Large-Scale Structural Covariance Networks Predict Age in Middle-to-Late Adulthood: A Novel Brain Aging Biomarker Free

Abstract

Introduction

Aging is a complex and dynamic process that is accompanied by cumulative age-related damage affecting various organs including the brain (Holliday 2004). Such organ damage may conduce to adverse behavioral and clinical outcomes, such as cardiovascular disease, cognitive decline, and increased risk of neurodegenerative diseases (Hindle 2010; Toepper 2017; Fajemiroye et al. 2018). Although these clinical conditions mostly manifest in later life, associated changes may present subtly years before clinical symptoms become evident (Wu et al. 2011; Fiandaca et al. 2014). Considering the irreversible natural courses of these debilitating diseases, more research has been focused on the feasibility of early detection of changes in organs as a potential biomarker for disease surveillance before disease onset, especially during middle to late adulthood (Sutphen et al. 2015; Chiu et al. 2017). Hence, investigating the best method to characterize individual health on the basis of the signature of organ changes during this critical period is a clinically important objective.

One promising signature of aging-related changes is the brain tissue integrity, which is measured by structural magnetic resonance imaging (MRI). Previous anatomical MRI studies have identified several aging-associated changes, including decreased brain gray matter volume (GMV) or/and cortical thickness, loss of tissue contrast, and impaired white matter (WM) integrity in healthy population (Sowell et al. 2003; Salat et al. 2009; Westlye et al. 2010; Lemaitre et al. 2012; Storsve et al. 2014; Cox et al. 2016). In addition, findings of both cross-sectional and longitudinal studies demonstrated that these well-identified aging-associated brain imaging patterns may be altered in patients with neuropsychiatric or neurodegenerative diseases (Driscoll et al. 2009; Nenadic et al. 2012; Fjell et al. 2013). These studies further proposed the concept of the “accelerated brain aging” may be an aging-related marker for differentiating patients from healthy controls, it could serve as a potential biomarker for evaluating neurological diseases. Hence, modeling these complex aging-related profiles of the human brain into a single quantitative index has attracted considerable attention in the past decade.

Biological age is a unitary quantitative index that can be estimated on the basis of biomarkers; this embodies the complexity of aging process for a given individual (Levine 2013; Belsky et al. 2015). Recently, several large cohort studies, which employed machine learning-based approaches to analyze the relationship between brain structural changes on MRI and chronological age in healthy subjects, yielded reliable imaging biomarkers of brain biological age. These voxel-wise or region-wise measures of GMV and WM integrity could not only accurately predict the chronological age but also act as potential biomarkers for investigating different kinds of clinical populations and public health issues (Franke et al. 2010; Brown et al. 2012; Erus et al. 2015). Using these neuroimaging-based brain age biomarkers, studies have also demonstrated the phenomenon of accelerated brain aging in several neurological diseases such as Alzheimer’s disease (AD), mild cognitive impairment (MCI), and schizophrenia (SCZ) (Franke et al. 2010; Gaser et al. 2013; Koutsouleris et al. 2014; Schnack et al. 2016). In addition, a link between increasing brain biological age and mortality has also been demonstrated (Cole et al. 2017). Therefore, measuring brain age might be potentially useful for identifying individuals’ risks of poor health outcomes in clinical settings.

Recent developments in our understanding of the human brain have led to the realization that both the structural and the functional aspects of the brain are organized as large-scale networks (Bullmore and Sporns 2009). Through a complex network architecture, these spatially distinct brain regions can perform a board range of cognitive functions and reflect individual differences in behavioral performance (Petersen and Sporns 2015). Structural covariance network (SCN) analysis was recently proposed as a surrogate approach to characterize large-scale brain structural network on the basis of the shared covariance of the brain morphologic features across study participants (Alexander-Bloch, Giedd, et al. 2013a). Although neurobiological mechanisms underlying large-scale SCNs remain obscure, recent studies have pointed out the SCNs may reflect heritable developmental coordination or synchronized maturation across distinct brain areas, in addition to being highly similar to gene expression profile and intrinsic functional network architecture of the human brain (Schmitt et al. 2008; Seeley et al. 2009; Zielinski et al. 2010; Alexander-Bloch, Raznahan, et al. 2013b; Sotiras et al. 2017; Romero-Garcia et al. 2018). On the basis of these fundamental properties, several studies have demonstrated the feasibility of using SCNs to investigate network-level changes in healthy aging populations and populations with neurodegenerative and neuropsychiatric diseases (Hafkemeijer et al. 2014, 2016; Gupta et al. 2015; Li et al. 2019). Such accruing evidence has suggested that large-scale SCNs could provide an alternative way to infer network-level brain changes; however, there has been limited research on using this network information to evaluate individual brain health in terms of brain age. Considering that the brain operates like an interactive network system and that SCNs have potential clinical utility, we hypothesized that analyzing large-scale brain structural networks with SCN approach may provide supplementary information for predicting individual brain age during the middle to late adulthood.

To fill the gap of predicting individual brain age from large-scale anatomical network perspective and to overcome the limitation of previous structure-based brain age studies, that is limited sample size of participants in middle-to-late adulthood, the present study aimed to develop a novel brain biological age estimation framework by integrating large-scale SCN analysis and machine learning approach in a large cohort of middle-aged and older adults. We hypothesized that this network-level framework could not only accurately predict individual brain age but also reveal accelerating brain aging pattern in neurodegenerative and neuropsychiatric diseases including AD, SCZ, and major depressive disorder (MDD). Furthermore, we believe that our results would provide a novel insight into how the brain changes with aging from a network perspective.

