al., 2023).

Graph theory is an approach that models the brain as a complex network composed of nodes, which represent brain regions or sensors, and edges, representing the connections between them. From this foundation, numerous metrics can be investigated that provide information regarding network function and information flow (Bassett and Sporns, 2017; Bassett et al., 2018; Bullmore and Bassett, 2011). Compared to other connectivity metrics, for which lower values indicate lower connectivity, there is more nuance to interpreting graph theory metric directionality. For example, a larger divisibility value reflects greater separation between nodes, and thus suggests less efficient, or lower, connectivity. Because interpretation is highly dependent on the specific metric, we report group patterns for metrics that had at least five supporting studies. Results for all significant metrics can be found in Table 3.

Graph theory metrics with at least five supporting studies include clustering coefficient, path length, small worldness, local efficiency, and global efficiency. Clustering coefficients provide an index of how closely connected a node is to its neighbors (Masuda et al., 2018; Miraglia et al., 2022; Rubinov and Sporns, 2010; Van Diessen et al., 2014). Path length describes the number of “steps” (i.e., edges) needed to get from one node to another; a longer path length is indicative of a less interconnected and less efficient network (Miraglia et al., 2022; Rubinov and Sporns, 2010; Thilaga et al., 2018; Van Diessen et al., 2014). Small worldness describes networks that are characterized by high local clustering coefficients and short average path length between nodes (Bassett and Bullmore, 2006; Bassett and Bullmore, 2017; Rubinov and Sporns, 2010). Networks high in small worldness have fewer long-distance connections. This topology is generally considered an efficient structure for neural network processing. Global and local efficiency describe how effectively information is transferred throughout the whole network and in local regions, respectively (Achard and Bullmore, 2007; Rubinov and Sporns, 2010).

Of the 124 total papers in this review, 31 reported significant effects using methods distinct from the previously described categories. Examples of these other methods include variations of mutual information (e.g., cross-mutual information, weighted symbolic mutual information), amplitude envelope correlation, and directed transfer function, amongst others, some of which are based in correlational approaches. A collective summary of results from these studies would be overly speculative. Thus, we show the methodological approach of each study in Supplementary Table S3 and the results of each individual study in Table 4. Here we summarize only a few overarching points. In MCI, 14 of 17 studies (82%) using these other methods found significant group differences. Most studies found reduced connectivity in MCI, although not all studies reported the directionality of the effect, and directionality varied across the methods, regions, and frequency bands, particularly when task-related. Of those with significant findings, four studies were conducted in source space (29%, 4/14). Three found only lower connectivity in MCI compared to HC, and the fourth had mixed directionality depending on the alpha sub-band. In AD participants, 19 of 25 studies (76%) reported significant connectivity effects. The majority of the results indicated lower connectivity in AD compared to HC, but with variability by method, region, and frequency band. Five of the studies with significant findings were conducted in source space (26%, 5/19). Three found only lower connectivity in AD compared to HC; the other two had mixed directionality, dependent on region and metric of interest. Of note, a longitudinal study of MCI participants who later progressed to AD reported greater broadband connectivity (1–100 Hz, via mutual information between sensors) only during the prodromal (i.e., MCI) stage, which was no longer evident at the time of AD diagnosis or at three years post-diagnosis (Bonanni et al., 2021). This evidence of hyperconnectivity in the prodromal stage was theorized to be indicative of plasticity (i.e., recruitment) that subsides with greater disease progression.

While many EEG research questions may be best addressed using sensor-level data (e.g., event-related potentials from cortical regions), connectivity analyses involve a higher risk of spurious connectivity when conducted at the sensor level (Chiarion et al., 2023; Mahjoory et al., 2017; Michel and Brunet, 2019; Schoffelen and Gross, 2009; Van de Steen et al., 2019). For this reason, relying solely on findings from sensor-level studies may over-estimate connectivity results. Specifically, sensor-level results may carry a greater risk of detecting patterns that may not actually be attributable to shared neural coupling or co-activation. Connectivity analyses in source space are better equipped to delineate the likely neural generators of EEG signals and more appropriately account for underlying sources of shared variance, thus increasing the likelihood of accurately modeling connectivity between two distinct brain regions. Despite these important advantages, only 29% (i.e., 36/124) of the current studies analyzed connectivity in source space. Trends from sensor- and source-level studies in the present review had overall similar patterns. That is, there was no discernible difference in preponderance or direction of effects using source space rather than sensor-level data. Yet, the subset of studies available with source-level findings was too limited to in any way be conclusive. We advocate for future analyses in source space, which has become particularly accessible and feasible with advances in user-friendly, open-source software and detailed published tutorials (e.g., Delorme and Makeig, 2004; Oostenveld et al., 2011; Penny et al., 2011; Tadel et al., 2011). This does require thoughtful consideration of the appropriate computation of sources. For example, some studies analyzed 84 ROIs computed from only 19 electrodes; a low-density EEG array is not ideal for delineating a large number of sources (Michel and Brunet, 2019; Song et al., 2015). Relatedly, thoughtful consideration should be given to using a priori brain regions of interest, rather than extensive or exhaustive numbers of comparisons.

