Beyond network connectivity: A classification approach to brain age prediction with resting-state fMRI

The brain is a complex, dynamic organ that shows differences in the same subject at various periods. Understanding how brain activity changes across age as a function of the brain networks has been greatly abetted by fMRI. Canonical analysis consists of determining how alterations in connectivity patterns (CPs) of certain regions are affected. An alternative approach is taken here by not considering connectivity but rather features computed from recordings at the regions of interest (ROIs). Using machine learning (ML) we assess how neural signals are altered by and prospectively predictive of age and sex via a methodology that is novel in drawing upon pairwise classification across six decades of subjects ’ chronological ages. ML is used to answer the equally important questions of what properties of the computed features are most predictive as well as which brain networks are most affected by aging. It was found that there is decreased differentiation among the neural signals of older subjects that are separated in age by the same number of years as younger subjects. Furthermore, the burstiness of the signals change at different rates between males and females. The findings provide insight into brain aging via an ROI-based analysis, the consideration of several feature groups, and a novel classification-based ML pipeline. There is also a contribution to understanding the effects of data aggregated from different recording centers on the conclusions of fMRI studies.


Introduction
An understanding of the changes that occur to a brain with increasing age is fundamental to human neuroscience.While variability in person-to-person brain aging is expected due to environment and life events, there are also variations in growth and structure that may not be as obvious.Resting-state fMRI (rsfMRI) can be collected relatively quickly and easily from subjects of different ages, demographics, and possible pathologies.Atypical brain aging is often reflective of, or a precursor to developmental disorders (Morgan et al., 2018).Furthermore, abnormal trajectories in brain age have been implicated in Parkinson's disease (PD) (Kim et al., 2017), Alzheimer's disease (AD) (Habes et al., 2021), as well as several other neural pathologies (Mattson and Arumugam, 2018) and even mortality (Cole et al., 2018).The simple deviation of a person's brain age from the trajectory of their chronological age may be a precursor to adverse clinical outcomes stemming from neurological disease, neurodegeneration, and cognitive impairment.Thus, predicted brain age is a valuable biomarker.
Due to its importance, age prediction from neural data has been undertaken with sizable datasets (Habes et al., 2021;Mennes et al., 2013;Webb-Vargas et al., 2017).Subsequent to correlational studies and graph-theoretic approaches (Biswal et al., 2010;Sporns et al., 2000), machine learning (ML) has been used to discover structure from neural data by performing either regression or classification on held-out data (Khosla et al., 2019;Vergun et al., 2013).There exists a growing number of works that use ML of fMRI to answer questions pertaining to brain aging.Dosenbach et al., (2010) used support vector regression (SVR) of rsfMRI data to estimate brain age (Dosenbach et al., 2010).Three groups of subjects were considered with the chronological ages lying between 6 and 35 years.Thus, the age ranges may be considered as rather limited in comparison to more recent works as well as the data used here (Cole, 2020;Siman-Tov et al., 2017;Nooner et al., 2012).The work computed 12,720 functional connectivity features via correlation among all regions of interest (ROIs) of a subject over the recording (Dosenbach et al., 2010).LASSO regression was used to predict brain age by considering over 14,000 healthy participants from the UK Biobank (Cole, 2020).Interestingly, it was reported that both resting and task-based fMRI are relatively uninformative in comparison to several variants of MRI, i.e.T1-weighted MRI, T2-FLAIR,T2*, and diffusion-MRI, at chronological age prediction (Cole, 2020).A review on the use of ML techniques for rsfMRI analysis is provided in (Khosla et al., 2019).The referenced studies provide background and highlight the importance of human age distinguishability from brain recordings.The deviation of a person's brain age from biological age may also be affected by sex as there is no reason to assume that the prospective changes are uniform across gender.Works are increasingly incorporating sex as either a predictor or a response variable when evaluating brain age.For instance, Pervaiz et al., (2020) considered over 9000 pipelines consisting of different hyperparameter values across a large (over 14,000) number of subjects from Human Connectome Project (HCP) and the UK Biobank (Pervaiz et al., 2020).Correlation-based measures of functional connectivity were computed and used as inputs for several classifiers including variants of deep learning (DL) architectures.It was noted that the general regularized regression method elastic net performs comparably with DL schemes for prediction of age, sex, and intelligence.
Attempting to decipher differences between male and female brains is an active area of research as noted in recent works (Weis et al., 2020;Sen and Parhi, 2020;Sanchis-Segura et al., 2022).Ritchie et al., (2018) considered rsfMRI data recordings from over 4000 participants across the age range of 44-77 years via the UK Biobank with an approximately even split between the sexes (Ritchie et al., 2018).The authors computed partial correlations between 55 ROIs per subject.While sex prediction was not undertaken, it was reported that the connectivity between sensorimotor, visual, and rostral lateral prefrontal areas was higher in males, whereas the strength of connectivity within the default mode network (DMN) was higher in females.Prior (Scheinost et al., 2015) and more recent (Cai et al., 2020) works have considered rsfMRI data from different cohorts of subjects than those mentioned so far and found significant differences in connectivity between sexes at various chronological ages.Similar to (Ritchie et al., 2018), the two works did not consider ML approaches nor address the question of at which ages male and female brains vary the most.More recent works such as Weis et al., (2020) have considered sex prediction from rsfMRI using over 1400 subjects from the HCP and 1000BRAINS datasets (Weis et al., 2020).Separate support vector machines (SVMs) were trained on data from individual brain parcels to differentiate between male and female subjects.Subjects in the 18-88 age range were considered.Interestingly, sex prediction via analyses restricted to specific brain regions was seen to provide performance levels comparable or higher than attained with the whole-brain connectome (Weis et al., 2020).Another recent work computed Pearson's correlation values across 85 regions from subjects in HCP and evaluated dynamic functional connectivity (FC) by repeating the computation along different time intervals within the recordings (Sen and Parhi, 2020).This yielded a three-way tensor that was used to predict sex via a random forest classifier at an AR of 0.94.However, there are several caveats with this high AR.The subjects' chronological ages were in the 22-36 year interval, and task-fMRI data from seven tasks (including an emotion task) were incorporated with the rsfMRI data.We shall consider a wider age range in the analysis.
The conventional analysis of utilizing fMRI data to predict brain age and sex has examined alterations in measures of FC across brains.Typically, patterns of correlated activity between ROIs are compared among subjects of different ages to assess whether there are differences (Dosenbach et al., 2010;Cole, 2020;Khosla et al., 2019;Sen and Parhi, 2020;Siman-Tov et al., 2017).We take an alternative approach by not considering FC but rather metrics derived from the measurements that would go into its calculation.Thus, our analysis is not connectivity-based in the sense of the prior works.Nevertheless, it is expected that changes in brain connectivity would manifest in the results since ROI signals share correlation and are perhaps causal to or compensatory for each other.This is a result of FC being derived from the underlying BOLD responses, and relationships between BOLD and neural activity (Tagliazucchi et al., 2012;Lahaye et al., 2003).Newer works have taken similar node-based approaches with insightful findings stemming from the local BOLD signal dynamics of human brains (Shafiei et al., 2020;Fallon et al., 2020).The ROI-centric approach provides spatially specific insight since it contains the identity of the brain region that the scrutinized signal originated from, rather than providing the relative network connections as would be the case with pairwise or group-based FC measures.There is also a reduction in the number of features that must be computed leading to savings in computation.This is increasingly desirable when considering a large number of ROIs and subjects.In this work the data is evaluated to assess how the neural signals are altered by and prospectively predictive of brain age and sex.A novel methodology is used that draws upon pairwise classification across six decades of subjects' chronological ages.A variety of features are computed from the per-ROI BOLD signals of the subjects.ML shall be used to probe whether statistical or paroxysmal properties of ROIs change with healthy brain aging.Aside from testing the signal properties, the spatial components of the analysis will provide insight into the brain regions and functional networks' contributions.We deviate from a conventional approach to age prediction of using chronological age as a continuous variable and training a machine to do regression.The binning of age groups and pairwise classification facilitates a categorical analysis between age and sex by focusing the analysis on specific age groups within the full spectrum.
The contributions of this work are summarized as follows.
• Following the per-ROI processing of subjects' BOLD signals into three feature groups, an ML analysis was used to perform pairwise classification.Accuracy rates are computed by comparing the predicted age group of subjects with their chronological age group.It is noted that there is decreased differentiation of neural signals from older subjects that are separated in age by the same number of years as younger subjects (Results -Predicting Comparative Brain Age via rsfMRI, Fig. 2, Figure S1).• The differentiation in age is highly affected by the distinguishability offered from ROI identities.This was noted by shuffling the ROIs of subjects prior to the ML analysis (Results -Predicting Comparative Brain Age via rsfMRI, Fig. 3, Figure S2).• The ML pipeline identified the default mode and visual networks as being the most distinguishing among various age groups (Results -Network and ROI Contributions to Brain Aging, Fig. 4, Figure S3).• The most predictive individual ROIs are isolated and have predictive capability that were either competitive or better than that attained with entire brain recordings (Results -Network and ROI Contributions to Brain Aging, Fig. 5, Table 3).• In sex classification, the accuracy of identifying a female brain was generally higher than identifying a male brain.However, the predictive capability across decades is not very impressive (Results -Sex-dependent Changes in Brain Aging, Fig. 6).• It is observed that the burstiness of the BOLD signal changes differently among sexes and is modulated by the subjects' age group (Results -Sex-dependent Changes in Brain Aging, Fig. 7, Figure S4).• Aggregation of subject data across recording centers generally provides superior predictive capability to that attained at a single-site (Results -Recording Center Influence on Predicting Brain Age, Fig. 8, Table S2).• A test center may introduce a fingerprint into its recordings.This can lead to a reliable recovery of the recording center's identity from its subjects' features (Results -Recording Center Influence on Predicting Brain Age, Fig. 9).