Materials and Methods

General Description of Participant Demographics

To construct a robust SCN-based brain age predictive framework with an appropriate sample size, the anatomical MRI scans of participants in middle-to-late adulthood were obtained from two large-scale on-going projects in our research group: the I-LAN Longitudinal Aging Study (ILAS) project and the Taiwan Aging and Mental Illness Neuroimaging (TAMIND) project. To meet our research objective, we selected only healthy participants aged 50–90 years from these two projects. For further clinical application, anatomical MRI scans of patients with neuropsychiatric (SCZ and MDD, part of TAMIND project (Guo et al. 2018; Rolls et al. 2018) and neurodegenerative (AD) diseases were obtained from our previous clinical brain imaging studies and on-going longitudinal study, respectively. To match the age distribution with that of healthy participants, we only included patients aged from 50 to 80 years. All study protocols were approved previously by the Taipei Veteran General Hospital (TVGH) and National Yang Ming University (NYMU) Ethics Committees on Human Research. All participants or their legally designated proxies provided written informed consent for each aspect of the study. A brief description of study participant characteristics is as follows: (1) healthy participants—721 from the ILAS project (age, 60.56 ± 7.3 years; 324 men, 397 women) and 188 from the TAMIND project (age, 68.85 ± 9.9 years; 99 men, 89 women); (2) clinical population—26 patients with SCZ (age, 56.0 ± 4.5 years; 12 men, 14 women), 30 patients with MDD (age, 61.5 ± 8.2 years; 13 men, 17 women), and 19 patients with AD (age, 72.6 ± 5.8 years; 6 men, 13 women).

Healthy Participants in Middle-to-Late Adulthood—The ILAS Dataset

The ILAS project is a community-based aging cohort study conducted in the I-Lan County of Taiwan that aimed to explore the inter-relationship among geriatric syndromes, cognitive functions, and brain structural and functional changes (Lee et al. 2013). In brief, community-dwelling adults aged ≥ 50 years from Yuanshan Township in I-Lan County were invited to participate in the study through household registrations by the government. Multimodality MRI, comprehensive cognitive assessments, blood samples collection, physical measurements, and personal health status evaluation were performed. Any participant who met any of the following criteria was excluded from the study: (1) inability to communicate and complete a clinical interview, (2) presence of an illness, such as malignancy, heart failure, or renal failure, with limited life expectancy (<6 months), (3) inability to complete a simple motor task (e.g., a 6-min walk test) or cognitive tasks (verbal memory, language, visuospatial, and executive function tests) because of functional disability, (4) institutionalization, (5) presence of contraindications for MRI such as ferromagnetic foreign bodies or metal implants, and (6) a global cognitive impairment (a Mini-Mental State Examination (MMSE) (Folstein et al. 1975) score of <24 in well-educated subjects (education years ≥6) or <14 in less-educated subjects (education years < 6) (Sun et al. 2014)). In addition, participants diagnosed with neurological and neuropsychiatric diseases were also excluded.

Healthy Participants in Middle-to-Late Adulthood—The TAMIND Dataset

The TAMIND project is a longitudinal brain imaging-based study, which aimed to investigate the relationship between several mental disorders and brain structural and functional changes across the entire period of adulthood (Huang et al. 2018). Healthy participants were recruited from local communities in the Northern Taiwan. The exclusion criteria for enrolling healthy participants were as follows: (1) inability to communicate in a clinical interview and complete the battery of cognitive tests, (2) presence of neurobiological or any Diagnostic and Statistical Manual of Mental Disorder-IV (DSM-IV) Axis I psychiatric disease, such as dementia, stroke, brain tumor, alcohol/substance abuse, psychotic disorder, (3) presence of medical illness, such as diabetes, malignancy, heart failure, or renal failure, (4) an MMSE score of <24 in all participants and a Clinical Dementia Rating (CDR) score of >0.5 in participants aged ≥65 years, and (5) presence of contraindications for MRI such as ferromagnetic foreign bodies or metal implants.

In summary, the analytical cohort, which comprised 909 healthy participants (486 women) aged 50–89 years, was used for constructing the large-scale SCN-based brain age estimator. Furthermore, to obtain an unbiased brain age estimator and to evaluate its generalizability, we randomly allocated healthy participants, obtained from two large-scale on-going projects (ILAS and TAMIND), into a training dataset (N = 800; age, 62.3 ± 8.7 years; 371 men, 429 women) and a testing dataset (N = 109; age, 62.0 ± 7.6 years; 52 men, 57 women) (matched for age and sex).

Clinical Datasets with Neurodegenerative and Neuropsychiatric Diseases

Diagnoses of patients with neuropsychiatric diseases (SCZ and MDD) were confirmed according to corresponding Structural Clinical Interview of the DSM-IV criteria. Hamilton depression rating scale (HDRS, 17 items) (Hamilton 1960) and positive and negative syndrome scale (PANSS) (Kay et al. 1987) were used for evaluating the disease severity of patients with MDD and SCZ, respectively. All the patients with AD fulfilled the diagnostic criteria recommended by the National Institute on Aging/Alzheimer’s Association (NIA-AA) workgroups in 2011. The severity of dementia was evaluated using CDR scale (Morris 1993), and all patients with AD had mild dementia with a CDR score of 1. None of the patients with AD had a history of major head injury, brain tumor, stroke, epilepsy, alcoholism, major psychiatric illness, and other systemic diseases that affect the cognitive function.

Image Acquisition

Scanning took place at two sites. Healthy participants and patients with neuropsychiatric disorders were subjected to identical image acquisition using an identical 3T MRI scanner (Siemens Healthcare, Magnetom TIM Trio) with a 12-channel phased-array head coil in NYMU. In contrast, brain images of patients with AD were obtained using another 3T MRI scanner (General Electric Medical Systems, Discovery MR750) with an 8-channel phased-array head coil in TVGH. Participants’ heads were constrained with cushions to minimize head motion during image acquisition. All images were acquired with alignment to the anterior/posterior commissure line, without in-plane interpolation and interslice gap. An experienced neuroradiologist inspected all anatomical scans to exclude participants with apparent image artifacts or brain abnormalities, including trauma, tumors, and hemorrhagic or infarct lesions. Healthy participants and patients with neuropsychiatric disorders underwent identical whole-brain T1-weighted anatomical scans performed using a three-dimensional magnetization-prepared rapid-acquisition gradient-echo (3D-MPRAGE) sequence (repetition time/echo time/inversion time = 3500/3.5/1100 ms; flip angle = 7°; field of view = 256 × 256 mm²; matrix size = 256 × 256; voxel size = 1.0 × 1.0 × 1.0 mm³; number of excitations = 1; and number of sagittal slices = 192). For patients with AD, a three-dimensional fast spoiled-gradient recalled echo (3D-FSPGR) sequence was used for acquiring whole-brain T1-weighted anatomical scans (repetition time/echo time/inversion time = 9.384/4.03/450 ms; flip angle = 12°; field of view = 256 × 256 mm²; matrix size = 256 × 256; voxel size = 1.0 × 1.0 × 1.0 mm³; number of excitations = 1; and number of axial slices = 172). Furthermore, to minimize the effect of WM lesions on brain tissue segmentation and volume estimation, T2-weighted fluid-attenuated inversion-recovery (FLAIR) scans of each individual were also acquired. A two-dimensional T2-weighted FLAIR multishot turbo-spin echo sequence (repetition time/echo time/inversion time = 9000/143/2500 ms; flip angle = 130°; field of view = 220 × 220 mm²; matrix size = 320 × 320; voxel size = 0.69 × 0.69 × 2.0 mm³; number of excitations = 1; and number of axial slices = 63) and a three-dimensional T2-weighted FLAIR sequence (repetition time/echo time/inversion time = 6000/128/1856 ms; flip angle = 90°; field of view = 256 × 256 mm²; matrix size = 256 × 256; voxel size = 1.0 × 1.0 × 1.0 mm³; number of excitations = 1; and number of sagittal slices = 180) were used with a Siemens MRI scanner and a GE MRI scanner, respectively.

