Brain health in diverse settings: How age, demographics and cognition shape brain function

Diversity in brain health is influenced by individual differences in demographics and cognition. However, most studies on brain health and diseases have typically controlled for these factors rather than explored their potential to predict brain signals. Here, we assessed the role of individual differences in demographics (age, sex, and education; n = 1298) and cognition ( n = 725) as predictors of different metrics usually used in case-control studies. These included power spectrum and aperiodic (1/f slope, knee, offset) metrics, as well as complexity (fractal dimension estimation, permutation entropy, Wiener entropy, spectral structure variability) and connectivity (graph-theoretic mutual information, conditional mutual information, organizational information) from the source space resting-state EEG activity in a diverse sample from the global south and north populations. Brain-phenotype models were computed using EEG metrics reflecting local activity (power spectrum and aperiodic components) and brain dynamics and interactions (complexity and graph-theoretic measures). Electrophysiological brain dynamics were modulated by individual differences despite the varied methods of data acquisition and assessments across multiple centers, indicating that results were unlikely to be accounted for by methodological discrepancies. Variations in brain signals were mainly influenced by age and cognition, while education and sex exhibited less importance. Power spectrum activity and graph-theoretic measures were the most sensitive in capturing individual differences. Older age, poorer cognition, and being male were associated with reduced alpha power, whereas older age and less education were associated with reduced network integration and segregation. Findings suggest that basic individual differences impact core metrics of brain function that are used in standard case-control studies. Considering individual variability and diversity in global settings would contribute to a more tailored understanding of brain function.

Recent findings underscore the significance of individual differences in demographic predictors, showing that brain-phenotype models primarily reflect these variables rather than the intended cognitive domains, particularly in non-stereotypical populations (Greene et al., 2022).Notwithstanding, most studies on brain health and aging (Jawinski et al., 2022;Yan et al., 2015), psychiatry (Wolfers et al., 2020;Elad et al., 2021), and neurodegeneration (Pinaya et al., 2021;Verdi et al., 2022) have typically controlled for these variables rather than investigating their impact on brain-phenotype associations.Incorporating the role of individual differences, including demographics and cognition, in functional brain dynamics across large, heterogeneous and underrepresented populations may provide a direct pathway to a better understanding of the impact of diversity on brain health.
The study of demographics and cognition as predictors of brain signals has predominantly relied on MRI or fMRI (Bethlehem et al., 2022;Elad et al., 2021;Di Biase et al., 2023;Rutherford et al., 2022;Habes et al., 2021;Rosenberg et al., 2020).A limited number of reports have incorporated MEG (Dimitriadis, 2022;Stier et al., 2023), and only one has used EEG (Hill et al., 2022).Age has consistently emerged as the most frequently studied demographic factor for predicting structural and functional changes in the human brain, spanning from post-conception to aging, both in healthy individuals (Bethlehem et al., 2022;Di Biase et al., 2023;Rutherford et al., 2022;Dimitriadis, 2022;Stier et al., 2023) and in patients with psychiatric (Elad et al., 2021) or neurodegenerative conditions (Bethlehem et al., 2022).While sex is considered a predictor of brain changes across the lifespan (Bethlehem   1 Equal contribution. H. Hernandez et al. et al., 2022;Rutherford et al., 2022;Stier et al., 2023), the impacts of other relevant variables, such as education and cognition (Rosenberg et al., 2020;Greene et al., 2020;Garo-Pascual et al., 2023;Cesnaite et al., 2023), have been less explored.Both education (Gordon et al., 2008;Jokinen et al., 2016) and cognition (Habes et al., 2021) have been associated with white matter changes and increased functional connectivity (Franzmeier et al., 2071a(Franzmeier et al., , 2017b)).However, while previous studies have investigated how demographic and cognitive factors influence brain signals, few have explored the effects of age, sex, education, and cognition simultaneously.This research gap highlights the need for studies that comprehensively explore the interplay between demographic factors, cognitive abilities, and brain signals using various imaging modalities, including EEG.These advancements would represent a crucial step towards robustly quantifying individual variations and their associations with signatures of brain functional organization or reorganization, especially in global research settings exhibiting larger heterogeneity.
Given the limited accessibility of neuroimaging or MEG measures on a large scale, more cost-effective and scalable techniques, such as EEG, may be better suited for studying individual differences in global settings.Neuroimaging measures often exhibit poor reliability and low power, requiring thousands of participants to demonstrate brainphenotype associations (Marek et al., 2022).The non-invasiveness, scalability, availability, temporal resolution, and low cost of EEG make it an attractive approach to meet this need (Prado et al., 2022).Although many studies have explored the association of EEG metrics with demographics and cognition, only one study have used these variables to predict EEG signals (Hill et al., 2022).Studies have identified age-related EEG changes in resting and task performance (Valsdóttir et al., 2022;Cesnaite et al., 2023;Al Zoubi et al., 2018;Carrier et al., 2001;Hinault et al., 2023;Javaid et al., 2022;Merkin et al., 2023;Meunier et al., 2009;Murty et al., 2020;Smith et al., 2023;Smits et al., 2016;Song et al., 2014;Stacey et al., 2021;Trammell et al., 2017;Trondle et al., 2023;Valdes-Sosa et al., 2021;Voytek et al., 2015;Zappasodi et al., 2015), including alpha amplitude, as well as slowdown and increases in global power (Cesnaite et al., 2023;Trondle et al., 2023;Gaal et al., 2010), topographic reorganization in delta and theta frequencies (Ishii et al., 2017;Rossini et al., 2007), and attenuated spontaneous gamma oscillations (Murty et al., 2020).Complexity changes across the lifespan with increases during young adulthood and decreases in the elderly population (Zappasodi et al., 2015) have been reported.Age also induces changes in the aperiodic components (Hinault et al., 2023;Trondle et al., 2023;Donoghue et al., 2020) and reductions in global and local efficiencies as well as small-worldness (Javaid et al., 2022;Meunier et al., 2009;Achard and Bullmore, 2007;Onoda and Yamaguchi, 2013).The association of sex and EEG metrics is less conclusive (Carrier et al., 2001;Pravitha et al., 2005;Miraglia et al., 2015).Conversely, the link between cognition and EEG signals is well-established, with theta (Finnigan and Robertson, 2011) and alpha (Lejko et al., 2020) associations, decrease in random and spontaneous neural activity (Smith et al., 2023;Ouyang et al., 2020;Pei et al., 2023), increased complexity (Smith et al., 2023;Ouyang et al., 2020;Pei et al., 2023), higher network integration (Finnigan and Robertson, 2011;Bullmore and Sporns, 2009;Iinuma et al., 2022;McBride et al., 2014), and small-world properties (Liao et al., 2017).The association between education and EEG metrics has been rarely reported, with limited evidence (Wilkinson et al., 2023).Furthermore, although EEG studies have improved preprocessing pipelines (Prado et al., 2022;Bigdely-Shamlo et al., 2015;Ballesteros et al., 2023;Prado et al., 2023), several methodological caveats continue to hinder their effective use in large-scale studies.The recording heterogeneity, different layouts of electrodes and amplifiers, lack of harmonization, disparate processing pipelines, and small sample sizes, have restricted broader application of EEG (Prado et al., 2022;Jovicich et al., 2019).Thus, there is a critical need for studies in large and diverse samples using demographic or cognitive variables to predict brain signals via EEG while addressing its sources of heterogeneity.
To address these gaps, this study explored EEG data from a large and diverse sample of healthy participants under resting-state conditions from regions in the global south (Argentina, Brazil, Colombia, Chile, Cuba) and the global north (Ireland, Italy, Turkey, United Kingdom).We aimed to validate the relationships between demographic (i.e., age, sex, and education; n = 1298) and cognitive factors (n = 725) with the EEG metrics across a diverse, heterogeneous, and global dataset.Additionally, we investigated their utility in predicting significant changes in brain function, even in the presence of large heterogeneity.We computed brain-phenotype models of local activity [i.e., power spectrum and aperiodic components (Hill et al., 2022;Varela et al., 2001)], and brain dynamics and interactions [i.e., complexity, and graph-theoretic measures (Bullmore and Sporns, 2009;Iinuma et al., 2022)].We selected these metrics based on their established relevance to demographic factors, their potential implications for understanding brain function, and their frequent use across different studies.These metrics have shown associations with the features assessed in our study, namely age (Cesnaite et al., 2023;Stacey et al., 2021;Gaal et al., 2010), sex (Carrier et al., 2001;Pravitha et al., 2005;Miraglia et al., 2015), education (Wilkinson et al., 2023), and cognition (Finnigan and Robertson, 2011;Bullmore and Sporns, 2009;Iinuma et al., 2022;McBride et al., 2014).Power spectrum metrics reveal dominant rhythms linked with cognitive processes and brain states (Klimesch, 1999), while spectral EEG metrics interpret non-rhythmic brain activity relevant to aging and cognition (Donoghue et al., 2020).Complexity metrics provide insights into the dynamics of different brain regions and their interactions, potentially serving as biomarkers for brain health disorders (Tononi et al., 1994;Lau et al., 2022).Graph-theoretic metrics illuminate functional connectivity patterns and network organization (Bullmore and Sporns, 2009).These metrics served as outcomes to evaluate the predictive value of each demographic and cognitive variable under two conditions: (a) when it served as the most robust predictor within the model, and (b) when another variable was considered the best predictor.Additionally, we explored the source space for each of brain-phenotype models.
We hypothesized that despite the sample diversity and heterogeneity, a subset of the demographic and cognitive variables would significantly predict each EEG metric.We expected that age (Rosenberg et al., 2020;Hill et al., 2022;Carrier et al., 2001;Javaid et al., 2022;Merkin et al., 2023;Meunier et al., 2009;Murty et al., 2020;Smith et al., 2023;Smits et al., 2016) and cognition (Carrier et al., 2001;Hinault et al., 2023;Smits et al., 2016;Song et al., 2014) would emerge as the most robust predictors of EEG metrics.We also anticipated that the combined effects of these two factors would result in models explaining a high proportion of variance of EEG measures.The influence of sex (Merkin et al., 2023;Stacey et al., 2021;Trammell et al., 2017) and education (Trondle et al., 2023) would be weaker.Our results suggest that individual differences in diverse settings impact core metrics of brain function that are used in standard case-control studies.The role of individual differences and diversity in contributing to brain function needs to be addressed more systematically in global contexts.