Materials and methods
Subjects and imaging acquisition.The subjects consisted of participants from the International Neuroimaging Data-sharing Initiative (INDI), 1000 Functional Connectomes' Project (1000FCP) (Mennes et al., 2013;Nooner et al., 2012).The cohort entailed N = 887 individuals from 17 centers with the requirement that a center provide a minimum of 10 participants with full brain anatomical and functional coverage.The ages of the subjects spanned 21-85 years, and 514 were female.All subjects were at rest and not performing any task during recording.Some subjects had their eyes open while some had their eyes closed, and each recording consisted of signals collected for 42 ROIs.Sixteen individuals from several centers were excluded due to excessive head motion during rsfMRI acquisition according to the described movement parameters criteria (Power et al., 2014).Information regarding the participants, number of time points collected for each recording, and the rsfMRI acquisition and epidemiological parameters of the different centers have been detailed in Table 1 as well as Tables 1  and 2 of (Siman-Tov et al., 2017).
Preprocessing and head motion correction.The functional images were preprocessed in (Siman-Tov et al., 2017) using the FSL software (FMRIB Software Library v. 5.0.1,Oxford, UK) and SPM software (Statistical Parametric Mapping software package, Wellcome Department of Imaging Neuroscience, London, UK) following conventional methods originally described in (Kahn and Shohamy, 2013;Kahn et al., 2008).The following standard steps (Kahn and Shohamy, 2013) were preformed: 1) slice-timing correction; 2) rigid-body motion correction; 3) registration to the MNI152 space; 4) regression of nuisance variables including ventricles, white matter, and global average signal 5) temporal bandpass filtering (0.01-0.08 Hz).The preprocessing included rigid body correction for motion within and across runs (FSL), normalization to the standard echo-planar imaging (EPI) template of the Montreal Neurological Institute (MNI), and compensation for slice-dependent time shifts (SPM).Several key acquisition parameters are shown in Table 1 with additional information such as the number of slices, voxel size, and subject handedness available in Table 1 of (Siman-Tov et al., 2017).
The preprocessed functional data (in atlas space) were temporally filtered to remove constant offsets and linear trends over each run while retaining frequencies below 0.08 Hz.Data were spatially smoothed using a 4 mm full-width half-maximum Gaussian blur.Sources of spurious or regionally non-specific variance were removed by regression of nuisance variables.This included six parameters obtained by rigid body head motion correction, the signal averaged over the whole brain (global signal), the signal averaged over the lateral ventricles, and the signal averaged over a region centered in the deep cerebral white matter (WM).Temporally shifted versions of these waveforms were removed by inclusion of the first temporal derivatives (computed by backward differences) in the linear model.Global signal regression (GSR) was applied as part of the preprocessing steps to account for global motion-and respiratory-related artifacts.Extensive analyses based on the same data following preprocessing with and without GSR yielded only minor differences in the results as reported in (Siman-Tov et al., 2017).
Head motion is a confound in fMRI.The analysis used in (Siman-Tov et al., 2017) to account for head motion is summarized as follows.The parameters were calculated for each participant as proposed by (Power et al., 2012).Framewise displacement (FD), which represents head displacement from volume to volume, was computed as the sum of the first derivative of the six rigid-body motion parameters estimated during standard volume realignment.Delta variation signal (DVARS), which represents the change in BOLD signal intensity from one frame to the next, was computed as the root mean square average of the first derivative of fMRI signals across the entire brain.A standardized version of DVARS was applied according to (Nichols, 2017).Volumes with FD value over 0.5 mm or DVARS value over 1.5 IQRs above the 75th percentile was removed with one prior and two subsequent volumes.Participants with less than five minutes of rsfMRI after censoring were excluded.
ROI definition and functional connectivity analysis.The analysis was based upon preprocessed data regarding brain ROIs obtained from a previous study.We provide an overview of the most relevant section from the methods of (Siman-Tov et al., 2017).Seed-based analysis was performed as previously described (Van Dijk et al., 2010;Fox et al., 2005;Kahn et al., 2008) to study four high-order cognitive networks: default mode network (DMN; (Andrews-Hanna et al., 2010)), salience network (SN; (Seeley et al., 2009)), dorsal attention network (DAN; (Van Dijk et al., 2010)), and frontoparietal control network (FPCN; (Brier et al., 2012)) as well as three primary sensory and motor networks: auditory network (AN; (Brier et al., 2012)), visual network (VN; (Dai et al., 2013)), and motor network (MN; (Seeley et al., 2009)).To delineate these networks, the following seed regions were used: left posterior cingulate cortex (LPCC) for DMN; right frontoinsula (RFI) for SN; right intraparietal sulcus (RIPS) for DAN; right superior parietal cortex (RSP) for FPCN; left auditory cortex (LAC) for AN; right visual cortex (RVC) for VN; and left motor cortex (LMC) for MN.Each seed was defined as a 6 mm radius sphere centered on previously published foci (Table S1).Correlation maps were produced by extracting the time course from each of the above seeds.Then, the Pearson correlation coefficient between the time course and the time course of each voxel across the whole brain was computed to create the voxel-wise connectivity maps.SPM8 software was used to compute statistical maps of each network across participants.Maps of a young age group (N = 458, 21-30 years) were used to identify peak coordinates of additional regions of interest (ROIs) representing each network.All regions were defined as mm radius spheres around the peak coordinate (Table S1).