Motion Assessment of T1-Weighted Anatomical Scans

To quantify the degree of head motion of the individual T1-weighted anatomical scans, the MRI Quality Control tool was used to calculate the motion-related index (https://github.com/poldracklab/mriqc; Esteban et al. 2017). The entropy focus criterion (EFC) index, estimated based on the Shannon entropy of voxel intensities of the T1-weighted scans, was used as a candidate indictor for describing the degree of head motion. This index has been suggested to indicate image ghosting and blurring induced by head motion (Atkinson et al. 1997).

Brain Age Analytic Framework

Figure 1 shows a general overview of the proposed analytical framework. Briefly, after estimating individual voxel-wise GMV maps, we conducted a multivariate spatial independent component analysis (ICA) with different component orders. We then estimated the corresponding network integrity indices (beta coefficient) using a spatial regression analysis in the training dataset. These network integrity indices were further used as input features for constructing the corresponding brain age estimators at different ICA model orders. After determining the optimal ICA order for the proposed brain age estimator, the corresponding un-thresholded ICA maps were then used as predefined spatial templates of the large-scale SCNs for extracting network integrity indices of the testing and clinical datasets. Subsequently, the brain age estimator that was constructed on the basis of the network integrity index of the corresponding large-scale SCNs of the training dataset was further applied to the testing dataset and the patients with AD, SCZ, and MDD to provide proof-of-concept evidence of hypothesized “accelerated brain aging” in neurodegenerative and neuropsychiatric disorders. Using this proposed analytical framework, the identical spatial ICA maps were applied to training, testing, and clinical datasets and the corresponding network integrity indices were calculated to estimate the individual predicted brain ages. Details of the proposed analytic pipeline are summarized below.

Figure 1

Analytical framework overview. (A) Data preprocessing: The VBM preprocessing pipeline is used to generate MNI-space modulated GMV maps for the training dataset, testing dataset, and clinical dataset. The preprocessed GMV maps of the training dataset with chronological age served as inputs for constructing the brain age estimator. (B) Feature extraction: The large-scale SCNs of the entire training dataset are estimated using spatial ICA approach. The spatial regression analyses with different ICA orders are then applied to estimate network integrity indices of the corresponded SCNs for each individual of the training dataset. (C) Model construction, validation, and evaluation: LASSO regression with nested 10-fold cross-validation scheme is used to construct the proposed brain age estimator from the training dataset with different ICA orders. The MAE and mean coefficient of determination (R²) were used to determine the optimal ICA order. (D) Feature extraction: For the testing and clinical dataset, spatial regression analyses with predifined ICA orders were applied to estimate network integrity indices of the corresponding SCNs. (E) Model generalization and its clinical application: The final large-scale SCN-based brain age estimator (established from training dataset) was used to assess generalizability (to testing dataset) and feasibility of clincal application (to clinical dataset). Abbreviations: AD, Alzheimer’s disease; GMV, gray matter volume; ICA, independent component analysis; LASSO, least absolute shrinkage and selection operator; MDD, Major Depressive Disorder; MNI, Montreal Neurological Institute; SCN, structural covariance network; SCZ, schizophrenia; T1, T1-weighted magnetic resonance imaging; and VBM, voxel-based morphometry.

Open in new tabDownload slide

Voxel-Wise Gray Matter Volume Map Estimation of the Healthy Participants

The voxel-based morphometry (VBM) analytical framework (Ashburner and Friston 2000) with Statistical Parametric Mapping software (SPM8, Version 6313, Wellcome Institute of Neurology, University College London, UK, http://www.fil.ion.ucl.ac.uk/spm/) and MATLAB R2010a (Mathworks, Natick, MA), was used to estimate voxel-wise GMV map for each participant. VBM was performed as described in our previous study (Chen et al. 2015); its brief description is as follows: (1) individual T2-weighted FLAIR scan was affine-registered to the corresponding T1-weighted scan and then served as the inputs for WM lesion identification and lesion-filling procedure, using the Lesion Segmentation Toolbox (LST, version 1.2.3, http://www.applied-statistics.de/lst.html) with default settings (Schmidt et al. 2012); (2) the image origin of lesion-filled T1-weighted scans was automatically reoriented using a center-of-mass approach to minimize imaging position variation across participants; (3) native-space lesion-filled T1-weighted scans were corrected for intensity inhomogeneity, segmented into distinct tissue classes (GM, WM, and cerebrospinal fluid), and further affine-aligned into the standard Montreal Neurological Institute (MNI) space using the VBM8 toolbox with default settings (version r445, http://dbm.neuro.uni-jena.de/); (4) study-specific affine-aligned GM and WM templates were generated from all the healthy participants (training and testing datasets) using the Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra approach (Ashburner 2007); (5) resultant flow fields were applied to warp the corresponding individual affine-aligned GM map and then modulated with corresponding nonlinear components to create the individual voxel-wise GMV map with a 1.5 mm³ voxel size (this nonlinear modulation step makes subsequent large-scale SCN analyses more sensitive to tissue volume than to density and accounts for individual variation in global brain size); (6) resultant MNI-space modulated GMV maps were smoothed by convolution with an 8 mm full-width at half-maximum isotropic Gaussian kernel. To avoid the potential partial volume effect, MNI-space unmodulated GM images were averaged across all participants (training and testing datasets) to generate the explicit mask at a threshold of 0.2 and further applied to the resultant dataset. All preprocessed data after each analytical step of the VBM pipeline were visually inspected to identify data with inaccurate preprocessing results (tissue mis-segmentation or image misalignment). After finishing preprocessing, the “check data quality” module of VBM8 toolbox was used to check sample homogeneity for the quality control of preprocessed data. This module calculates between-subject covariance across all input anatomical volumes of the given group. Following the instructions of the manual, we used the individual modulated and normalized GM tissue segments as inputs and visually inspected the preprocessed scans with an overall covariance below two standard deviations. No subjects were excluded from this quality control procedure. Finally, voxel-wise GMV maps of the training dataset were concatenated as a 4D dataset to identify large-scale SCNs.