Participants
This multicentric study involved 1298 healthy adult participants (age: mean = 46.60,SD = 20.76,range 18-91 years;; years of formal education: mean = 13.77,SD = 4.38; sex: M = 597, F = 701) representing diverse populations (Fig. 1A) from the global south (Argentina, Brazil, Colombia, Chile, Cuba) and the global north (Ireland, Italy, Turkey, United Kingdom).Demographic characteristics and sample sizes for each country are provided in Table 1.Participants had no history of psychiatric and/or neurological disorders, alcohol/drug abuse, significant visual and/or auditory impairments.Furthermore, participants were not using any ferromagnetic implant.No participant reported subjective cognitive complaints or functional impairments.Data on the general cognitive state of 725 participants were available.The study protocol was approved by the Institutional Ethics Committee at each participating center, and all participants provided written informed consent in accordance with the Declaration of Helsinki.The data and analysis codes are freely available at the following GitHub link https ://github.com/euroladbrainlat/Brain-health-in-diverse-setting.
The data in the repository has been anonymized and pre-processed.
We used the G*Power 3.1.9.7 software (Faul et al., 2007) for power analysis, specifically employing an F-test for multiple linear regressions.Our study's sample size, comprising 725 subjects with cognitive measurements, surpassed the required minimum for conducting multiple linear regressions with four variables.This allows for the detection of small effect sizes (0.02 (Cohen, 1988)) with a statistical power of 0.88.Additionally, the larger total sample of 1298 subjects was adequate for performing multiple linear regressions with three variables, effectively identifying small effect sizes with a statistical power of 0.99.

Demographics
The demographic information includes age at the time of assessment (in years), sex (male or female), and years of formal education.Data on gender identity was not available.

Cognition
The general cognitive state was evaluated using the raw total score from the Mini-Mental State Examination (MMSE) (Folstein et al., 1975).This test is a widely recognized screening tool for cognitive impairment and assessing general cognitive functioning.It evaluates various cognitive domains, including orientation to time and place, short-term memory recall, working memory, language abilities, visuoconstructional skills, and basic motor commands.This instrument consists of items, and each correct answer adds one point to the total score, which ranges from 0 to 30.A score of 24 or more indicates normal cognitive functioning (Creavin et al., 2016;Mukaetova-Ladinska et al., 2022;Cebi et al., 2020;Foderaro et al., 2022;Kochhann et al., 2010).The dataset comprised a total of 1298 participants, with cognitive data available for 725 individuals.Regression models tuned with cross validation and data partition were developed using years of education, age, sex, and cognitive state as predictors.(C) EEG data preprocessing: The preprocessing steps included electrode re-referencing, noise removal through signal filtering, resampling to Hz, artifact removal (e.g., blinks and eye movements), and transformation into a common source space using the Automated Anatomical Labeling (AAL) atlas.(D) Outcomes: EEG data in the source space was used to compute four groups of metrics, categorized into power spectrum, aperiodic, complexity, and connectivity metrics.(E) Data analysis: Linear regressions were applied, utilizing data partitioning with an 80% training sample and a 20% testing set, cross-validated (k = repetitions).We reported the importance of predictors and the results on the influence of each predictor on specific brain regions.

EEG data acquisition, processing, and harmonization
The EEG acquisition parameters at each center are detailed in Supplementary Table 1.Participants were situated in comfortable chairs within dimly lit, electromagnetically quiet rooms, and were advised to stay still and alert.Resting-state EEG (rsEEG) with closed eyes was recorded using various systems, encompassing distinct sensor types, calibrations, and electrode configurations (Supplementary Table 1).In some instances, EEGs with open eyes were also captured.Due to differing recording durations across centers, EEG analyses were limited to the first five minutes of each recording.
The EEG data was processed offline using a customized, automated, and validated pipeline that incorporated harmonization protocols specifically designed to mitigate batch effects and methodological variations in multi-center EEG studies (Ballesteros et al., 2023;Prado et al., 2023).The processing workflow involved several steps, including data preprocessing, EEG rescaling, spatial normalizations, and EEG source localization.

Pre-processing
Raw EEG data was filtered between 0.5 and 40 Hz using an 8th order zero-phase shift Butterworth filter.Data were re-sampled to 512 Hz and referenced using the reference electrode standardization technique (REST) (Hu et al., 2018).Blinking, ocular artifacts and myogenic activity were corrected using two distinct ICA techniques (Delorme and Makeig, 2004): ICLabel (a tool for the classification of EEG independent components into signals and different categories of noise) (Pion-Tonachini et al., 2019) and EyeCatch (a tool for identifying eye-related ICA scalp maps) (Bigdely-Shamlo et al., 2013).ICA techniques [Infomax ICA (Górecka and Walerjan, 2011)] were exclusively used to remove ocular artifacts (blinking and eye movements).These artifacts typically fall among the first components of ICA.After ensuring artifact correction through ICA techniques (specifically, ICLabel and EyeCatch), we conducted a visual inspection of the EEG.Malfunctioning channels were replaced using weighted spherical interpolations (Kothe and Makeig, 2013).