Table 1
Recording center demographics for the N = 887 subjects from the 1000 Functional Connectomes' Project (1000FCP) that were used in the analysis.Additional information such as the number of slices, voxel size, and subject handedness can be found in Table 1 of (Siman-Tov et al., 2017).Computed features.Each subject's recording consists of signals collected for 42 ROIs.Each signal was summarized into five statistical features consisting of its mean, variance, median, min, and max.We also computed two paroxysmal features from each signal.More specifically, a signal was Z-scored and a deviation-low was declared whenever the signal fell below its − 2σ value.Conversely, a deviation-high was declared whenever the signal exceeded its 2σ value.The total number of deviation-low and deviation-high counts for a signal are referred to as ND low (ND: number of deviations) and ND high, respectively.It should be noted that the features have been selected to be simple, reflective of the temporal properties of the BOLD signal, while not being redundant of one another.More specifically, none of the features can be faithfully reproduced from the remainder of the features.Combining the computed statistical and paroxysmal features across the 42 ROIs results in a total of 294 features per recording -210 statistical and 84 paroxysmal.In examining the burstiness of the BOLD signal of subjects at various age groups, the distribution of the paroxysmal features was considered to assess whether a trend was due to the presence of a relatively few extreme outliers.An outlier was defined as a value that is more than three scaled median absolute deviations (MAD) from the median.The MATLAB function isoutlier with a default value of MAD = 3 was used to locate outliers.
ML analysis.The features computed from each participant were used either as training or test data in the ML analysis.The ML technique used was an SVM with a linear kernel, and a binary classifier.The ML and cross-validation techniques were implemented in R via the packages e1071 and kernlab.For comparison between two groups, a leave-oneout analysis was conducted with Monte-Carlo (MC) sampling of the subjects from the two groups.The MC technique consisted of 1000 random iterations with one subject from each of the two groups left out at each iteration.More specifically, let N 1 denote the number of subjects in the young group and N 2 the number in the aged group, the following steps have been taken.
Step 2) Train an SVM on m randomly selected recordings from the N 1 young and the N 2 aged subjects.The feature vectors will consist of the computed features (i.e.statistical, paroxysmal, or combined) across all ROIs of a subject.The label reflects whether the subject is from the aged or young group.
Step 3) Use a left-out recording from the set of aged and the set of young subjects as the test data.The trained machine is to predict whether each of the two residual recordings are young or aged.
Step 4) Compare the labels of the two left-out recordings to the two respective predictions and record the accuracy of the predictions.
Step 5) Repeat Step 2 to Step 4 for 1000 iterations to attain an accuracy rate (AR) for predicting a young and aged subject.
It should be noted that the decisions at every iteration were mutually independent of each other.The above steps are depicted in Fig. 1.The resultant AR may be viewed as the performance of a trained machine being tested on left-out data for each respective pairwise comparison.In assessing the statistical significance of the ARs, a value greater than 0.65 will be referred to as statistically significant since it is approximately one-sigma greater than the chance value of 0.5 (binary classification).The choice of 0.65 follows from the combination of the empirical rule in statistics and a central limit theorem (CLT) argument since the AR is a summation of many indicator functions over mutually independent trials.An AR of 0.5 or lower is equated to a futile classifier as it has not exploited any structure beyond what could be achieved through uninformed guessing.It is valuable to compare the findings with ML to a conventional statistical approach.The AR is a single number rather than a series of values or a distribution, thus it is not possible to compute a canonical statistical test on its significance.Nonetheless, the ML pipeline described above was repeated for a lower (500) and higher (1500) number of iterations to attain AR values.In all cases the attained AR was negligibly different from the numerical value reported with 1000 iterations.Two-sample Kolmogorov-Smirnov (KS) tests were used on each pair of compared age groups to determine if the features from the subjects are sufficiently different.
Methodology for brain age, sex, test center classification.The steps in the described ML analysis and Fig. 1 have been for brain age Fig. 1.An ML pipeline to evaluate whether features extracted from rsfMRI recordings can be used to differentiate the subjects' chronological age group or sex.The sets A and B will be assigned to either two age groups or two sexes depending on the study.At any iteration of the pipeline one random data point from each set will be left out and used as the test data (orange).Furthermore, the decisions at every iteration are mutually independent of each other.The pipeline is iterated to attain a classification accuracy rate (AR).
prediction by identifying a subject as young or aged.The analysis may be referred to as an age group pair classification task.After each classification, the prediction was compared to the ground truth label of whether the recording belonged to a relatively young or old subject.Rather than using all ROIs, the analysis was also performed on a perbrain network basis.We consider the seven networks discussed in (Siman-Tov et al., 2017) (see Table S1 for the MNI coordinates) with Fig. 1 applied separately to each network.A more fine-grained investigation was conducted by considering the predictive accuracy offered by individual ROIs across the brain.Specifically, the ML pipeline of Fig. 1 was applied to features belonging to single ROIs from the extrema age groups.Namely, 38 brains were randomly selected as the test data from each of the two groups -21-30 years and 71+ yearswith one aged and young subject held-out.A second analysis was conducted for the more challenging scenario of the subjects belonging to the median age groups of 41-50 (young) and 51-60 (aged).
The sex prediction study entailed the ML pipeline of Fig. 1 applied to each of the age groups in Table 2 with sets A and B denoting male and female subjects.The subjects can be viewed as undergoing an agematched comparison with a subject's sex representing its label in the supervised analysis.Lastly, an altered version of the pipeline from what has been discussed so far was used to investigate if it is possible to differentiate between pairs of recording centers via the BOLD signal features of the subjects.In conducting a binary classification analysis between the recordings coming from two centers, two-thirds of the recordings in the center with the fewer subjects were used as the number of training samples taken from the sets A and B. The remaining held-out recordings comprised the test data.The machine's predictions on the test data were compared to the ground truth labels to attain an accuracy rate.
Ethics statement and code availability.All datasets used in the present study have been obtained from previously published studies which have been approved by their respective institutional review board or relevant research ethics committee.The data was further processed Fig. 2.An evaluation of the capability to distinguish between young and aged subjects via rsfMRI recordings and an ML algorithm trained on one of three feature groups.The analysis may be referred to as an age group pair classification task.The subjects from 17 recording centers are combined in the analysis.A) The mean AR as a function of the age gap separating the subjects in the classification analysis.B) Designation and naming of the age groups used in the 15 pairwise comparisons.The Δ value of 0 (no gap), 10, 20, and 30 years are shown with the indices 1 to 5 denoting the location of the Δ along the spectrum of the considered chronological ages.C) A comparison of the ARs in differentiating young from aged brains with the three classes of features for the 15 different combinations of groups.The scenarios are ordered from those that are the closest proximity in chronological age (i.e.no gap) to the furthest apart in age (i.e.extrema).The dashed, red line indicates the chance predictive value of 0.5 associated with binary classification.The binarized heatmap identifies the ARs that are statistically significant (greycolored) for a Δ value and a feature group.

S.K. Sorooshyari
and aggregated by Siman-Tov et al. prior to being used in this work.The aforementioned study was approved by its institutional review board and research ethics committee (Siman-Tov et al., 2017).The MATLAB code for computing the features, and the R code used for the ML processing are available at: https://github.com/sorooshyari/NeuroImage_2024.