Large-Scale Structural Covariance Network Identification in the Training Dataset

We identified large-scale SCNs of the training dataset using multivariate spatial ICA, which enabled us to identify naturally grouped, maximally spatial independent sources of GMV across study participants without the predefined regions of interest (Xu et al. 2009). Unlike a seed-based correlation approach, multivariate ICA could minimize inaccurate seed placement and capture the morphometric covariance pattern of the human brain (Alexander-Bloch, Giedd, et al. 2013a; Yu et al. 2017). The Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC; FSL v5.0.9; http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) tool was used to decompose the concatenated 4D GMV dataset (subject-by-voxel) into a set of spatially distinct components with corresponding weighted scores. More specifically, the ICA decomposed signal intensity across the concatenated 4D GMV data into subjects-by-components mixed and components-by-voxel source matrices. The mixing matrix exhibited the weights of each component for each subject and the source matrix exhibited the covariation of GMV among subjects, which represented the corresponding spatial distribution of brain network (i.e., SCN) (Xu et al. 2009). As there is no general consensus on the number of ICA orders that would be appropriate for estimating individual brain age, different ICA orders (6–150 components at intervals of 10) were extracted to evaluate the predictive performance of the investigational brain age estimator; this range was consistent with that used in previous SCN studies (Segall et al. 2012; Douaud et al. 2014; Hafkemeijer et al. 2014). We visually checked all the spatial locations of the large-scale SCNs and removed nonbiological meaningful components before constructing the brain age estimator. Independent components (ICs) with significant spatial overlap with non-GM areas (including ventricles, large vasculature, WM, and brainstem), or located at the boundaries between different tissue types or sharp edges near brain boundaries were removed (Xu et al. 2009). After visual inspections, retained SCNs were used as input features for constructing the brain age estimator. To visualize the spatial location of the optimum set of SCNs, the spatial maps of each SCN were scaled into a mixture model z-score map and further thresholded at an absolute z-score of >4.

Network Integrity Index Estimation of Corresponding Structural Covariance Network

To construct the large-scale SCN-based brain age estimator and further apply to the unseen dataset, we needed to define a quantitative index to represent the integrity of the corresponding SCNs and then served as input features for the following brain age estimation. Using these requirements, we used the recently proposed unbiased method, spatial regression analysis with large-scale SCNs as an additional step, to calculate the SCN integrity indices of the corresponding SCNs (Hafkemeijer et al. 2016). A general linear model (FSL command-line tool “fsl_glm”) was used for the spatial regression of the 4D GMV dataset against unthresholded SCN maps with different ICA orders. The beta weights of spatial regression analyses (positive or negative value) represent the network integrity level of corresponding SCN for each participant (Hafkemeijer et al. 2016). A larger network integrity index indicated a stronger expression of the corresponding SCN in that participant and vice versa. Previous clinical studies have also demonstrated its ability in identifying large-scale structural network abnormalities in clinical population (Zeighami et al. 2015; Foster-Dingley et al. 2016; Hafkemeijer et al. 2016; Koini et al. 2018). Hence, all subsequent large-scale SCN-based brain age estimator simulations, construction, and related clinical applications were performed based on these network integrity indices.

Statistical Analyses

Demographic and Clinical Characteristics

All demographic data analyses were performed using SPSS Statistical software for Windows, Version 20.0 (IBM Corp.). Continuous (chorological age, years of education, EFC index, and MMSE) and categorical variables (sex) were compared between the training and testing dataset using the two-sample Student t test, analysis of covariance (ANCOVA), and chi-squared tests. Furthermore, differences among clinical and testing dataset were investigated using analysis of variance (ANOVA; age and years of education), ANCOVA (EFC index and MMSE), and chi-squared test (sex). Bonferroni corrected P-value of <0.05 was considered statistically significant.

Construction of SCN-Based Brain Age Estimator and Evaluation in the Training Dataset

To evaluate the effect of different ICA orders on brain age estimation and to determine the optimal ICA order further, the optimal regularization λ value and the predictive performance of the brain age estimator on different ICA orders in the training dataset were determined using the nested 10-fold cross-validation scheme (Ambroise and McLachlan 2002; Varoquaux et al. 2017).

Specifically, training datasets were divided into a training set and test set. In the training set, the optimum penalization parameter λ was determined on the basis the minimum mean-squared error of the inner 10-fold cross-validation loop. The training model was then applied to the test set to evaluate the predictive performance of the brain age estimator in the outer 10-fold cross-validation loop. The mean absolute error (MAE) and mean coefficient of determination (R²) were used as quantitative indices for model evaluation. Finally, we used the inflection point of the mean MAE and R² values to determine the optimal ICA order of the brain age estimator; the inflection point was defined as the maximum difference of the slope by computing the first derivative of the mean MAE and the mean R² value between successive ICA orders. After selecting the optimal ICA order for the training dataset, the whole training dataset was used to construct the final large-scale SCN-based brain age estimator; it was further applied to not only the unseen testing dataset but also the AD, SCZ, and MDD groups. In addition, these un-thresholded ICA maps were used as the optimal predefined spatial templates of the large-scale SCNs for extracting individual network integrity indices of the testing and clinical datasets. Furthermore, additional hierarchical clustering analysis using the Caliński-Harabasz method was performed to visualize putative subnetwork systems of the brain age estimator (Calinski and Harabasz 1974). The optimal set of SCN maps can be downloaded via NeuroVault (https://neurovault.org/collections/5848/). To further examine the relative contribution of each SCN and the corresponding subnetwork systems in estimating individual brain age, each SCN was ranked according to the absolute weight of the constructed brain age estimator; the contribution of subnetwork systems was determined by summing the weights of each SCN that are involved in corresponding subnetwork systems.