Normalization
The normalization process adhered to multicentric studies guidelines (Prado et al., 2022).To minimize variability across centers, Z-score transformations of EEG time series were implemented (Ballesteros et al., 2023;Prado et al., 2023aPrado et al., , 2023b)).The normalization consisted of computing the mean voltage of each EEG channel and Z-transforming the corresponding voltage samples, considering the mean and the standard deviation of the distribution.This normalization reduces the electrode-by-electrode variability and was conducted independently for each recruitment center, therefore reducing the inter-site variability.Furthermore, a spatial normalization of EEG was performed by combining different electrode configurations, creating virtual electrodes calculated from topographic interpolation transforms.This method has been effectively utilized in studies investigating the variance between acquisition systems in contrast to between-subject and between-session variances (Melnik et al., 2017).The technique projects electrode positions onto a mesh-head model consisting of 1082 points and interpolates EEG activity.We adjusted the EEGLAB headplot (Prado et al., 2023) function to map the original EEG onto a 6067-point mesh-head model (Kothe and Makeig, 2013;Melnik et al., 2017).

EEG source estimation
EEG source generators were estimated using the standardized lowresolution electromagnetic tomography (sLORETA) method (Pascual--Marqui, 2002).This method estimates the standardized current density at specific virtual sensors located in the cortical gray matter and the hippocampus of an average brain (MNI 305,Brain Imaging Centre,Montreal Neurologic Institute).This estimation is based on a linear, weighted summation of a unique scalp voltage configuration or the EEG cross-spectrum at the sensor level.Essentially, sLORETA serves as a distributed EEG inverse solution technique that extends from a standardized version of minimum norm current density estimation.It effectively addresses challenges associated with estimating deep sources of EEG activity and ensures precise localization, even in cases with significant correlation among nearby generators (Asadzadeh et al., 2020).
Electrode layouts were aligned with the MNI152 scalp coordinates (Mazziotta et al., 2001).When computing the sLORETA transformation matrix, a signal-to-noise ratio of 1 was selected as the regularization method.Standardized current density maps were produced using a three concentric spheres head model, in a predefined source space of 6242 voxels (with a voxel size of 5 × 5 × 5 mm) of the MNI average brain.The brain was segmented into 82 brain regions using the Automated Anatomical Labeling (AAL) atlas (Rolls et al., 2015).Standarized current densities were estimated for each of the 153,600 Vage distributions comprising the five-minutes of rsEEG (sampled at 512 Hz).Standarized current density time series estimated in voxels belonging to the same AAL regions were averaged, leading to a mean time series for each brain area (Prado et al., 2023;Cruzat et al., 2023;Herzog et al., 2022).

EEG descriptors
We computed four categories of EEG descriptors: (i) power spectrum and (ii) aperiodic spectral metrics, (iii) complexity, and (iv) graphtheoretic measures.These categories reflect neural activity (i.e., power spectrum or aperiodic spectral metrics) (Cesnaite et al., 2023;Donoghue et al., 2020), and brain dynamics and interactions (i.e., complexity, and graph-theoretic measures) (Zappasodi et al., 2015;Vecchio et al., 2022).Power spectrum metrics are relevant in EEG data analysis, offering insights into brain activity by characterizing how signal power is distributed across various frequencies (Wang et al., 2015).These metrics establish a baseline for typical spectral activity, serving as a valuable benchmark for identifying pathological conditions (Buzsáki, 2006) The aperiodic spectral components of EEG represent non-rhythmic brain activity.These components facilitate physiological interpretations related to aging and cognition (Donoghue et al., 2020).For example, these aperiodic metrics has been linked to excitation/inhibition balance (Medel et al., 2023;Martinez-Canada et al., 2023), and electrophysiological noise (Voytek et al., 2015).Complexity metrics enable the evaluation of dynamical brain complexity, which have been explored as potential biomarkers for diagnosing mental health disorders (Tononi et al., 1994;Lau et al., 2022).Lastly, graph-theoretic metrics shed light on the structure of functional connectivity patterns and network organization (Bullmore and Sporns, 2009).

Power spectrum metrics
Power spectrum analyses were conducted in both the canonical EEG frequency bands and the subject-specific EEG frequency bands (Babiloni et al., 2020).Canonical bands provide standardized ranges ensuring consistency and comparability across studies.However, these bands may not effectively capture individual differences (Doppelmayr et al., 1998;Bazanova and Vernon, 2014).Subject-specific bands enhance the sensitivity to detect specific physiological, cognitive, or behavioral states that might be masked when using generic bands (Klimesch, 1999).For instance, the individual alpha frequency peak (IAF) increases during childhood and slows in middle and older age (Turner et al., 2023), decreases with neurodegeneration (Moretti et al., 2004), and increases with cognitive demands (Haegens et al., 2014).Analyzing power with canonical and subject-specific frequency bands allows for a comprehensive assessment of EEG data.
Subject-specific frequency bands were determined using the IAF, defined as the frequency with maximum power in the alpha-frequency band, and the theta/alpha-frequency transition (TF), defined as the frequency with minimum power in the second half of theta frequency range (Martinez-Canada et al., 2023;Babiloni et al., 2020;Ince et al., 2017).The subject-specific frequency bands are defined as: δ (TF-4 to TF-2), θ (TF-2 to TF), α low (TF to IAF), and α high (IAF to IAF+2) (Babiloni et al., 2020).The β and γ frequency bands corresponded to the canonical division.
We computed both the power spectral density (PSD) and the normalized PSD (nPSD) using Welch's method with 1-second Hanning windows with 50% overlap.Subsequently, we calculated the mean of the nPSD in each frequency band (equivalently percent power) (Li et al., 2007) and the percent of the nPSD of a given frequency band relative to the total nPSD (relative power density) (Wang et al., 2015;Babiloni et al., 2020).

Aperiodic spectral metrics
We computed the aperiodic components of the EEG power spectral density (PSD): the 1/f slope (a), knee (k), and offset (b) (Martinez-Canada et al., 2023;Pathania et al., 2021;van Nifterick et al., 2023).We applied the fitting oscillations and one-over-F (FOOOF) algorithm to the PSD, covering the frequency band from 0.5 to 40 Hz, peak width limits from 1 to 6 Hz, maximum number of peaks of 6, minimum peak height of 0.2, and peak threshold of 2.0 (Martinez-Canada et al., 2023;Pathania et al., 2021;van Nifterick et al., 2023).The FOOOF algorithm was designed to model the aperiodic component of the PSD (Donoghue et al., 2020), using a Lorentzian function as follows: where A is the aperiodic component, b is the offset, x is the exponent, and x=-a.

Complexity metrics
Complexity descriptors of the EEG source space provide insights into the dynamics of different brain regions and their interactions.A high complexity value in an EEG signal indicates an increased degree of irregularity, indicating that the EEG is less predictable.This is often associated with a healthy and alert brain state (Medel et al., 2023).Conversely, a low comlexity values suggests a more regular or repetitive signal, could be observed in pathological states (Sun et al., 2020) anaesthesia or deep sleep (Sarasso et al., 2021;Boncompte et al., 2021).Fractal dimension (FD), permutation entropy (PE), Wiener Entropy (WE), and spectral structure variability (SSV) (Sun et al., 2020;Burns and Rajan, 2015) were computed to assess the complexity of EEG signals, with code sourced from a GitHub repository (https://github.com/tfburns/MATLAB-functions-for-complexity-measures-of-one-dimensiona l-signals) (Burns and Rajan, 2016).We include measures of predictability (FD) as well as regularity (PE, WE, SSV), which represent the two key aspects of brain dynamics (Lau et al., 2022).Predictability refers to the ability to anticipate the temporal evolution of the system's states, while regularity measures the frequency and pattern of repetitions in the system's trajectory (Lau et al., 2022).Changes in FD are linked to variations (Zappasodi et al., 2015) and cognition (Smits et al., 2016;Hemmati et al., 2013).Similarly, both PE and WE have been associated with individual differences in age (Al Zoubi et al., 2018;Shumbayawonda et al., 2018) and cognition (Parbat and Chakraborty, 2021;Seker et al., 2021).These metrics collectively capture broad conceptions of complexity (Burns and Rajan, 2015).The equations for complexity metrics are described below.
Fractional dimension: The FD is a statistic index of complexity details in a patter and can be expresed as: where a, k y NLD 0 , was set to the suggested parameters values: 1.9079, 0.18383 y 0.097178, respectively (Kalauzi et al., 2009).The NLD was calculate as: where y n (i) represents the ith signal sample after amplitude normalization.