Predicting comparative brain age via rsfMRI
We seek to investigate whether features attained from rsfMRI recordings can be used to predict if a brain belongs to a younger or an older subject.In making comparisons between subjects, two of six age groups in Table 2 were selected with one group considered as aged and the other as young.The ML pipeline of Fig. 1 was separately applied to all pairs of age groups listed in Table 2 to arrive at the results in Fig. 2. The ARs for young and aged classification decisions are also contained in Figure S1.We note that in considering the predictive accuracy with the three classes of features among the 15 combinations of age groups, the combined features provided the best accuracy rate (mean AR = 0.637) while the paroxysmal features provide the worst accuracy (mean AR = 0.556).The statistical features provided a mean AR of 0.633.In general, the accuracy in differentiating between a young and aged brain decreased with the years separating the considered groups (Fig. 2A).In fact, for the combined and statistical features, the best differentiation occurred for the 21-30 and 71+ age groups (AR of 0.848 and 0.8345, respectively).With the paroxysmal features the highest accuracy was noted between the 31-40 and 71+ age groups (AR = 0.6585).This reinforces the capability of the synergistic rsfMRI and ML framework as being able to distinguish brain age at higher accuracy when provided with increasing temporal distance between the chronological brain ages.We consider the comparative performance of the three groups of features across the pairwise analyses.The statistical features provided the highest AR in 8 out of the 15 cases while the combined and paroxysmal features provided the best AR in 5 and 2 cases, respectively (Figure S1A).The results of two-sample KS tests between pairs of compared age groups also show increased disparity with an increasing number of years between age groups (Figure S5).However, the statistical testing reveals relatively small percentages of the combined features being sufficiently different (p < 0.05) for any of the comparisons.The scenario that yielded the highest percentage of statistically significant features (35.3 %) was the 21-30 to 61-70 year comparison.
While the above results on higher differentiability with increased temporal gap are not surprising, it was interesting to note that the accuracy in differentiating among younger brains was higher than that attained among older brains (Figure S1A).For instance, when considering the combined features, we are able to distinguish between the 21-30 and 31-40 groups at a AR of 0.6225.However, the accuracy rate for distinguishing between 61 and 70 and 70+ age groups as well as between the 51-60 and 61-70 groups did not exceed 0.55 (0.4675 and 0.547, respectively).This indicates that aged brains appear more similar, in terms of the rsfMRI features, than younger brains that are separated by the same number of years.The seemingly increased differentiation of neural signals from brains that are separated by the same number of years but belonging to younger rather than older subjects is important from a human neuroscience perspective and deserves scrutiny.Via Fig. 2B we assign labels to the 15 groups according to the gap (in years) that separates the brains constituting the young and aged.The index 1, 2,…, 5 denotes the groups of subjects that are compared who are separated by the same gap but are of different agesa higher index denotes older groups of subjects.A general upward trend in AR is observed in Fig. 2C across the three feature groups.We also observe a general decline in AR with an increasing index for each gap.This is a trend that was noted for all feature groups except the paroxysmal features.The statistical features provided the highest number of statistically significant AR values (> 0.65) among the classification scenarios.Also, the Δ value in the age spectrum located towards the youngest subjects (i.e.gap 1) provides the highest number of significant AR values (Fig. 2C).
While we have determined the proficiency of the ML analysis in differentiating between young and aged brains based on attributes computed from rsfMRI recordings, it is important to assess whether the predictive accuracy is driven by region-specific differences between the brains, or if the discernment is insensitive to the brain regions and more dependent on the feature values.Accordingly, the ROIs of each recording are randomly shuffled prior to following the ML steps in Fig. 1.The 42 ROI identities have physical significance via the MNI coordinates in Table S1.By shuffling the order of ROIs of each subject prior to the ML analysis, the location of the ROIs within the feature vector are randomized, i.e. their identities are destroyed.Rearrangements such as the example in Fig. 3A aim to remove potential orthogonality (i.e.dissociation or unrelatedness) that may exist among the features of various brain regions and discards ROI identity from the predictive analysis.A null hypothesis can be viewed as the predictive capability not being affected by the shuffling.A different randomization sequence was applied for each subject.The ML pipeline was repeated but with the ROIs of the subjects randomly rearranged to attain the results shown in Figure S2 and summarized in Fig. 3B.In comparing the results with the shuffled ROIs to the accuracy rates in Figure S1 and Fig. 2 it is obvious that the differentiation in age is affected by the distinguishability offered from the ROIs.Mean AR levels of 0.5652, 0.5495, and 0.5527 are noted with the combined, statistical, and paroxysmal features corresponding to respective decreases of 11.27 %, 13.19 %, and 0.59 % in comparison to the mean rates attained in the unshuffled case.Similar to the nonshuffled ROI analysis, the AR increased with the years separating the groups of brains in the pairwise analysis.More specifically, the highest AR for the combined features was 0.740 for differentiating between the 21-30 and 71+ age groups.For the statistical features the highest attained AR was 0.682 when considering the 21-30 and 61-70 age groups, and 0.667 with the paroxysmal features for the 31-40 and 71+ age groups.It is interesting that the differentiability attained with the paroxysmal features was the least affected when comparing the shuffled to the baseline scenario of in-tact ROI identities.Nevertheless, with the ROI identities destroyed, we note via Fig. 3C that only the extrema and Δ = 30 year scenarios provide AR values that are statistically significant in differentiating among subjects.

Network and ROI contributions to brain aging
Although the human brain is elaborately wired with interneurons providing communication among seemingly unrelated brain regions, the constituent networks are often deemed as modular.In fact, networks often viewed as dedicated for various sensory, motor, and cognitive function have been delineated in works such as (Seeley et al., 2009;Van Dijk et al., 2010;Fair et al., 2009).This has translational application in isolating brain regions that may be targeted or monitored via invasive as well as non-invasive techniques.A group of closely spaced ROIs are frequently equated to a brain network that will function by interacting with other networks to induce a brain state.This was done to view networks as separate groups of ROIs and investigate which networks are the most predictive of brain age.From a systems neuroscience perspective, such networks may not show the first signs of neurodegeneration, but rather be the most indicative of biological age.The accuracy rate noted with individual networks is shown in Figure S3 for all pairwise comparisons, while Fig. 4 contains the identity of the network with the highest AR for each differentiation.When considering the combined, statistical, and paroxysmal features, the brain network that was implicated the greatest number of times is the DMN (6/15), VN (6/15), and VN (7/15), respectively.Conversely, the following networks were not the most predictive ones in any of the 15 comparisons for the respective features: MN and VN (combined), FPCN (statistical), and FPCN and DMN (paroxysmal).The three heatmaps in Fig. 4 are not summary plots since they reflect maximal rather than average values.Nevertheless, the seemingly natural trend of higher distinguishability with increasing age gap between the groups is apparent.In comparing the identities of the most predictive networks in Fig. 4, it is surprising that there is high overlap (13/15) with the statistical and paroxysmal features.Conversely, the most predictive networks according to the combined features share no overlap (i.e.0/15) with those of the other two feature classes.
While we have noted that the signals collected across the brain and the constituent networks of subjects can provide an accurate account of age, we investigate the fidelity of the signal provided by the buildingblock, namely a ROI.Whether considering single or multiple brain regions, we expect that the difference in predicting brain aging will be most pronounced for brains further apart in years.Mean AR values of 0.6948, 0.6518, and 0.6756 were noted across the individual ROIs for the combined, statistical, and paroxysmal features, respectively.The classifier accuracy rates in Fig. 5 indicate the best ROI performance as 0.841 (L-ACC of AN), 0.818 (L-ACC of AN), and 0.727 (R-FEF of DAN) for the combined, statistical, and paroxysmal features, respectively.The Fig. 3.A random shuffling of the ROIs prior to applying an ML analysis to differentiate brain age.The study is performed to determine whether the predictive accuracy is driven by region-specific differences between the brains, or if the discernment is insensitive to the brain regions and more dependent on the BOLD feature values.By shuffling the order of ROIs of each subject prior to the ML analysis, the physical identities of the brain regions are destroyed.A different randomization sequence was applied for each subject.A) An example of the random shuffling of the order of ROIs prior to applying an ML analysis.B) The ARs in distinguishing between age groups via the ML pipeline.The ROIs of the subjects have been shuffled to remove their identities and make the classification only dependent on the feature values.The binarized heatmap identifies the ARs that are statistically significant (grey-colored) for a Δ value and a feature group.
minimum accuracy attained by any ROI with the combined and paroxysmal features remained above 0.55 indicating a relative robustness regardless of the brain region.The AN and VN networks provide the ROIs that are the most predictive while the single ROIs with the combined features yielded the highest AR when comparing the three feature groups.
Conducting the single ROI analysis for the more challenging scenario of differentiating between the two median groups, a single ROI did not provide significantly above chance predictive capability.Fig. 5 shows that the AR values with the three feature classes rarely exceed 0.55.The highest AR of an ROI was 0.531 (RMC of MN), 0.532 (LPCC of DMN), and 0.566 (LIPS of DAN) for the combined, statistical, and paroxysmal features, respectively.It is interesting that the paroxysmal features provide the single ROIs with the best predictive capability in this casethe ROIs belong to the DAN and SN networks.When considering the statistical or paroxysmal features, there was an apparent bias in prediction towards one of the two decisions.The bias was lower, but still present with all the features.The accuracy rates attained with the best five ROIs are listed in Table 3 for the extrema and median group analyses.It is rather surprising that the most predictive single ROIs in Table 3 performed either better or comparable to (within 3.5 %) the predictive capability attained with the entire brain.Perhaps more interesting, the paroxysmal features of the ROIs with the highest ARs exceeded the predictive capability attained when that feature group included the entire brain.
Identification of the seven networks entailed the selection of seed regions.Subsequently the regions were used to determine the additional ROIs via recordings from young adults (21-30 years).We investigate the possibility that the connectivity-based seeds identified in the brains of young adults may have introduced bias to the finding that the age of younger subjects are more readily distinguishable than that of older subjects separated by the same number of years.The ML pipeline of Fig. 1 was applied to all of the subjects' data while restricting attention to the seven original network seeds.The pipeline was then separately used on all the subjects' data that were recorded from the 35 extra connectivity-based seeds.The two analyses provided AR values for the 15 cases where the pairwise classification was performed.The comparative results demonstrate the same findings with respect to age discriminability (Figure S6) and agree with the results attained when all 42 ROIs are considered.