Generalization of the SCN-Based Brain Age Estimator for the Testing Dataset

To evaluate the generalizability of the SCN-based brain age estimator constructed using the training dataset to the unseen testing dataset, the same spatial regression analysis, with 40 predefined, un-thresholded ICA spatial maps (obtained from the training dataset), was used to calculate the corresponding network integrity indices for the testing dataset individually; the MAE and R² values were calculated similarly.

Testing “Accelerated Brain Aging” Hypothesis in Clinical Populations

The same VBM processing pipeline with predefined, standard MNI space tissue templates (established from healthy participants) and network integrity estimation procedures were also applied to the AD, SCZ, and MDD groups before applying the constructed large-scale SCN-based brain age estimator. The brain age gap (Franke et al. 2010; Cole et al. 2017), defined as the difference between chronological age and predicted brain age of each individual, was calculated. To minimizing the confounding effects of nuisance variables, ANCOVA was used to compare the brain-age gaps between healthy controls (testing dataset) and patients with AD, SCZ and MDD, with sex, chronological age, years of education, and EFC index were entered as nuisance variables. This nuisance covariate adjustment approach has been widely used in previous large-scale neuroimaging-based brain age studies (Cole et al. 2015; Kuhn et al. 2018; Le et al. 2018; Liang et al. 2019; Sone et al. 2019). Bonferroni corrected P-value of <0.05 was considered statistically significant.

Investigating Potential Relationship Between Brain Age Gap and Clinical Profiles

Partial Pearson’s correlation analyses were used to investigate the potential association between individual brain age gap and clinical profiles (including MMSE, HDRS, and PANSS scores) in each disease group. Participants’ chronological age, sex, years of education, and EFC index were included as nuisance variables. The threshold for statistical significance was set at a two-tailed Bonferroni corrected P-value of <0.05.

Exploring Potential Selective Vulnerability Profiles of Large-Scale SCNs in Neurodegenerative and Neuropsychiatric Diseases

To explore the possible differential contributions of various SCNs to different diseases, an exploratory univariate ANCOVA was used to compare the network integrity index of the corresponding large-scale SCNs between healthy controls and clinical groups with participants’ sex, chronological age, years of education, and EFC index entered as nuisance variables. False discovery rate (FDR) corrected P-value of <0.05 was considered statistically significant.

Figure 2

Predictive performance of the SCN-based brain age estimator as a function of ICA orders. The predictive performance of the brain age estimator increased with ICA order size; MAE ranged from 4.92 to 3.74 and R² from 0.49 to 0.69. Increasing predictive performance inflected at 40 ICs (MAE = 3.97 years, R² = 0.66). Red/blue shading: ±1 standard error. Abbreviations: MAE, mean absolute error; and R², coefficient of determination.

Open in new tabDownload slide

Evaluating Potential Confounding Effects of the SCN-Based Brain Age Estimator

To evaluate the influence of confounding factors on constructing the SCN-based brain age estimator, two additional analyses were performed. First, to evaluate the sample inhomogeneity effect of healthy participants, we exclusively used the ILAS dataset as the primary healthy cohort. Briefly, we randomly allocated the ILAS dataset into training (N = 635) and testing (N = 86) datasets to evaluate the predicted performance of the SCN-based brain age estimator. The same nested 10-fold cross-validation scheme was used to evaluate the relationship between predictive performance of brain age estimator and ICA order size. The following inflection point analysis was also conducted to determine the optimum ICA orders of the proposed brain-age estimator. Finally, the direct group comparison on MAE and R² value (across the 10-fold cross-validation) between two brain-age estimators that were constructed from combined the ILAS and TAMIND projects and exclusively ILAS project was conducted. In the second additional analysis, we evaluated whether using different machine learning regression algorithms may substantially influence prediction accuracy on individual brain age. Additional L2-norm regularized ridge regression (Hoerl and Kennard 1970) and combined L1-norm and L2-norm regularized elastic-net regression (Zou and Hastie 2005) were used.

Results

Demographics and Clinical Characteristics of Study Participants

Table 1 summarizes the demographic data and clinical profiles of the training and testing dataset and those of the patients with AD, SCZ, and MDD. There were no statistically significant differences in sex (χ²₁ = 0.068, P = 0.794), age (t = 0.36, P = 0.719), years of education (t = −0.45, P = 0.652), EFC index (F_1,905 = 0.175, P = 0.676), or MMSE (F_1,900 = 1.113, P = 0.247) between the training and testing dataset (healthy participants). Statistically significant differences in chorological age (F_3,180 = 19.78; P < 0.001), years of education (F_3,180 = 5.66; P = 0.001), EFC index (F_3,178 = 55.16, P < 0.001), and MMSE (F_3,177 = 21.69; P < 0.001) were observed among testing dataset and disease groups. Subsequent posthoc tests revealed that patients were significantly younger in the SCZ group than in the other three groups (Bonferroni corrected P < 0.05) and significantly older in the AD group than in the other three groups (Bonferroni corrected P < 0.05). The SCZ group also revealed a significantly higher years of education than the testing dataset and AD group (Bonferroni corrected P < 0.05). Furthermore, the AD group showed a significantly higher mean EFC index and lower mean MMSE than the other groups (Bonferroni corrected P < 0.05). In addition, no differences in sex were observed among healthy controls and disease groups (χ²₃ = 1.751, P = 0.626).