Permutation entropy:
The PE is an ordinal-based non-parametric metric of the temporal dependence structure in linear or non-linear time series.Therefore, PE can be expressed as follows: Considering a time series represented by x t , con t = 1, …, T, and the embedded vector where n is the embedding dimension and l is the lag.Then, the vector X t is arranged from smallest to largest.Finally, PE was defined as: where p(π) = f(π)/(T − (n − 1)l) y f(π) represents the frequency of the symbols of length n derived from the ordinal relationships between Xt. (Ouyang et al., 2013) Wiener entropy and spectral structure variation: The WE, alternatively referred to as spectral flatness or tonality coefficient, signifies the uniform distribution of signal energy across the frequency domain.This metric provides a quantifiable measure of how closely the signal mirrors a sinusoidal function, as opposed to exhibiting characteristics reminiscent of noise.The spectral flatness measure (SFM) was used to quantify spectral structure (Singh, 2011), and the equation for its calculation is as follows: where N is the number of points in the Fourier transform, and S(t, f) is the associated power at each frequency component.Finally, the Wiener entropy is calculated as the average of SFM(t), and the spectral structure variability is calculated as the variance of SFM(t).

Connectivity graph metrics
Calculating graph metrics required determining functional connectivity across the entire space of different brain regions.Functional connectivity was assessed in the time-domain using information theoretic measures, which were calculated with custom Matlab codes (Prado et al., 2023;Herzog et al., 2022).These measures were derived using the Gaussian copulas approximation (Ince et al., 2017), a robust computational framework that combines copulas statistical theory with an analytic solution for the entropy of multivariate Gaussian distributions.The incorporation of Gaussian copulas-based methods strengthens the robustness and cost-effectiveness of our methodology, resulting in a reduction of computational time required for functional connectivity calculations.We computed both pairwise and high-order interactions metrics, specifically mutual information (MI) (Prado et al., 2023;Herzog et al., 2022;Santamaria-Garcia et al., 2023), conditional mutual information (CMI) (Prado et al., 2023;Ince et al., 2017), and organizational information (O_info) (Prado et al., 2023) metrics (Supplementary Data S1).MI gauges the shared information between two random variables (or time series), while CMI assesses the shared information between two random variables (or time series) with respect to a specified third variable.Finally, O_info, an extension of Shannon's mutual information, enhances our understanding of the crucial characteristics of multivariate systems, especially those involving high-order interactions (Herzog et al., 2022).Information theory metrics surpass traditional measures like coherence or phase locking value because they can capture non-linear interactions (Imperatori et al., 2019;Kida et al., 2016) and provide better results(109).This is crucial for exploring brain dynamics, where complex non-linear interactions among neural regions shape patterns of functional connectivity (Imperatori et al., 2019;Kida et al., 2016).
Measures of segregation, integration, and global metrics were included to measure complementary dimensions of network topology (Bullmore and Sporns, 2009;Rubinov and Sporns, 2010).Global Efficiency provides insights into the efficiency of information exchange across the entire network, crucial for global information processing (Rubinov and Sporns, 2010).Transitivity measures the prevalence of clustered connectivity, reflecting localized, interconnected communities key for local processing and network segregation (Rubinov and Sporns, 2010).Small-worldness captures the balance between random and regular networks, indicating an optimal organization of localized and distributed processing, and integration and segregation, enhancing network efficiency and robustness (Bassett and Bullmore, 2017).Lastly, density quantifies the overall connectivity of the network, offering a direct assessment of the complexity and connectivity level (Rubinov and Sporns, 2010).All these measures have been shown to be sensitive to individual differences in age (Javaid et al., 2022;Meunier et al., 2009;Achard and Bullmore, 2007;Onoda and Yamaguchi, 2013), sex (Miraglia et al., 2015) and cognition (Finnigan and Robertson, 2011;Bullmore and Sporns, 2009;Iinuma et al., 2022;McBride et al., 2014;Liao et al., 2017) .The equations of these metrics are described below.
Connection weight: The connection weight expresses the strength with which nodes are linked, where w ij is the connection weight between node i and node j.
Weighted shortest path length: The shortest path length is a fundamental metric to evaluate integration, which is defined as: where f serves as a mapping from weight to length, and gi ↔ j represents the shortest weighted path between i and j.
Weighted characteristic path length: The characteristic path length is a measure employed to evaluate communication efficiency in a weighted network.The equation for its calculation is as follows: where n represents the number of nodes in the network, and d ij is the length of the shortest path between node i and node j.

Number of triangles:
The number of triangles is a fundamental metric to evaluate segregation, which is defined as: where t i is the weighted geometric mean of triangles around i, j and h are indices representing the neighboring nodes of i in the weighted matrix.
Weighted degree: The weighted degree is a measure employed to evaluate the importance or connectivity of a node in a weighted network.This metric can be calculated as follows: where k i is the degree of a node i.
Weighted clustering coefficient: The clustering coefficient is a metric used to evaluate the tendency of nodes in a network to form local groupings or clusters.The equation is as follows where t i is the number of triangles around i and k i is the weighted degree of a node i.
Weighted global efficiency: Global efficiency was employed to assess integration, crucial in quantifying the effective sharing of information across the brain, essential for coordinated cognitive activity.This metric evaluates the network's ability to exchange information globally (Yu et al., 2021;Stanley et al., 2015).The equation for its calculation is as follows.
where n represents the number of nodes in the network, and d ij is the length of the shortest path between node i and node j.
Weighted transitivity: Transitivity was selected to measure segregation due to its reflection of local interconnectivity, indicating the H. Hernandez et al. tendency of nodes to form closed clusters or triads.It serves as an indicator of local connectivity (Rubinov and Sporns, 2010) and can be expressed as follow.
where t i is the number of triangles around i and k i is the weighted degree of a node i.
Weighted density: Density served as global measure, indicating the overall connectedness of the network addressing complexity and resilience of brain functioning (Rubinov and Sporns, 2010).The equation for its calculation is as follows.
where w ij represents the weight of the connection between node i and node j, and n is the total number of nodes in the network.
Weighted small-worldness: Small-worldness was utilized to evaluate both integration and segregation, capturing the optimal balance between specialized processing within clusters and the global integration essential for comprehensive brain functionality.A network exhibits small-world characteristics if it displays a high clustering coefficient (reflecting segregation) and a short characteristic path length (indicating integration) (Rubinov and Sporns, 2010;Bassett and Bullmore, 2017;Yu et al., 2021).
The commonly used measure to quantify small-worldness in a network includes the clustering coefficient (C) and the characteristic path length (L).The formula for small-worldness is expressed as the ratio of the observed clustering coefficient in the network (C) to the expected clustering coefficient in a random graph with the same number of nodes and edges (C r ), normalized by the observed characteristic path length in the network (L) and the expected characteristic path length in the random graph (L r ).
A crucial difference between transitivity and the clustering coefficient is found in the normalization procedure.Clustering is normalized individually for each node, while transitivity is normalized across the entire node set (Rubinov and Sporns, 2010).

Simplification of the EEG analytical domain
To streamline the analysis of spectral and complexity results from EEG data, we adopted two complementary approaches.Initially, anatomically and functionally related AAL regions were merged to form more consolidated regions of interest (ROIs), as detailed in Supplementary Table 3. Subsequently, we refined the EEG analytical space by applying specific statistical criteria.The methodology for these processes is elaborated in the subsequent sections.