Sex-dependent changes in brain aging
Aside from consisting of diverse age groups, the considered recordings are divided by sex.We investigated whether there are differences between male and female brains at various ages.The analysis is crucial in assessing the importance of accounting for sex when arriving at conclusions from brain imaging studies.It will also shed light on the fMRI features that are possibly different between men and women at the various ages.The question of at which age group male and female brains appear the most different concomitantly provides insight into if it is possible to differentiate sex based on BOLD features.The results in Fig. 6 indicate that accuracy in identifying a female brain was in general higher than identifying a male brain.This held for 5/6 age groups when considering the combined features and 4/6 age groups for the statistical and paroxysmal features.Interestingly, in the 21-30 age group there existed a similar, marginal accuracy in predicting sex for male and female recordings across the three feature classes.The male and female brains showed the largest difference in the 41-50 age group when considering the combined features (AR = 0.632) and statistical features (AR = 0.606).Similarly, both feature classes provided the lowest accuracy with the 61-70 year age group.The highest predictive capability with the paroxysmal features (AR = 0.605) was noted for the 31-40 age group while the lowest (chance) was noted for the 21-30 group.In summary, the rsfMRI data and ML analysis reveals differences of varying degrees between male and female brains across the age groups.The highest accuracy rate attained for any age group (AR = 0.632) was not very impressive.It was not possible to reliably predict sex based on Fig. 5.A scrutiny of individual brain regions' reflection of aging.The ML pipeline was applied using the A) combined, B) statistical, and C) paroxysmal features.An AR is shown for each ROI in considering two cases.In the first case, we discerned between the extrema age groups that consist of subjects 21-30 (young) and 71+ years old (aged).Secondly, the analysis was undertaken for the median age groups consisting of subjects 41-50 (young) and 51-60 years old (aged).In addition to the aggregate AR (i.e.all), in each plot the accuracy of separately differentiating young and aged brains is shown.The dashed, red line indicates the chance predictive value of 0.5 associated with binary classification.rsfMRI and the SVM analysis considered here as the accuracy is sensitive to the age and feature group.A natural caveat exists that ML with other brain imaging techniques such as MRI may be used to predict sex very accurately as done in (Sanchis-Segura et al., 2022).Furthermore, the presented findings may not directly translate to other algorithmic techniques, different recording modalities (e.g.MRI), or perhaps other fMRI datasets.An algorithmic technique such as deep learning may sufficiently capture non-linearities in the considered dataset and provide more accurate sex prediction results, but, due to peculiarities in its training examples, potentially overfit a different dataset to provide marginal predictive capability.The SVM classifier was considered in this work due to its simplicity, general robustness, and interpretability.A comparative study with more advanced methods is a future avenue of investigation.
Functional connectivity studies with rsfMRI have provided evidence for a dedifferentiation theory of aging where large cross-correlation values among ROIs were found to become more diffuse and less network-concentrated across the brain with increasing age (Siman-Tov et al., 2017;Fair et al., 2009).While also encompassing the entire brain, the number of paroxysmal events per-ROI is a more localized measure of the BOLD signal that does not reflect network connectivity.The paroxysmal events associated with each ROI are not computed with respect to a global reference, but rather with respect to the BOLD signal of the ROI.We investigated whether the resting-state brain becomes increasingly bursty with age.The youngest group (21-30) was taken as the reference Table 3 Identification of brain regions that are the most reflective of aging for the extrema and median group studies.The five ROIs with the highest AR and their respective network are listed from left-to-right.The extrema groups consist of differentiation between subjects 21-30 and 71+ years old while the median groups entailed 41-50 and 51-60 year old subjects.The ROIs in red denote regions that have provided a higher differentiation than that attained when considering the entire brain.when conducting a 2-sample t-test among ND low, ND high, and the combined number of paroxysmal events computed across the ROIs of each subject.In Fig. 7A significant decreases in the number of paroxysmal events were noted when considering the age group of 41-50 and higher (p < 10 − 4 ).The standard deviation (s.d.) in the number of deviations per subject also showed a decreasing trend with increasing age.In considering the same analysis separately for male (Fig. 7B) and female (Fig. 7C) subjects, the reduction in burstiness with increasing age occurred much earlier in females (41-50) than male (61-70) subjects.The same decrease in the s.d. of the deviations with increasing age was Fig. 7.The burstiness of the BOLD signal is evaluated between subjects at various chronological age groups.The paroxysmal features, in other words, ND low (low), ND high (high), and the sum of the two quantities (total) are averaged across all subjects in the respective age groups.The youngest age group (21-30) is taken as the reference when conducting 2-sample t-tests to determine significance (p < 10 − 4 ) of the change in the number of paroxysmal events of subjects with increasing chronological age.The s.d. was calculated to assess the relative variability in the measure.The decreasing trend in burstiness and the s.d. of the number of paroxysmal events with increasing age is seen for A) all subjects as well as separately in B) male and C) female subjects.
noted between male and female subjects, although the decrease is more pronounced across females.In the three considered scenarios of Fig. 7a similar reduction in the number of total, high, and low paroxysmal events is noted.This indicates that the trends encompass a comprehensive account of the burstiness.In summary, the per-ROI measure BOLD signal burstiness decreases with age at different rates between male and female subjects.