Predictive Performance of the SCN-Based Brain Age Estimator with Different ICA Orders

The predictive performance of the brain age estimator increased with ICA order size; MAE ranged from 4.92 to 3.74 and R² from 0.49 to 0.69. However, the predictive performance improved little beyond 40 ICs. As demonstrated, the simulation results suggested that the brain age estimator with 40 ICs showed a satisfactory balance between model complexity and prediction accuracy in the training dataset (MAE = 3.97 years, R² = 0.66) (Fig. 2). Therefore, we chose 40 ICs as the optimum network order for constructing the final brain age estimator with the entire training dataset. After LASSO estimation, the ICA order of the final brain age estimator was 37 (with nonzero beta coefficient, Supplementary Table 1). Figure 3 depicts the detailed spatial patterns and locations of these 37 large-scale SCNs, which are further divided into the following distinct subnetwork systems after hierarchical clustering analysis: (1) the cerebellar and subcortical network system; (2) the posterior default-mode network (pDMN), posterior fronto-parietal network (pFPN), and motor network system; (3) visual network, salience network, anterior fronto-parietal network (aFPN), and anterior default-mode network system (aDMN); and (4) temporal network system. By analyzing the absolute weight of the SCNs that contribute to the final brain-age prediction model, we found the subcortical and cerebellar subnetwork systems (top 1 of the four subnetwork systems), which include the putamen (IC 23, absolute weight: 587.79), caudate nucleus (IC 17, absolute weight: 539.46), hippocampus-amygdala complex (IC 29, absolute weight: 320.26), and cerebellum (IC 05, absolute weight: 190.22) play relatively important roles in individual brain age estimation (Supplementary Table 1 and Supplementary Fig. 1).

Figure 3

Four distinct subnetwork system of the established brain-age estimator revealed by hierarchical clustering analysis. The 37 SCNs (with nonzero beta coefficient) are grouped into the four distinct subnetwork systems using hierarchical clustering analysis. The spatial maps of each SCNs are shown in hot colors and thresholded at z > 4. Abbreviations: aDMN, anterior default mode network; aFPN, anterior fronto-parietal network; IC, independent component; pDMN, posterior default mode network; pFPN, posterior fronto-parietal network; and SCN, structural covariance network.

Open in new tabDownload slide

Generalizability of the SCN-Based Brain Age Estimator to the Testing Dataset

The constructed brain age estimator had satisfactory generalizability to the testing dataset, with an MAE = 3.66 years and R² = 0.64 (Table 2 and Fig. 4).

Figure 4

Generalizability of the constructed brain age estimator. The scatter plot showing the detailed data distribution of chronological age and predicted brain age in the testing dataset. The MAE is 3.66 years and the R² is 0.64 in the testing dataset. The shaded area demonstrates the 95% confidence interval. Abbreviations: MAE, mean absolute error; and R², coefficient of determination.

Open in new tabDownload slide

“Accelerated Brain Aging” in Patients with AD, SCZ, and MDD

The estimated brain age gap quantifying individual brain aging relative to the global network pattern differed significantly among testing dataset and disease groups (F_3,176 = 8.64; P < 0.001, Table 2). In posthoc comparisons, the estimated brain age gap was significantly higher in patients with AD (Bonferroni corrected P < 0.05) and SCZ (Bonferroni corrected P < 0.05) than in healthy controls; however, the estimated brain age gap did not differ significantly between healthy controls and patients with MDD (Fig. 5).

Figure 5

Comparison of brain age gap between healthy controls (testing dataset) and clinical groups. Patients with AD and SCZ showed significantly higher estimated brain age gaps than healthy controls. The estimated brain age gap did not significantly differ between healthy controls and patients with MDD. The raw data of the brain age gap in the four groups are plotted on the upper panel. Each mean difference in brain age gap of each group, compared with healthy controls, is plotted the lower panel. Mean differences and 95% confidence interval are indicated with point estimates and vertical error bars, respectively (Ho et al. 2019). Abbreviations: AD, Alzheimer’s disease; HC, healthy controls; MDD, major depressive disorder; SCZ, schizophrenia.

Open in new tabDownload slide

Association Between Brain Age Gap and Clinical Profiles in Clinical Datasets

No statistically significant associations were found between the individual brain age gap and cognitive/disease severity assessments in each disease group. More specifically, there were no statistically significant correlations between MMSE score and brain age gap in each clinical group. In addition, the brain age gap of patients with MDD and SCZ demonstrated statistically insignificant associations with multiple disease severity assessments including HDRS, PANSS Positive scale, PANSS Negative scale, PANSS General Psychopathology scale, and PANSS Total scale.

Differential Contributions of the Large-Scale SCNs in Clinical Datasets

Statistically significant differences in network integrity index were observed in 12 large-scale SCNs among healthy controls and clinical groups (FDR corrected P < 0.05). This included the subcortical network system (ICs 17, 29, and 37); the pDMN, pFPN, and motor network subsystem (ICs 4, 14, 27, 30, 38, and 39); and the visual and salience subnetwork system (ICs 7, 15, and 28) (Table 3 and Fig. 6). Subsequent posthoc tests demonstrated that patients with SCZ had lower network integrity indices in the hippocampus (IC 29), posterior frontal parietal (ICs 14 and 39), visual (IC 7), parietal operculum (IC 15), and salience (IC 28) regions than healthy participants. In contrast, the patients with SCZ also exhibited a higher network integrity index of the caudate nucleus (IC 17). Patients with AD also differed from healthy participants in network integrity indices in several large-scale SCNs. Relative to healthy participants, networks with lower network integrity indices were observed in the hippocampus (IC 29), posterior default-mode network (IC 4), posterior frontal parietal (IC 14, 27, and 39), visual (IC 7), and parietal operculum (IC 15) regions in patients with AD. Alternatively, networks with higher network integrity indices were found in the thalamus (IC 37), posterior frontal parietal (IC 30), and motor (IC 38) regions in patients with AD. Finally, compared with healthy participants, the network integrity indices of seven SCNs, including the hippocampus (IC 29), thalamus (IC 37), posterior frontal parietal (IC 30 and 39), visual (IC 7), parietal operculum (IC 15), and salience network (IC 28) were lower in patients with MDD.