Region selection for analysis
We combined regions from the AAL atlas to generate ten cohesive ROIs.This approach involved two critical criteria: (i) grouping together brain regions associated with a specific cortical gyrus (e.g., superior, middle, and inferior orbital gyri) into a single ROI, and (ii) assembling neighboring regions with established functional coupling, such as the Rolandic operculum and insula.The proposed method restructures EEG analysis by concentrating on ROIs that include both structurally and functionally related regions, rather than analyzing each small region in isolation.We utilized a mean averaging approach, calculating the average value of the metrics derived from the signals of the regions within the AAL atlas that are associated with each ROI.This simplification enhances data interpretation and facilitates the identification of patterns or significant features in brain activity.

Statistical reduction of spectral and complexity features
To refine the analysis of spectral and complexity metrics, we applied statistical criteria.Specifically, ROIs demonstrating statistically significant relative power density and equivalent percent power (Wang et al., 2015;Jeong et al., 2021;Moretti et al., 2004) were identified through mean-vs-zero non-parametric permutation tests (α = 0.05; 5000 randomizations (Manly, 1997)).This refinement, applied to both canonical and individual EEG band classifications (Babiloni et al., 2020), was carried out for each frequency band.The results were then adjusted for multiple comparisons using the Benjamini and Hochberg False Discovery Rate (FDR) method.ROIs representing the Individual Alpha Frequency (IAF) (Klimesch, 1999;Moretti et al., 2004) were those displaying statistically significant α activities.Similarly, ROIs indicative of the θ-α transition (TF) (Klimesch, 1999;Moretti et al., 2004) were those with statistically significant θ activity.Moreover, the analytical space of aperiodic (Martinez-Canada et al., 2023;Pathania et al., 2021;van Nifterick et al., 2023) and complexity (Sun et al., 2020;Burns and Rajan, 2015) metrics was further reduced by employing mean-vs-zero non-parametric permutation tests, followed by correction using the Benjamini and Hochberg FDR method.

Multiple linear regression models
We used multiple linear regression models to understand the relationship between predictors and outcome variables (Bishop, 2006).The regression models were constructed using the four predictors outlined in Section 2.2.We constructed an individual regression model for each ROI.Regression models were generated and categorized based on the highest-rated predictors.This approach allowed us to separate the results into three distinct groups according to the top-rated predictor.The first group prioritized age as the top-rated predictor, the second emphasized education, and the third focused on cognition.None of the models identified gender as the most important predictor.The mathematical representation of the linear regression model is: where Y is the outcome variable, B 0 is the intercept, B = [B 1 , B 2 , …, B n ] is the coefficients vector of the independent variables matrix X = [x 1 , x 2 , …, x n ] and ϵ represents the error term.The coefficients B 1 , B 2 , …, B n are determined using the least squares method, which aims to minimize the sum of the squared differences (errors) between the observed values (actual values) and the values predicted by the model.The formulas to calculate the coefficients are: where X T is the transpose of matrix of vector X .
The EEG parameters obtained from the chosen ROIs (see Supplementary Table 3) and the connectivity outcomes estimated in wholebrain analyses were utilized as inputs for the multiple linear regression models described above.

Data partition
Models were trained on a training sample (80%) and tested in a testing set (20%), with k = 10 folds (Muller and Guido, 2018).For each iteration, we computed the estimation coefficients for the predictors, R-squared, Cohen's f 2 , Fisher's F of the model, and the model's significance.We reported the mean estimation values for each predictor along with their standard deviation, average R-squared and Cohen's f 2 (Selya et al., 2012), and overall Fisher's F. To determine the overall model significance, we combined the model's p-values obtained in each iteration using the Fisher method (Fisher, 1992).
H. Hernandez et al.

Statistical criteria for the reduction of feature space for frequency and complexity metrics
A further reduction of the analytical space was conducted using statistical criteria.To this end, ROIs with statistically significant relative power density and equivalent percent power (Wang et al., 2015;Moretti et al., 2004;Jeong et al., 2021) were selected by implementing mean-vs-zero non-parametric permutation tests (α = 0.05; 5000 randomizations (Manly, 1997)).This analysis was conducted for each frequency band, using both canonical and individual (Babiloni et al., 2020) EEG band classifications.Results were corrected for multiple comparisons using the Benjamini and Hochberg FDR method (Benjamini and Hochberg, 1995).Representative ROIs for IAF (Klimesch, 1999;Moretti et al., 2004) were those for which θ and α activities were statistically significant.Likewise, ROIs for TF (Klimesch, 1999;Moretti et al., 2004) were those for which θ activity was statistically significant.Furthermore, mean-vs-zero non-parametric permutation tests followed by Benjamini and Hochberg FDR was the method selected for further reducing the analytical space of the aperiodic (Martinez-Canada et al., 2023;Pathania et al., 2021;van Nifterick et al., 2023) and complexity (Sun et al., 2020;Burns and Rajan, 2015) metrics.

Quality of the signal
We conducted additional analyses to confirm that the number of electrodes does not impact signal quality and does not interfere with the observed effects.We calculated the Overall Data Quality (OQD) index in the source space, using the methodology proposed by Zhao et al. (2023), and conducted a linear regression to verify that the number of channels cannot predict signal quality.We used the number of channels as a predictor and the quality of the final signals as the outcome.The method for calculating OQD segmented into 1-second epochs, each labeled as 1 for low-quality epochs or 0 for high-quality epochs.The OQD represents the percentage of EEG epochs with good quality, ranging from 0 for signals where all epochs were classified as low quality, to 100 for signals where all epochs were classified as high quality (Zhao et al., 2023).

Results
For each type of EEG descriptor (spectral, complexity, and connectivity EEG outcomes), we categorized the regression models based on their top-rated predictors, resulting in three distinct groups according to the best-evaluated predictor.The first group prioritized age as the bestevaluated predictor, the second group emphasized education, and the third group centered on cognition.None of the models identified sex as the most important predictor.For each group of models and sets of EEG outcomes, we reported the three models with the highest R 2 values.Additionally, we reported the adjusted R 2 .In the Supplementary Material (Supplementary Figures 1-3), we provided an assessment of the model fit quality using Q-Q plots and residual vs. fitted values plots.In most cases, the results for the total sample and the subsample with available cognitive data were similar; therefore, we report the results for the latter.Further details on the results for the total sample are provided in Supplementary Tables 1 to 7.
Some participants had portions of the recording with their eyes open (Supplementary Table 1).Therefore, we analyzed data only from participants who underwent eye-close rsEEG exclusively.These results were comparable to those reported with the entire sample, and are detailed in Supplementary Tables 8 -11.In addition, we checked for collinearity among the predictors (see Supplementary S2 and Supplementary Figure 4) and conducted supplementary analyses using Ridge regressions (Supplementary Tables 12-15), which is a suitable method for predictors exhibiting collinearity (Tsigler and Bartlett, 2024).The results obtained with Ridge were comparable to the previous results.
We conducted additional analyses to ensure that variations in the number of channels across recruitment centers did not influence the results.We applied a linear regression to check if the number of channels predicts signal quality in the source space.The results showed that the model was not significant (Supplementary Table 16).Finally, we conducted an analysis of the correlation across the outcomes reported in the main results (see Supplementary S3, Supplementary Figure 5, and Supplementary Tables 17-20) to assess the effect of their correlation on the reported effects.