Recording center influence on predicting brain age
The considered dataset entailed 17 centers across continents with each recording facility having its own scanner and demographics.Additionally, the sampling rate, number of slices, and voxel size were not consistent across any two centers.It is natural to inquire whether it is possible to decipher the center from the recordings.The question may be asked in an alternative manner of whether a recording center maintains a signature in the BOLD signals that are amassed in forming a crosscenter dataset.A negative answer to this question would be beneficial in indicating that the features and neural signals are unaffected by the center where the recording was made.Our analysis for predicting comparative brain age was repeated while restricting attention to subjects from the NKI-Rockland (NKI-RS) center.The center was selected because it contained the largest number of participants as well as the largest spread of subjects' chronological ages.Although NKI-RS contained approximately 34 % of the subjects considered in the analysis, this number is significant when considering the remainder of the recordings come from 16 centers with the next largest participant pool (i.e.Beijing) contributing approximately 13 % of the total.Table 2 contains an itemization of the number of subjects in NKI-RS that belonged to the considered age groups.
The leave-one-out with MC sampling pipeline in Fig. 1 was applied to the 307 subjects from the NKI-RS center.A comparison of the results to those attained for all centers will indicate whether the findings exist on a per-center basis.In Fig. 8A it is noted that with the combined and statistical features, the highest overall accuracy (AR = 0.722 and 0.739, respectively) was seen in differentiating the youngest from the most aged group.The statistical features provided the highest accuracy in 8/ 15 comparisons while the paroxysmal and combined features provide the highest AR in 4 and 3 cases, respectively.Surprisingly, the paroxysmal features provide their best reliability in differentiation between the 21-30 and 31-40 age groups while yielding below-chance performance when distinguishing between the extrema age groups.The comparative age-gap plot in Fig. 8B shows similar trends in differentiability to what was noted in Fig. 2B when subjects from the centers were aggregated.In other words, with the combined and statistical features the accuracy in differentiating among younger subjects is higher than that of differentiation among older subjects that are separated by the same number of years.Specifically, with no age gap the highest AR was noted between the 21-30 and 31-40 groups (0.7, combined features), while for 10, 20, and 30 year gaps, the highest ARs are noted between the 21-30 and 41-50 (0.677, statistical features), 21-30 and 51-60 (0.745, statistical features), and the 21-30 and 61-70 groups (0.6525, combined features), respectively.Similar to what was noted with the subjects aggregated among all recording centers, Fig. 8B shows that the Δ value in the age spectrum located towards the youngest subjects (gap 1) yields the highest number of significant ARs.It is interesting that the feature groups showed increased variability across the 15 cases when comparing the ARs attained with subjects at all centers (Fig. 2) to the ARs for subjects at NKI-RS (Fig. 8).An increase in variance of 38 %, 17 %, and 119 % was noted for the combined, statistical, and paroxysmal features.
We compared the accuracy in age prediction with data aggregated from all centers to a single recording center (i.e.NKI-RS).It was observed that the classification capability is generally increased by pooling subjects across the centers (Table S2).Except for three of the scenarios where the groups were not separated by more than 10 years (no gap 1, no gap 3, and no gap 5), the differentiation of the 51-60 age group from the 71+ group (10-y gap 4), and the differentiation of the 41-50 from the 71+ group (20-y gap 3), the pooling of subjects across centers provided superior results to a within-site consideration.However, the group of features did have a role in this result, since the performance with the paroxysmal features was less improved by data aggregation.In fact, there was only one scenario (no gap 4) when classifying between groups with less than a 20 year gap resulted in a significant improvement when all centers were included in the analysis (Table S2).
We assessed whether the aggregation of data between centers obfuscates the differences that may exist between the male and female brains.The ML pipeline of Fig. 1 is applied to each of the six age groups of the NKI-RS subjects.It is noted in Fig. 8C that the accuracy of identifying a female brain is in general higher than a male brain.This holds for all age groups when considering the statistical and combined features and in 3/6 cases with the paroxysmal features.Such results are consistent with those noted in Fig. 6 with subjects pooled from all centers.Conversely, the 21-30 age group that exhibited poor accuracy in the differentiation of sex with the pooled subjects show an AR greater than 0.6 with all three feature groups.Interestingly, all of the feature groups reflect an indistinguishability between male and female brains for 61-70 year-olds.Although more pronounced for the NKI-RS center, a similar trend was noted when considering all subjects in Fig. 6 as it is not possible to accurately distinguish between male and female subjects in the 61-70 age group.In summary, the rsfMRI data and ML analysis reveals differences of varying degrees between male and female brains at a single centerthe overall distinguishability, however, was still not impressive.
A different analysis was undertaken to investigate whether it is possible to differentiate between recording centers via the neural signals of the subjects at the centers.While studying a subject's neural fingerprint via fMRI has been extensively explored, the question of recording centers' fingerprint has not been scrutinized.The classification was pursued independently for young (i.e.21-26 year old) and aged (i.e.44-65 year old) data in order to circumvent the subjects' chronological ages being a determining variable.A negative finding would be encouraging from the perspective of the analysis being agnostic to the recording center.The Beijing (N = 119), Cambridge (N = 96), and NKI-RS (N = 54) centers are considered in the analysis of young subject data.The analysis was repeated with aged neural recordings considered from the NKI-RS (N = 145), Milwaukee (N = 43), and COBRE (N = 21) centers.The number of subjects in the young and aged category are shown in Table 4 and three combinations of pairwise analyses are considered via the pipeline in Fig. 9A.The AR computed across the heldout subjects are shown in Fig. 9B.The identity of a recording center is highly conspicuous.The capability to distinguish the center for young and aged brains exceeded 0.9 for at least two feature classes in each of the three combinations.The distinguishability was marginal (AR < 0.6) with only the paroxysmal features of the young subjects between the NKI-RS and Cambridge centers.In summary, the analysis indicates that the recording center can be identified at relatively high accuracy and thus introduces a fingerprint when conducting a pairwise comparison.

Discussion
Extraction of information from the time-varying dynamics of brain regions is a fundamental facet of neuroscience.Inferences from BOLD signals have typically involved correlational analysis with pairs or groups of ROIs considered.An ROI-centric feature calculation is different from the conventional FC approach of computing correlations between brain regions and using their normalized or transformed versions as features.While the implications and neural mechanisms at work between the per-ROI analysis and more conventional FC are different, the former provides several advantages.It can be verified that the number of features that must be computed and used in the ML training is a fraction of that necessary for an FC analysis (approximately 1/3 less with the combined features).This savings in computation is particularly significant for considering thousands of subjects and more fine-grained parcellation methods that yield hundreds (rather than 42) of ROIs.The non-FC analysis provided insight from the perspective of identifying implicated brain regions rather than relative network connections involved in a phenomenon.While the comparative utility is debatable and perhaps an unresolved avenue, it cannot be discounted (Shafiei et al., 2020;Fallon et al., 2020;Damoiseaux et al., 2006).The three references provide a unified description of BOLD recordings, the statistical properties of brain regions' time-series, and FC.
The ROI-based analysis is more amenable to the potential discarding of an ROI that may be deemed noisy or dysfunctional.Such an ROI could contain voxels that have very low signal-to-noise ratio (SNR) values, or may be associated with a region of the head that experienced excessive motion during a recording.In the case of correlation-based FC, potentially numerous entries in a connectivity matrix would need to be discarded whereas the ROI analysis would involve the removal of features associated with a single unit.For instance, a simple calculation shows that the case of 3 out of 42 ROIs being discarded would lead to approximately 7 % of the features being discarded with the per-node analysis.This is rather favorable in comparison to the roughly 14 % of discarded FC features that would be necessary.Also, the identification of an excessively noisy ROI is likely to be easier when examining features computed directly from ROIs than correlational values of the ROI with other regions.
The difficulty in differentiating older brains in comparison to younger recordings was noted in a prior work that used several ML techniques with anatomical brain measurements from 3144 subjects spanning an age range from 8 to 96 years (Valizadeh et al., 2017).Similarly, Aycheh et al., (2018) considered various regression analyses on anatomical data from older subjects who spanned 45-91 years (Aycheh et al., 2018).The accuracy in the predicted brain age and the biological age was not very impressive in comparison to previous works.The authors conjectured that this may have been due to considering older subjects whose anatomical structure provided more complexity and hence increased variance.With respect to rsfMRI recordings, Song et al., considered a group of young (18-35 years) and older (55-85 years) adults and compared the reliability of the test-retest collected FC between the two groups (Song et al., 2012).Statistical tests and regression analysis were used, and the results reflected an age-related difference with the older subjects having reduced reliability and stability in FC.The aforementioned studies have motivated the complexity and several of the intricacies of deciphering chronological age from brain recordings.
The ML pipeline in our presentation encompassed the division of subjects into six age groups and a pairwise classification analysis with the training and testing steps repeated for each pair of age groups.Although this deviates from the conventional methodology of having age as a continuous variable, it enabled a categorical analysis between chronological age and sex by focusing the analysis on specific age groups