Figure 6

Difference in the large-scale SCNs among healthy controls (testing dataset) and clinical patients. The distributions of the network integrity indices of each SCN in the healthy controls and patients with MDD, SCZ, and AD are visualized using violin plots. The asterisks reveal statistical differences between healthy controls and clinical groups after adjusting for sex, chronological age, years of education, and EFC index effect (FDR-corrected P < 0.05). Mean and standard deviation intervals are indicated with point estimates and vertical error bars. The differences of each SCN are shown in hot colors and thresholded at z > 4. Abbreviations: AD, Alzheimer’s disease; aDMN, anterior default mode network; aFPN, anterior fronto-parietal network; a.u., arbitrary units; EFC, Entropy focus criterion; HC, healthy controls; IC, independent component; MDD, major depressive disorder; pDMN, posterior default mode network; pFPN, posterior fronto-parietal network; SCZ, schizophrenia; and SCN, structural covariance network.

Open in new tabDownload slide

Sampling Inhomogeneity Effect of the SCN-Based Brain Age Estimator

Using the same nested 10-fold cross-validation scheme with the healthy participants in the exclusively ILAS project, the predictive performance of the brain age estimator also increased with ICA order size (MAE: 5.01 to 4.02, R²: 0.46 to 0.65) in the training dataset (Supplementary Fig. 2A). The following inflection point analysis of MAE and R² values also suggested that the brain age estimator with 40 ICs had a satisfactory balance between model complexity and prediction accuracy in the training dataset. In addition, the constructed brain age estimator demonstrated satisfactory generalizability to the testing dataset with an MAE of 3.58 years and R² of 0.49 (Supplementary Fig. 2B). Finally, the direct group comparisons using MAE and R² values demonstrated the predictive performance of the brain age estimator that was constructed using the exclusively ILAS project did not demonstrate more improvements compared with the original ones (MAE value: combined ILAS and TAMIND projects = 3.97 ± 0.41 years, purely ILAS project = 3.92 ± 0.34 years, and P = 0.799; R² value: combined ILAS and TAMIND projects = 0.66 ± 0.08, purely ILAS project = 0.55 ± 0.07, and P = 0.004).

Effect of the Different Machine Learning Algorithms in Predicting Individual Brain Age

Using the same nested 10-fold cross validation scheme, the predictive performance of the three constructed brain age estimators also increased with ICA order (Ridge regression: MAE ranged from 4.94 to 3.76 and R² ranged from 0.49 to 0.69; Elastic-net: MAE ranged from 4.92 to 3.74 and R² ranged from 0.49 to 0.69; and LASSO: MAE ranged from 4.92 to 3.74 and R² from 0.49 to 0.69) in the training dataset (Supplementary Fig. 3). These three algorithms were performed similarly, exhibiting large range of overlap in brain age prediction accuracy regardless of the ICA orders. The following inflection point analysis of MAE and R² values also suggested that the brain age estimator with 40 ICs had a satisfactory balance between model complexity and prediction accuracy for all three machine learning regression algorithms. Finally, the constructed brain age estimator with three regularized regression algorithms were consistently generalized to the testing dataset (Ridge: MAE = 3.74 years, R² = 0.626; Elastic-net, MAE = 3.67 years, R² = 0.636; and LASSO, MAE = 3.66 years, R² = 0.64). These results indicate the prediction performance and generalization ability of the proposed large-scale SCN-based brain age estimator are stable across different machine learning regression algorithms.

Discussion

In this study, we applied large-scale SCNs as a network-based feature to predict individual brain age in a large sample of middle-aged and older adults. We also systematically evaluated the effect of ICA orders on predicting individual brain age and demonstrated the feasibility of using structural MR images for constructing an SCN-based brain age estimator to predict brain age in healthy people, with well-balanced complexity versus accuracy. Our results indicated that the subcortical–cerebellum related SCNs might play an important role in age prediction of participants in middle-to-late adulthood. We further evidenced the hypothesis of “accelerated brain aging” in neurodegenerative and neuropsychiatric diseases by using SCN. In summary, the proposed network-level analytical framework might facilitate the prediction of individual brain age from large-scale structural network perspective.

Optimal ICA Orders for Predicting Individual Brain Age in Middle-to-Late Adulthood

Structural alteration of the human brain in the entire aging trajectory involves not only the regional brain area but also the reorganization of large-scale brain networks (Liu et al. 2017). Previous studies have demonstrated the feasibility of using spatial ICA approach in identifying multiple brain networks on the basis of covariant tissue volume and density information across all study participants (Segall et al. 2012). Despite being data-driven in nature, subsequent studies also provided possible biological meanings to these covariant GMV pattern of the human brain (Schmitt et al. 2008; Zielinski et al. 2010; Alexander-Bloch, Giedd, et al. 2013a). The spatial patterns of structural networks identified in the current study were highly similar to those identified in previous SCN-based studies (Hafkemeijer et al. 2014; Liu et al. 2017). However, previous studies typically used arbitrary number of components (9–70 components) to indicate that aging is associated with a structural network-level decline (Douaud et al. 2014; Hafkemeijer et al. 2014; Liu et al. 2017) and to explore the possible between-group differences in large-scale SCNs (Coppen et al. 2016; Hafkemeijer et al. 2016; de Schipper et al. 2017). The ambiguity of such data-driven approach lies in the determination of the optimal order of SCNs for specific research topics. In this study, we systematically evaluated the inter-relationship between number of component orders and predictive accuracy of brain age estimation in terms of multiple quantitative indices (MAE and R²). Although the predictive accuracy of SCN-based brain age estimators increased with ICA order number, the predictive performance improved modestly beyond 40 ICs. In general settings of neuroimage-based prediction model, the number of independent variables (features) usually far exceeds the number of dependent variables (subjects) and hence increases the occurrence rate of the overfitting problem. Therefore, we suggest using lower ICA orders to avoid model overfitting and to improve prediction accuracy and speed of computing. In contrast, the number of latent features retrieved using data-driven approach largely depends on selective methods, data characteristics, and sample size (Abraham et al. 2017). Under the setting of the current study design, we suggested that the constructed brain age estimator with 40 ICs may be the optimal order number for Chinese participants in middle-to-late adulthood. Further systematic evaluations will be needed when applying the proposed predictive framework for participants with demographic characteristics including age range and races.