Age, education and cognition as the best predictors of power spectrum metrics
Statistical descriptors for all the regression models reported here are shown in Table 2.
The two regression models predicting age with the highest R 2 values had subject-specific α low equivalent power as the EEG outcome measure.
The third one had cannonical γ equivalent power.The first model reached the highest values in the left occipital region, the second in the left parietal region, and the third in the left orbitofrontal region (Table 2A and Fig. 2A).For all three models, the most important and significant predictors were age and cognition.In the model for the left occipital region, sex was the third most important predictor.For the other two models, education was the third significant predictor.When analyzing the total sample, the results for the three models remained significant, although R 2 values decreased (Fig. 2A, and Supplementary Table 4A).
For education as the best rated predictor, the three models with the highest R 2 values had IAF as the outcome measure.The model with best scores corresponded with the right medial frontal gyrus, and revealed education, cognition, age, and sex as significant predictors (Fig. 2B and Table 2B).In the second and third models, the highest values were reached in the left orbitofrontal cortex and the left inferior frontal gyrus, respectively, education and age were the significant predictors displaying similar values.Sex and cognition were significant in the best model, but not in the other two.In the total sample, R 2 values were low, but the models remained statistically significant (Fig. 2B and Supplementary Table 4B).
When cognition was the best-evaluted predictor, the two best models estimated canon α 2 relative power metrics more accurately.The third one performed better in estimating subject-specific αlow equivalent power.The first two models were in the left occipital regions and the third one in the right parietal region (Fig. 2C and Table 2C).For the three models, cognition and age emerged as the most important and significant predictors.Sex reached significance in the three models, but with lower estimates.Education was not statistically significant in any model.In the total sample, the models were also statistically significant, but showing a decrease in the R 2 values (Fig. 2C and Supplementary Table 4C).

Age and cognition as the best predictors of aperiodic spectral metrics
Statistical values for all models reported here are shown in Table 3.
For age as the top predictor, the three best-evaluated models had slope as the outcome metric.These models were placed in the left temporal region, the right temporal region, and the left hippocampus, respectively (Fig. 3A and Table 3A).In all three models, age and cognition emerged as the most important and significant predictors.In the first model, sex also reached significance, but not in the second and the third one.Education was not statistically significant in any model.In the total sample, the models were also statistically significant, but with lower R 2 values (see Fig. 3A and Supplementary Table 5A).
The best model with cognition as the best-evaluated predictor, estimated slope metrics more accurately.The second and the third models performed better in estimating offset metrics.The best model of cognition was in the left occipital region, the second in the left temporal region and the third in the left hippocampus (Fig. 3C and Table 3B).It is worth noting that, although reaching significance, all these models H. Hernandez et al. showed R 2 values lower than 0.1.In the three models, cognition and age were the most important and significant predictors.Sex and education did not reach significance in any model.In the total sample, the models were also statistically significant, but with lower R 2 values (Fig. 3C and Supplementary Table 5B).
The models in which education was the best-rated predictor were not statistically significant (Fig. 3B).

Cognition as the best predictor of complexity metrics
Statistical values for all models reported here are shown in Table 4.There were no models where age, education or sex were identified as the most highly valued predictors.
For cognition as the top predictor, the FD and the WE metrics in the whole brain were the main outcomes in the two best models.The third model performed better in spectral structure variation.In all the models, cognition showed much higher estimates compared to other predictors.For the best two models, age, and sex were also significant predictors.In the third model, sex was significant.Education did not reach significance in any model (Table 4A and Fig. 4C).In the total sample, the models were also statistically significant but with an important decrease in R 2 values (Fig. 3C and Supplementary Table 6A).

Age and education as the best predictors of graph-theoretic measures
Statistical values for all models reported here are shown in Table 5.When age was the top predictor, the best model estimated transitivity in the CMI matrix more accurately.The second and the third ones performed better in estimating global efficiency and small-worldness in the MI and CMI matrices, respectively.For all the three models, age and education were the most important and significant predictors.In the first model, cognition was also significant.Sex did not reach significance in any model (Fig. 4A and Table 5A).In the total sample, the models were also statistically significant with almost the same performance as in the subsample (Fig. 4A and Supplementary Table 7A).
For education as the top predictor, the best model had transitivity as outcome, the second one global efficiency and the third one smallworldness, all of them in the organizational information.The three models revealed education and age as the most important and significant predictors.Cognition and sex were not statistically significant in any model (Table 5B and Fig. 4B).The models were also statistically significant in the total sample (Fig. 4B and Supplementary Table 7B).
There were no models where cognition was identified as the most highly valued predictor.

Discussion
This study aimed to characterize EEG-derived endophenotypes based on demographic and cognitive factors across a diverse sample of participants.Despite multimodal diversity and heterogeneity (diverse populations, data acquisition, multicentric assessments, amplifiers, number and type of electrodes), individual differences shaped electrophysiological brain dynamics.Age emerged as the most robust and systematic predictor of EEG signals, followed by cognition.Education and sex were less influential predictors.Power spectrum activity and graph-theoretic measures were the most sensitive in capturing individual differences.Results are relevant for better understanding the individual differences that lead to diversity.The use of more affordable and scalable measures, such as EEG metrics, could be instrumental in creating future brain charts.The present findings may challenge traditional interpretations of case-control differences in brain signatures, emphasizing the influence of demographic and cognitive factors in brain-phenotype associations.
Age was associated with changes power spectrum and aperiodic spectral activity and less integrated and segregated networks.Age predicted a decrease in alpha power at parieto-occipital regions (Trondle et al., 2023;Donoghue et al., 2020), while the opposite direction was observed for gamma power in orbitofrontal regions (Rempe et al., 2023;Hunt et al., 2019).This pattern of results could be explained by the gradual loss of cholinergic function in the basal forebrain with age  (Schreckenberger et al., 2004) which is associated with diminished cholinergic input in the thalamus (Hindriks and van Putten, 2013;Pfurtscheller and Lopes da Silva, 1999), leading to decreased power in alpha oscillations (Trondle et al., 2023;Schliebs and Arendt, 2011).Conversely, increased gamma in the orbitofrontal cortex suggests a potential compensatory mechanism associated with age-related changes in frontal lobes, regions susceptible to developmental and aging processes (Rempe et al., 2023).The decrease in the aperiodic slope with increasing age aligns with previous results (Hill et al., 2022;Merkin et al., 2023;Trondle et al., 2023) and supports the proposed association between older ages and increased asynchronous background neuronal firing (Trondle et al., 2023;Donoghue et al., 2020).As age advances, there is an increase in random and spontaneous neural activity (Cremer and Zeef, 1987) and a decrease in the signal-to-noise ratio (Voytek et al., 2015).This reduction in signal-to-noise may stem from heightened spontaneous/baseline neural spiking activity (Cremer and Zeef, 1987), disrupting neural communication fidelity and potentially contributing to typical age-related cognitive decline (Voytek et al., 2015).Furthermore, the findings of age-related decreases in alpha power and a flattened aperiodic slope are highly compatible.These two phenomena may reflect aspects of the same neurobiological changes (Trondle et al., 2023).Decreased thalamic inhibitory control over cortical areas, due to impaired cholinergic input, is reflected in diminished cortical alpha