Table 4
The number of subjects from three centers used to determine if it is possible to differentiate the center where the recordings were taken via the BOLD signals' features.The top table provides an itemization of the young subjects from three centers while the bottom table provides the corresponding information for aged subjects.within the full spectrum.The choice to use an SVM as the machine learning technique rather than a large-scale neural network method such as a DL architecture is justified from works such as (He et al., 2020).The work found similar performance of three DL techniques to kernel regression in predicting age, sex, and intelligence from resting-state FC patterns while requiring significantly lower computational costs.The use of an SVM within this context has also been motivated by other works such as (Dadi et al., 2019).
Considering the results attained for predicting age and sex from brain recordings, we note comparable performance among the statistical and combined features when a common ML pipeline is applied.The paroxysmal features generally did not lead to a differentiability that is better than that attained with the other two feature groups.Nevertheless, they comprise a subset of the combined features and have a physical interpretation as a measure of the burstiness of the activity at an ROI.The distribution of paroxysmal features for each subject was considered to assess whether our trends were due to the presence of a relatively few extreme outliers (Figure S4).While outliers were observed in four of the six age groups, they did not deviate extremely from the values noted across all subjects of various ages.The two age groups void of outliers were the two youngest groups.This is in accordance with the finding of the younger subjects being more easily differentiable than the older subjects.
It is challenging to compare the predictive accuracy presented here to prior literature because of data and preprocessing differences as well as the novel classification-based approach that was used.In fact, direct comparisons are largely non-existent due to the heterogeneity of the studies.For instance, we were unable to compare to a seminal work such as (Dosenbach et al., 2010) that attained an AR exceeding 0.9 since they solely considered the discrimination of subjects 7-11 from 24 to 30 years Fig. 9.A quantitative assessment of a recording center's prospective fingerprint in the various features computed from its subjects' rsfMRI recordings.The subjects were selected to be relatively age-matched when performing a young (21-26 years) and an aged (44-65 years) analysis.A) An ML pipeline for testing the influence of the recording center on the results.An SVM is trained on 2/3 of the minimum number of samples from the two centers that are considered with ⌈.⌉ denoting the ceiling function.The residue samples from both centers are used as the test data.The procedure is repeated in Monte Carlo (MC) fashion to arrive at a mean accuracy rate.B) The capability to distinguish recording centers based on the neural recordings from their respective subjects.In each of the three pairwise comparisons the ML pipeline was applied 1000 times to arrive at an AR in determining whether the center could be identified from the recordings.old.Our findings are comparable to (Vergun et al., 2013) since the authors performed age prediction by SVM and SVR analyses on rsfMRI data.They concluded that olderparticularly middle agebrains were more difficult to predict (Fig. 6 in (Vergun et al., 2013)) thus agreeing with our results.Although an AR of 0.84 was reported, only one classification was performed between a young (19-37 years) and old (61-85 years) group of subjects.In light of the 24-66 year age gap, the AR is comparable to what we have attained in Fig. 2. The authors also noted that sex classification was not nearly as accurate as age prediction which agrees with our findings and that of (Allen et al., 2011).A recent work noted a ~5.5 year mean absolute error (MAE) in prediction when using regression to predict ages of subjects across a broad spectrum (22 to 97 years old) (Zhou et al., 2023).Other works have found similar MAE values such as 5.2 years (Cole, 2020).These results are not immediately comparable to our results because of the different metrics used, however, they are similar with closer scrutiny.For instance, the lowest ARs that we have reported are for subjects less than 10 years apart in chronological age (Fig. 2).A non-negligible MAE (e.g.~5 years) would be expected to have an increasingly pronounced effect on the classification as the age gap between subjects decreases.Thus, we expect that AR values of chance (i.e.0.5) to 0.6 would be attained if the related works were to classify subjects that are less than 10 years apart.Similarly, the effect of the MAE in the related works would be less pronounced if the techniques were to differentiate subjects with larger age gaps.The differentiation in such cases would correspond to our AR values that exceed 0.7 and 0.8 in Fig. 2 and Fig. 8.It is also infeasible, and perhaps counter-productive, to compare the results to prior brain age prediction works such as (Pervaiz et al., 2020) that express accuracy as a correlation between subjects' chronological ages and their predicted age.Nevertheless, it is important to note that in this work the intent was not to optimize AR during age prediction, but rather to perform a comparative analysis between discernability as a function of age, sex, and different ROI-based features.
Comparing our sex prediction results to existing works is more feasible since the outcome is categorical and its predictive accuracy is usually expressed via a percentage.In the work of Casanova et al.,. (2012) (Casanova et al., 2012) the highest accuracy rate for distinguishing 21 year-old males and females via rsfMRI recordings was 0.65 which is nearly identical to what we noted when considering combined features at a single (NKI-RS) recording center (Fig. 8C).An important caveat is that (Casanova et al., 2012) considered a single age while our analysis consisted of subjects 21-30.Although the authors in (Weis et al., 2020) evaluated subjects of the same ages as us, the sex classification question was investigated across the entire age spectrum rather than on a more fine-grained (e.g.per-decade) level.The AR spanned the 0.6-0.7 interval which is comparable to the results here.However, significantly more processing was used in (Weis et al., 2020) since, rather than a single ML algorithm, a separate SVM was trained and tested on every brain parcel of a subject.In Ritchie et al., (2018) statistical tests were used to distinguish males and females of ages 44-77 years via rsfMRI recordings (Ritchie et al., 2018).It was found that several signals were more prominent at various brain regions in one sex over the other.The authors reported their findings to agree and disagree with the literature, thus pointing at the heterogeneity of the analyses, pipelines, and results.
In the present work the entire brain recording was used to differentiate sex.It is conceivable that certain regions can provide higher accuracy than we have reported in Fig. 6 and Fig. 8.An extension to this work is to pursue sex differentiation on a per-network or per-ROI basis.Works such as (Joel et al., 2015;Sanchis-Segura et al., 2020) have considered more limited age ranges and foundvia MRI recordingsthat male and female brains could not be differentiated very accurately.This was apparently in disagreement with studies that have reported relatively high ARs.However, as shown in (Sanchis-Segura et al., 2020), the high ARs were reduced to approximately 0.6 when total brain size was properly controlled for.Although not infeasible, it is unlikely that the resting state BOLD signal would introduce information about a subject's sex that is not already encoded in MRI and anatomical recordings (Cole, 2020).Thus, the ARs presented in this work with rsfMRI, per-ROI features, ML pipeline, and per-decade division are comparable to what has been reported in the literature.Lastly, we remark that our findings support the observations in (Scheinost et al., 2015;Cai et al., 2020) by showing that the differentiation of sex can vary significantly along the age spectrum.The two references considered statistical analyses while Fig. 6 and Fig. 8C present this result via ML and prediction.
There is a premium on using dimensionality reduction techniques to derive conclusions from data.It is customary to use principal and independent component analysis (PCA and ICA) to parse what may be sparse sets of high dimensional neural data.However, the techniques have various drawbacks such as requiring unjustified assumptions on the statistical properties of the data and their resultant outputs not being as interpretable as the inputs.We have scrutinized entities by isolating them during the analysis.This has led to reducing dimensionality by considering individual network and ROI contributions to brain aging.It is important to be cognizant of the aging dynamics of brain regions because the corresponding ROIs may constitute associated functions that are the most prone to being compromised.With respect to the predictive capability attained with single ROIs on comparatively young and aged subjects it was noted that, in general, the trained machine provided very good accuracy for one age group while performing exceptionally poor for the other age group.It may be that cooperation between brain regions persists in the portrayal of a phenotype as fundamental as chronological age.However, the results indicate that the cooperation may be destructive or interfering since single ROIs can serve as biomarkers.The presented results are novel in providing per-network, predictive capability of brain age across several classes of features at a detail that did not previously exist.The comparison of seven networks across three feature groups sheds light on the networks that are altered by or reflective of aging.The use of ML to study whether single ROIs can serve as a more accurate biomarker of chronological age than having access to the entire brain yielded surprising results.Namely, the most predictive single ROIs were highly competitive biomarkers in reflecting aging in comparison to the corresponding features recorded across the entire brain.
While prior works have investigated sex prediction across different chronological age ranges, the question of during which decades restingstate signals from male and female brains appear the most similar or different has not received attention.By using a classification approach to investigate if the BOLD signal characteristics can distinguish sex at various ages, it was noted that the accuracy in identifying a female brain was higher than identifying a male brain.Holistically, the results revealed the likelihood of male and female brains looking similar at rest since it was not possible to reliably predict sex based on rsfMRI.While prior fMRI studies have provided evidence for cross-correlation values between the BOLD signal of various ROIs becoming more evenly distributed across the brain with increasing age (Siman-Tov et al., 2017;Fair et al., 2009;Chan et al., 2014), little was known about the paroxysmal nature of the BOLD signal's alterations with aging.We observed significant decreases in the burstiness of BOLD activity in subjects 41-50 years and higher.The variation in the burstiness also showed a decreasing trend with increasing age.This effect was noted to occur earlier and be more pronounced for females.The results further vindicate the wisdom of not generalizing across sex in neuroimaging studies.It is productive to hypothesize on the significance of the noted alternations in burstiness.Prior works have shown that in healthy subjects, individual vessel flows, total cerebral blood flow, and variability in flow in distal vessels decline with age (Amin-Hanjani et al., 2015;Stoquart-ElSankari et al., 2007;Ainslie et al., 2008).The findings herein serve as further evidence since a decreased flow would lead to a lower possibility for large, abrupt variations in the BOLD signal.The decrease in cerebral circulation with age will have definite effects on neural activity and may be a precursor to neurodegeneration or dementia.
Interestingly, a greater rate of decrease in cerebral blood flow velocity has been reported among females with increasing age (Alwatban et al., 2021;Tegeler et al., 2013).It is feasible to concoct this reflecting the higher rate of decrease in burstiness that we have noted in females.The implications of sex on paroxysmal properties of brain signaling may be indicative of metabolic and hormonal differences between males and females during aging.Such complex biological processes have been shown to be nonlinear and it is conceivable that they resonate gender differences in neurovascular signaling.
In considering the results for determining chronological age and sex from brain recordings we noted comparable classification performance among the statistical and combined features with the ML pipeline.The paroxysmal features generally did not lead to a differentiability that is better than that attained with the other two feature groups.This points towards the likelihood of male and female fMRI recordings looking similar at rest.When performing the classification analysis on subjects from a single recording center, similar trends appeared to those when aggregating subjects across all centers.Namely, for the combined and statistical features the accuracy in differentiating between younger subjects was higher than that of differentiation among older subjects separated by the same number of years.The trends also persisted in examining the differences between male and female brains at a single recording center.The accuracy of identifying a female brain was higher than a male brainan overall distinguishability, however, was not noted.The comparison of age prediction with data aggregated from all recording centers to a single center showed that performance increased by pooling subjects (Table S2).Although this result held for the three classes of features, the improved accuracy was less apparent for the paroxysmal features.
A novel methodology was used to investigate if a recording center can be identified based on the neural signals of the subjects at the center.Our use of ML to conduct a recording center fingerprint study provided discouraging findings since the rsfMRI signal properties were not agnostic to the recording center.The identity of the center was highly conspicuous for a young and an older group of subjects.The findings are important in the discussion of whether fMRI recordings should be treated uniformly and irrespective of the center even if nearly identical experimental protocols have been followed.Our methodology can be implemented in any recording paradigm where experiments are conducted at multiple locations with the intent of results being agnostic to the site.The salience of the fingerprints of the considered recording centers presents a potential caveat to the interpretability of our findings on age and sex prediction.However, it is important to note that the age and sex prediction each considered different labels in the ML analysis as well as additional recording centers than were used in the fingerprint study.In discriminating between subjects of relatively close ages but from one of two centers, an ML algorithm is not necessarily using similar mechanisms as an ML algorithm that was trained to discriminate between disparate age groups (e.g.20 year gap) of subjects from a larger number of recording centers.A further motivation for considering the NKI-RS center separately when conducting the age and sex prediction was to compare the results and trends to what we attained with data aggregated from all centers.The findings were similar.Although there was an increase in the variance of the ARs attained with NKI-RS, this may be attributed to the higher dynamic range in predictive capability offered by the aggregating of subjects from different centers.The relative similarities mitigate the notion that the model is using site-specific features to enhance age or sex prediction.
The contributions of this work are listed as follows.We have proposed an ML-based methodology, referred to as an age group pair classification task.It bins subjects into age groups and uses ML to measure group difference based on the classification accuracy between each pair of groups.An SVM was used with a leave-one-subject-out scheme as the cross-validation method to generate predictions.Following the processing of the BOLD signals on a per-ROI basis, three groups of features were formed: statistical, paroxysmal, and combined.
The use of ML for age prediction via a classification-based approach led to an observation that the statistical features generally provided the highest accuracy rate.We note that there is decreased differentiation of neural signals from older subjects that are separated in age by the same number of years as younger subjects.The ML analysis was also undertaken individually on seven brain networks comprising distinct groups of ROIs.The results are compared with the predictive capability that we attained with the entire brain as well as an analysis that was done on a single ROI basis.A byproduct was discovering the most predictive single ROIs to differentiate among various age groups.Sex classification within each age group was also performed by the ML to answer a question of during which decades resting-state signals from male and female brains appear the most different.In general, it was discovered that it is not possible to reliably predict sex based on rsfMRI since the accuracy is sensitive to the age and feature group.The paroxysmal events, which quantify the burstiness of the BOLD signal at an ROI, were compared between each age and sex group.It was concluded that the number of paroxysmal events change through the aging process in different rates between males and females.
The presented study had several limitations.First, recordings from subjects who had their eyes open were combined with those who had eyes closed.This is not uncommon among related works.In addition to (Siman-Tov et al., 2017), relatively recent studies have considered multiple datasets with a mixture of age-matched subjects with eyes open and eyes closed during the rsfMRI data collection to conduct a regression analysis (Gbadeyan et al., 2022;Goni et al., 2014).For sex prediction Weis et al., (2020) used multiple datasets consisting of recordings where there was a mixture of subjects with eyes open and closedthe AR between the datasets were not significantly different (Weis et al., 2020).Nevertheless, the combining of data from the two conditions is a limitation of this work.Separate ML analyses of the subjects with eyes open and eyes closed is an avenue of future investigation.The subsequent comparison of the results would assess if the findings generalized.Second, in light of our quantification of a recording center's prospective fingerprint, the use of harmonization techniques such as those used in (Yu et al., 2018;Pomponio et al., 2020) should be explored as a means of determining whether the presented results are altered.Third, we have not optimized or conducted a broad search on the feature space.The consideration of other features/metrics may produce different results and constitutes an avenue of future research.An initial undertaking may be a comparison of the presented per-node features with FC measures computed from correlation coefficients.This analysis has been undertaken with Pearson correlation coefficients (Sorooshyari, 2023).Interestingly, the results consist of cases where the considered ML pipeline with FC features had superior predictive capability, while in other scenarios the three per-ROI classes of features yielded a higher AR.The thorough comparative account will be a future work that is under preparation.
In conclusion, we have departed from the canonical analysis of evaluating the effects of brain aging via FC to examine features computed on a per-ROI basis.The extracted features were used in a new, classification-based ML pipeline to predict brain age and sex.The findings shed light on the properties of the rsfMRI recordings that are most modulated by aging.It was found that aged brains appear more similar than younger brains that are separated by the same number of years.Increasingly detailed results are presented by identifying the brain networks and individual ROIs that are the most affected by aging.Interestingly, the findings were largely conserved when the analysis was repeated on a constituent center.The ML analysis also discovered the presence of a recording center fingerprint since a center could be identified based on the neural signals of the subjects at the center.The pernode analysis provided insight beyond what is typically attained with CPs by scrutinizing various features' characteristics.The paroxysmal properties of the recordings indicated a decrease in the burstiness with aging.The rate and variability of the decrease were notably different between males and females.