The majority of the papers reviewed here analyzed EEG connectivity during the resting state. Given the cognitive limitations inherent in studying MCI and AD, this tendency toward resting state analysis is understandable. However, the current findings are certainly biased toward documenting patterns within the default mode network. Future complementary study is needed using tasks that tap specific cognitive processes, with attention to high or comparable task accuracy across groups (at least in milder forms of cognitive impairment). High accuracy tasks discern neural activity patterns that are not conflated with group differences due to error or cognitive demand (Reuter-Lorenz and Cappell, 2008). Some examples include tasks that tap aspects of crystallized intelligence (e.g., vocabulary; Ferré et al., 2019; Salthouse, 2014), semantic access (Nielson et al., 2006; Pistono et al., 2019; Seidenberg et al., 2009; Sugarman et al., 2012), stop-signal tasks (Elverman et al., 2021; Hsieh and Lin, 2017a, 2017b; Paitel and Nielson, 2021, 2023), and oddball paradigms (Invitto et al., 2018; Iragui et al., 1993; Schiff et al., 2008; Stevens et al., 2000).

The most common approaches to analyzing connectivity used metrics that are based on EEG frequencies (e.g., coherence, phase lag index, etc.). While there are advantages to investigating connectivity within specific oscillatory bands, these approaches are limited in temporal resolution (Cohen, 2014; Luck, 2014). A primary strength of EEG is its precise recording of summated post-synaptic potentials with millisecond temporal resolution (Luck, 2014; Slotnick, 2017). However, most frequency domain approaches average activity across several seconds, thereby compromising that temporal precision. Such an approach is more appropriate for resting state analyses, in which participants are not completing a specific task. However, in the context of active task engagement, temporal precision provides crucial information regarding temporally specific neural network connectivity. In using longer time windows of several seconds to allow for higher frequency resolution, the ability to provide temporal precision about cognitive subprocesses is compromised. For example, Li et al. (2019) used a digit span task, but across epochs of 17 s, which was inclusive of two seconds pre-stimulus, the ten second duration of the stimulus presentation, and five seconds post-stimulus. Thus, rather than specifically analyzing neural activity relevant to the holding or manipulating of information in working memory, this lengthy window encompassed a broad number of cognitive processes underlying task performance.

The vast majority of the studies analyzed connectivity within a specific frequency band. Yet, oscillatory activity in the brain simultaneously occurs across multiple frequency ranges. For this reason, there has been a recent increase in prevalence of cross-frequency approaches that analyze the relationship between activity across different bands. Two categories of cross-frequency coupling are phase-amplitude and phase-phase coupling. Recent work with these methods points to their potential import for understanding the neuroscience of cognition (Abubaker et al., 2021; Canolty and Knight, 2010; Riddle et al., 2021), as well as for application in contexts such as MCI and AD (Dimitriadis et al., 2015; Musaeus et al., 2020). Of the studies in the present review that analyzed cross-frequency coupling, most found lower coupling in MCI or AD compared to HC, including theta-gamma coupling in MCI (Vanneste et al., 2021), and delta-theta, delta-alpha, delta-beta (Cai et al., 2018) and multiplex (across four bands and two bands) network features (Cai et al., 2020) in AD. In contrast, the only study during active task engagement found greater theta-gamma coupling for frequent (non-oddball) trials in MCI compared to HC during an olfactory oddball task (Sedghizadeh et al., 2022). Thus, while this area of research is young, it is likely that unique insights may be obtained by pursuing application in MCI and AD.

In contrast to the frequency domain, connectivity approaches in the time domain, such as those based in correlations, allow for shorter time windows, and thus superior temporal resolution. The primary limitation of such correlational methods is that most do not account for volume conduction. However, this is not a concern that is unique to correlational methods. For example, magnitude squared coherence is also sensitive to volume conduction (Bastos and Schoffelen, 2016; Khadem and Hossein-Zadeh, 2014; Ruiz-Gómez et al., 2019a; Srinivasan et al., 2007). Measures such as imaginary coherence have been specifically developed to better account for those effects (Nolte et al., 2004). Furthermore, with recent advances in computational processing, approaches that allow for evaluating temporal–spatial dynamics are becoming more feasible. For example, Crook-Rumsey et al. (2023) used a novel network modeling approach (i.e., spiking neural networks) during working memory and prospective memory tasks to interrogate connectivity within three separate time windows: 200-400 ms (cue detection and monitoring), 400-800 ms (deeper contextual and memory processes), and 0-1,000 ms (full epoch). They found lower connectivity in MCI compared to HC that was specific to the 200 to 400 ms window in a central cluster during a 1-back task, and in lateral frontal clusters during a perceptual prospective memory task. Importantly, these effects were not significant using the full 1,000 ms time window.