Fig. 2. Power spectrum metrics results. (A)
The three best models, where age was the best-evaluated feature, estimated low-frequency metrics more accurately.Age and cognition emerged as the most significant predictors with the highest evaluation.While sex was statistically significant, its impact in the models was minimal.When analyzing the total sample, R 2 values decreased, but the predictors remained significant.(B) The three best models, where education was the bestevaluated feature, exhibited low R 2 values but maintained statistical significance.These models IAF metrics more accurately.Both education and age demonstrated significance and displayed similar values in the models.Sex and cognition were significant in the best model.In the total sample, R 2 was very low, but the models remained statistically significant.(C) The three best models, where cognition was the best-evaluated feature, had a similar structure to panel A, but they estimated high and canonical frequencies more accurately.Cognition was the best-evaluated feature, but age reached almost the same value.Education did not reach statistical significance.For the total sample, there was a significant decrease in R 2 .The brains in the right column represent the intensity of predictors in the respective brain regions for statistically significant models, for the subsample with cognitive data.The models in occipital regions were the most significant ones, with parietal regions also demonstrating significant models.Some additional regions in the left hemisphere emerged, including the inferior, middle and orbital frontal gyri.The statistical values for the models with the complete sample are provided within each box.power, leading to higher cortical excitability, an increased excitation-to-inhibition ratio, and a higher level of neural noise (Trondle et al., 2023).These effects, coupled with decreased network integration (reduced global efficiency and small-worldness) (Javaid et al., 2022;Meunier et al., 2009;Achard and Bullmore, 2007;Onoda and Yamaguchi, 2013), suggests disruptions in local and network communication effectiveness.Age was also associated with a loss of functional specialization (segregation), manifesting as smaller and more local modules across brain networks (Song et al., 2014).Thus, aging may be characterized by an increased asynchronous activity and reduced network integration and segregation.
Better cognition was associated with increased cortical activation, and decreased asynchronous neural activity.Improved cognition predicted enhanced information processing capacity, as reflected in increased alpha power across occipital electrodes (Hindriks and van Putten, 2013;Pfurtscheller and Lopes da Silva, 1999).Better cognition was also associated with an increase in general spiking activity (offset) (Pei et al., 2023;Zhang et al., 2023;Waschke et al., 2021) and a decrease in neural noise (slope) (Smith et al., 2023;Ouyang et al., 2020;Pei et al., 2023) in occipital, temporal, and hippocampal regions.This aligns with previous results showing that aperiodic activity reflects dynamic adjustments of metacognitive states crucial for successful cognitive performance (Zhang et al., 2023).A positive association with whole-brain entropy and a negative with FD and spectral structure confirmed the role of complex dynamics in cognition (Iinuma et al., 2022;Parbat and Chakraborty, 2021).Thus, increased cognitive performance is consistently associated with both local and global dynamics, involving increased cortical activation, enhanced alpha oscillations, and complex brain dynamics.
Fewer years of education correlated with higher individual alpha frequencies in frontal regions and reduced integration and segregation of brain networks.However, these associations were weaker in comparison with other metrics, as reported with fMRI (Raz et al., 2005).While direct assessments of education as a predictor of EEG metrics have been limited in prior studies, previous research including education as a covariate, showed no significant effects on periodic frequency (da Cruz et al., 2020) or graph-theoretic measures (Tan et al., 2019).Similarly, sex did not emerge as the most influential predictor in any model.Sex differences have been observed (Carrier et al., 2001;Pravitha et al., 2005) and not detected in EEG studies (Trondle et al., 2023;Rempe et al., 2023;Gaubert et al., 2019).Our results suggest that sex is a less influential individual predictor, although it influences signals when combined with other demographic or cognitive variables.The impact of education and gender on brain phenotypes should be further investigated by exploring more specific measures capturing population diversity.
Across all models, periodic frequency and graph-theoretic measures demonstrated the highest sensitivity in capturing individual differences.Older age, worse cognition, and being male predicted lower alpha power, while older age and lower education were associated with less integrated and segregated networks.Even when combined with age, education was not found to be an influential predictor, supporting a weak relationship between this factor and brain signals (Raz et al., 2005;da Cruz et al., 2020;Tan et al., 2019).Older age and worse cognition, when combined, are linked to reduced alpha power and increased neural noise, aligning with the neural noise hypothesis of aging (Voytek et al., 2015;Cremer and Zeef, 1987), stating that with increasing age, neural noise rises, and the reliability of neural communication diminishes, contributing to cognitive decline.Therefore, age-related cognitive decline might be attributed to these combined effects, where decreased alpha oscillatory activity allows for more neural noise, which, in turn, disrupts neural communication (Voytek et al., 2015).As complexity and aperiodic components can emerge by intrinsical modulation of brain dynamics, or as a consequence of metabolic or energetic impairments (Medel et al., 2023;Kluger et al., 2023), future studies should investigate the influence of individual differences in demographics and cognition on changes in these metrics.
It is worth noting that some models demonstrated low R 2 values (R 2 < 0.1); the residuals of these models did not show an approximately normal distribution and exhibited poor fit.However, due to the intrinsic complexity of neuroimaging data and the inherent variability in brainbehavior associations, even models with low R 2 can provide valuable insights into the effects among variables.Additionally, the variation in R 2 values provide important information about the strongest effects.

Limitations
Several limitations of our study should be acknowledged.Cognitive data were only available for a subsample of participants.Additionally, cognition was assessed using a screening tool.Although the MMSE has been widely used as a reliable measure of general cognitive state (Folstein et al., 1975), it may not fully encompass the spectrum of cognitive abilities.Contrary to prior findings associating graph metrics with cognition (Bullmore and Sporns, 2009;Yu et al., 2021;Stanley et al., 2015), our results indicated that MMSE scores did not possess strong predictive value.Limited performance variability in MMSE scores, with all participants scoring above 24, affects the range and reduces variability.We acknowledge the limitations of the MMSE as a tool for assessing cognition, particularly in healthy populations.However, we

Table 3
Results for aperiodic spectral metrics on the subsample with available cognitive data.
A. Models with age as the best evaluated feature Left temporal region (slope), F = 50.12,p < 1e15, R 2 = 0.12, R 2 adjusted = 0.12, CI = 0.06, F 2 = 0. employed the MMSE because it remains a valuable cognitive screening in both clinical and research settings.Its widespread availability, ease of use, and brevity render it a useful tool for initial cognitive assessment in environments necessitating rapid and accessible evaluations (Creavin et al., 2016).Further assessments should explore the role of individual differences in cognition using more comprehensive measures across various cognitive domains.Moreover, the measure of years of education may not be sufficiently sensitive to capture population diversity in this factor.Future studies should consider incorporating more detailed measures that provide information on the quality and nature of education received by participants.As we did not inquire about gender identities, a factor that can significantly influence diversity among populations, future studies should systematically address individual differences related to gender identities to provide a more comprehensive understanding of the associated brain signatures.Also, although country-level analyses are highly relevant, we included only a limited number of nations with unbalanced sample sizes, thereby reducing the possibilities for cross-country interpretations.Although these effects are beyond the scope of this work, future global approaches with larger, balanced, and more diverse samples should explore country-level effects effectively.
It is also worth noting that the choice of parceling scheme and methodology for defining the ROIs can influence the outcomes.In our study, we opted to employ ROIs for two primary reasons: (i) To mitigate potential effects resulting from employing different electrode configurations and quantities, given the EEG's low spatial resolution and the Fig. 3. Results on the models using aperiodic spectral metrics as outcomes.(A) The three best models when age was the best-evaluated feature identified age and cognition as the metrics with the highest evaluation, both of which were statistically significant.Education and sex did not reach statistical significance.In the total sample, R 2 experienced a decrease, but the models remained statistically significant.(B) The models when education was the best-evaluated feature were presented with transparency because none of them were statistically significant.(C) The three best models when cognition was the best-evaluated feature had an R 2 lower than 0.1, but they were still statistically significant.Education and Sex did not attain statistical significance in any model.In the total sample, the models experienced a significant decrease in R 2 .The brains in the right column represent the intensity of predictors in the respective brain regions for statistically significant models, for the subsample with cognitive data.The models in hippocampus and temporal regions were the most significant ones.The statistical values for the models with the complete sample are provided within each box.
variance in estimation error that arises from using low and high-density electrode configurations.This approach facilitates comparison of results across a broad spectrum of electrode arrangements and quantities.(ii) To reduce the number of regressions.Despite this reduction, we still had to construct a substantial number of models.On the other hand, using ROIs requires finding consistent effects in broad regions, and as our results demonstrate, they displayed robust effects.Future studies using only high-density arrays should explore the impact of spatial variation using finer areas.In adittion, we observed variations in the number of channels across different centers, ranging from 21 to 132 electrodes.To address the potential effects of these variations, we employed a mesh model approach for integrating the electrode layouts (Melnik et al., 2017).We applied spatial normalization to ensure that the EEG inversion process remains unaffected by the number of channels (Prado et al., 2023).This was verified by confirming that the number of channels did not predict the signal quality in the source space.Furthermore, our results were reported with an MNI average brain for source estimation, given that individual MRI data was unavailable.Source space provides more adequate spatial resolution with 64+ channels.Although challenging, source analysis can be efficiently done with low-density electrodes (Nguyen-Danse et al., 2021;Soler et al., 2020;Baroumand et al., 2018).We constructed a regression model to predict signal quality based on the number of channels, which yielded non-significant results, suggesting that the number of channels did not impact the reported associations.We avoided analyzing small regions and focus only on larger effects.However, using the MNI average brain for source estimation with low electrode density represents a limitation.Future studies should conduct further analyses of the impact of the number of channels on source variability and its associations with brain phenotypes.
Moreover, the correlation among the reported outcomes exhibited expected values.Metrics that were spatially close showed high levels of correlation, as did the connectivity metrics.However, the regressions of predictors with the outcomes were independent of the relationships among the outcomes.We opt to use different measures as is typical in the literature (Cesnaite et al., 2023;Zappasodi et al., 2015;Donoghue et al., 2020;Vecchio et al., 2022) and emphasize their individual effects.Finally, we identified collinearity between cognition and education in the sample with cognitive data, while there were no collinearity in the complete sample.Collinearity can cause issues with estimated accuracy, high coefficient variances, interpretation challenges, among others (Snee, 1983).However, our models showed a high level of consistency and precision in their estimates across both samples (sample with cognitive data and the total sample).The results yielded the same effects when employing a collinearity-robust model such as Ridge regression (Tsigler and Bartlett, 2024).