Declaration of competing interest
The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Fig. 4 .
Fig. 4. The individual brain networks' reflection of chronological age.In each comparison the ML pipeline was applied to individual networks to attain an AR.The network that provided the highest AR is shown for the three classes of featuresthe results for all networks are shown in Figure S3.AR values with the A) combined, B) statistical, and C) paroxysmal features are shown for all subjects.

Fig. 6 .
Fig. 6.The effect of sex on the chronological age dynamics of the neural signals processed by the ML analysis.The accuracy in determination of sex from rsfMRI brain recordings when considering the A) combined, B) statistical, and C) paroxysmal features with the ML pipeline of Fig. 1.For each feature class a black (red) square above the aggregate bar denotes the age group for which sex was most (least) accurately distinguished.The dashed, red line indicates the chance predictive value of 0.5 associated with binary classification.

Fig. 8 .
Fig.8.A comparative evaluation of the capability to distinguish between young and aged subjects at a single recording center (NKI-RS) via the ML analysis and the three feature groups.The ML pipeline was separately applied for the combined, statistical, and paroxysmal features to attain the subfigures.A) The ARs attained for distinguishing subjects in one of two age groups indicated via the x and y axes.B) A comparison of the ARs in differentiating young from aged brains with the three classes of features for the 15 different combinations of groups.The scenarios are ordered from those that are the closest proximity in chronological age (i.e.no gap) to the furthest apart (i.e.extrema).The binarized heatmap identifies the ARs that are statistically significant (grey-colored) for a Δ value and a feature group.C) The accuracy in determination of sex from recordings of subjects in the NKI-RS center when considering the three classes of features.For each feature class a black (red) square above the aggregate bar denotes the age group for which sex was most (least) accurately distinguished.The dashed, red line indicates the chance predictive value of 0.5 associated with binary classification.

Table 2 (
Top) An itemization of the number of subjects per age group.The data spans recording centers.(Bottom) The distribution of chronological ages for the subjects at the NKI-RS center.The itemization is provided as part of a cross-center consistency evaluation aimed at assessing the agreement of results with subjects pooled among all centers to the analysis performed on a constituent center (i.e.NKI-RS).