Implications and future directions
Despite the diversity of our sample and the heterogeneity in data acquisition across centers, our results revealed that demographic and cognitive factors robustly predicted EEG modulations.These findings reflect the potential of using more affordable and scalable techniques, such as EEG, for the study of individual differences contributing to diversity and brain signatures.Although we did not include EEG microstates in this work, as it was beyond the scope of our study, previous reports have shown sex-specific changes in microstate dynamics during adolescence as well as at older age (Tomescu et al., 2018).Additionally, changes in microstates have been associated with age, from childhood (Hill et al., 2023) to adulthood (Koenig et al., 2002).As EEG microstates can inform the temporal dynamics of large-scale brain networks across millisecond timescales, future studies should use them to explore the predictive value of demographic and cognitive factors across large and diverse populations.Similarly, given that other complexity metrics like multiscale entropy have shown sensitivity to variations in age (Wang et al., 2016) and cognition (Maturana-Candelas et al., 2019), future research should leverage these tools in large and diverse samples.
Future studies should employ EEG metrics in normative modeling to create brain charts as anchor points for standardized quantification of brain functioning over the lifespan, considering individual differences contributing to population diversity.Normative modeling with neuroimaging measures (Bethlehem et al., 2022;Elad et al., 2021;Di Biase et al., 2023;Rutherford et al., 2022;Habes et al., 2021;Rosenberg et al., 2020) has clarified individual differences in the context of brain development or aging, brain health and disease, and mapping variations across multiple cognitive domains.The development of has brain charts challenged traditional interpretations of case-control differences (Marquand et al., 2016), which become problematic in domains such as neurodegeneration and psychiatry where disorders are diagnosed based on symptoms that overlap between disorders, often yielding heterogeneous clinical groups.As normative modeling does not require categorical partitioning, applying this approach using EEG signals will allow for a more global applicability in parsing heterogeneity and diversity in brain health, psychiatry, and neurodegeneration.
Our findings endorse EEG as a potential tool that, in the future, should be critically incorporated into case-control studies in clinical settings (Rossini et al., 2020).While EEG offers numerous benefits for investigating physiological changes linked to pathology, it often relies on group-level data to identify associations or differences, which limits its utility at the individual diagnostic level.Established methods like MRI, PET, and cerebral spinal fluid analysis offer undeniable value in individual diagnosis; however, they are often expensive and have limited accessibility, particularly in resource-constrained settings.EEG offers complementary accessible tools that can aid in revealing physiological changes linked to early pathology and the risk of dementia (Rossini et al., 2020;Parra, 2022).

Conclusion
Variations in EEG metric typically used in case-control studies are influenced by individual differences in demographics and cognition.Older age, poorer cognition, and being male were associated with reduced alpha power, whereas older age and less education were linked to less integrated and segregated networks.Moreover, older age and worse cognition were linked to reduced alpha power and increased neural noise.Our findings pave the way for the future use of EEG-based brain charts for standardized quantification of brain functioning over the lifespan, considering individual differences contributing to  In terms of connectivity metrics, cognition never emerged as the bestevaluated feature, whereas in complexity metrics, neither age nor education stood out as the top-rated predictors.(A) The three best models when age was the primary feature exhibited significant results with high R 2 values and effect sizes for measures of segregation, integration, and small-worldness.Age and education were the best evaluated and significant predictors in the models.In the total sample, the models maintained almost the same performance.(B) The three best models when education was the primary feature yielded a low R 2 but remained statistically significant.Education and age were the best-evaluated predictors, whereas cognition and sex did not hold significance.(C) Results for complexity metrics.The three best models when cognition was the primary feature for predicting connectivity metrics emerged with several complexity metrics: fractal dimension, spectral structure variability, and Wiener entropy.Cognition had a much higher value compared to other predictors.In the total sample, the models significantly decreased R 2 values (see values inside the box).Brains in this panel indicate that the most predominant regions were the whole brain, along with the occipital, inferior frontal gyrus, and temporal lobe.These regions prevailed in both hemispheres.The brains to the right of panels (A) and (B) represent the utilized connectivity metrics for the subsample with cognitive data.The statistical values for the models with the complete sample are provided within each box.
H. Hernandez et al.

Fig. 1 .
Fig. 1.Study design.(A) Sample: Participants were recruited from a multicenter study encompassing regions in the global south (Argentina, Brazil, Chile, Colombia, Cuba), and the global north (Italy, Ireland, Turkey, United Kingdom).(B) Predictors:The dataset comprised a total of 1298 participants, with cognitive data available for 725 individuals.Regression models tuned with cross validation and data partition were developed using years of education, age, sex, and cognitive state as predictors.(C) EEG data preprocessing: The preprocessing steps included electrode re-referencing, noise removal through signal filtering, resampling to Hz, artifact removal (e.g., blinks and eye movements), and transformation into a common source space using the Automated Anatomical Labeling (AAL) atlas.(D) Outcomes: EEG data in the source space was used to compute four groups of metrics, categorized into power spectrum, aperiodic, complexity, and connectivity metrics.(E) Data analysis: Linear regressions were applied, utilizing data partitioning with an 80% training sample and a 20% testing set, cross-validated (k = repetitions).We reported the importance of predictors and the results on the influence of each predictor on specific brain regions.

Fig. 4 .
Fig.4.Results on the models using connectivity and complexity metrics as outcomes.In terms of connectivity metrics, cognition never emerged as the bestevaluated feature, whereas in complexity metrics, neither age nor education stood out as the top-rated predictors.(A) The three best models when age was the primary feature exhibited significant results with high R 2 values and effect sizes for measures of segregation, integration, and small-worldness.Age and education were the best evaluated and significant predictors in the models.In the total sample, the models maintained almost the same performance.(B) The three best models when education was the primary feature yielded a low R 2 but remained statistically significant.Education and age were the best-evaluated predictors, whereas cognition and sex did not hold significance.(C) Results for complexity metrics.The three best models when cognition was the primary feature for predicting connectivity metrics emerged with several complexity metrics: fractal dimension, spectral structure variability, and Wiener entropy.Cognition had a much higher value compared to other predictors.In the total sample, the models significantly decreased R 2 values (see values inside the box).Brains in this panel indicate that the most predominant regions were the whole brain, along with the occipital, inferior frontal gyrus, and temporal lobe.These regions prevailed in both hemispheres.The brains to the right of panels (A) and (B) represent the utilized connectivity metrics for the subsample with cognitive data.The statistical values for the models with the complete sample are provided within each box.

Table 1
Demographic and cognitive state information.

Table 2
Results for the power spectrum metrics on the subsample with available cognitive data.A.Models with age as the best evaluated feature Left occipital region (subj spec α low equivalent power), F = 129.36,p < 1e-15, R 2 = 0.27, R 2 adjusted = 0.26, CI =0.07, F 2 =0.36