Evaluating Generalizability of the Constructed SCN-Based Brain Age Estimator

Generalizability and hyperparameter tuning are two important issues involved in predictive model construction. Different cross-validation schemes were proposed for achieving these objectives. Although traditional leave-one-out cross-validation scheme is popular in the neuroimaging field, one recent study pointed out that this approach will produce unstable and biased predictive power (>10% prediction errors) (Varoquaux et al. 2017). Unlike the leave-one-out cross-validation approach, the nested cross-validation scheme has two cross-validation loops that tune model-parameters in the inner loop and evaluate the predictive performance in the outer loop, which could avoid systematic biases. Under this consideration, we used the 10-fold nested cross-validation scheme in the current study. To further confirm our result and to perform empirical evaluation of the predictive power of the SCN-based brain age estimator, we used an additional unseen testing dataset for evaluating the generalizability of the proposed analytical framework. Commensurate predictive accuracy in the independent testing dataset affirmed the generalizability of the constructed SCN-based brain age estimator, which was comparable with the structure-based prediction models in previous studies that evaluated brain age with high accuracy, on the basis of different brain anatomical features such as GMV (Franke et al. 2010; Cao et al. 2015; Cole et al. 2017), WM integrity (Mwangi et al. 2013; Lin et al. 2016), cortical thickness (Khundrakpam et al. 2015; Aycheh et al. 2018), and multimodal features (Brown et al. 2012; Erus et al. 2015; Liem et al. 2017). However, unlike previous studies, our study could fill the gap by including a larger sample of participants in middle-to-late adulthood and further extending the knowledge on network-level information for brain age prediction in this critical period of human aging process.

Insights on Large-Scale SCNs from the Perspective of Individual Brain Age Estimation

The knowledge of relative contribution of the corresponding SCNs to the brain age estimation could provide potential insights on biological aging of the human brain. In our study, the subcortical and cerebellar SCN, which include the striatum (caudate, putamen), hippocampus-amygdala complex, and cerebellum, contribute substantially to the prediction of individual brain age and these are further grouped as a single subnetwork system. Using various data-driven approaches, previous structural and functional neuroimaging studies have also demonstrated that these distinct anatomical regions could be grouped as a single network system (Huo et al. 2016; Bagarinao et al. 2019). Furthermore, previous studies also indicated that this subcortical–cerebellum subnetwork system has rich bidirectional anatomical and functional connections within itself and that this network could be assembled as an integrated neuronal network involved in different cognitive functions, including motor control, emotion regulation, and reward-based learning (Bostan et al. 2010, 2013; Bostan and Strick 2018). These different domains of cognitive function have also been reported with strong aging-associated changes in healthy elders (Seidler et al. 2010; Eppinger et al. 2011; Ebner and Fischer 2014). In addition to the potential functional changes in the subcortical–cerebellum subnetwork in the aging process, several structural neuroimaging studies also demonstrated the aging-related neuroanatomical changes within this subnetwork system. During the aging trajectory, especially during middle-to-late adulthood, the shrinkage and accelerated regional GMV atrophy rate were noted in the hippocampus, cerebellum, caudate nucleus, entorhinal cortex, and orbitofrontal cortex (Raz et al. 2005, 2010). In addition to its regional significances, age-related alteration of structural covariance profile and intrinsic functional connectivity of the hippocampus, caudate nucleus, thalamus, and cerebellum have also been identified (Tomasi and Volkow 2012; Liu et al. 2017). A recent large-scale SCN study has also demonstrated the altered covariance between the striatum and hippocampus in late adulthood and suggested that this pattern is associated with a compensatory mechanism of individual memory performance and dopaminergic modulation during aging (Li et al. 2018). Together with our findings, we suggest that the subcortical–cerebellum subnetwork could act as a potential aging-associated network signature in middle-to-late adulthood and that it could serve as a predictive biomarker in future normal and pathological aging studies. In contrast, our analytical framework also demonstrated that multiple cerebral SCNs contribute to individual brain age estimation. Although the weighting value of these cerebral SCNs is relatively minor to the subcortical–cerebellum subnetwork system, the alteration in these cerebral SCNs, which include default-mode network, executive control network, sensorimotor network, attention network, and temporal network, were also evidenced in the aging process (Li et al. 2013; Hafkemeijer et al. 2014; Liu et al. 2017). The difference in relative contribution between the cerebral and subcortical–cerebellar SCNs may be partially in line with a previous study, which suggested that the changing rate of GMV may be more accelerated in the subcortical regions than in other cortical areas in middle-to-late adulthood (Fjell et al. 2009). These essential differences may further promote a higher performance of the subcortical–cerebellum related SCNs than that of other cerebral SCNs in predicting individual brain age. However, biological process based only on model coefficients, which are estimated from various machine learning approaches, need to be interpreted cautiously (Haufe et al. 2014). Hence, in the future, related experiential designs will be needed to decipher the importance and potential biological meanings of the cortical and noncortical SCNs in human aging trajectory, respectively.

Accelerated Brain Aging in Clinical Populations from the Perspective of Large-Scale SCNs

The hypothesis of accelerated brain aging in AD and SCZ has been tested from various clinicopathological perspectives, including brain morphometry, gene expression, metabolic problems, and cognitive functions (Kirkpatrick et al. 2008; Spulber et al. 2010; Saetre et al. 2011). Studies that estimated individual brain age using structural MRI clearly demonstrated that the pattern of accelerated brain aging may exist in patients with AD and SCZ but not in those with MDD (Franke et al. 2010; Franke and Gaser 2012; Gaser et al. 2013; Kochunov et al. 2013; Lowe et al. 2016; Schnack et al. 2016; Nenadic et al. 2017; Shahab et al. 2019). Furthermore, a recent study used structural MRI data from 45 615 individuals (aged 3–96 years) demonstrated the pattern of apparent brain aging in several neural disorders including dementia and schizophrenia, but not MDD (Kaufmann et al. 2019). Although there are substantial methodological differences between these studies (single site versus multisite study design; regional- versus network-level brain morphometrical features as predictors; and age range of the clinical populations), they consistently demonstrated apparent brain aging in clinical populations. Collectively, using regional and network-level brain morphometrical measurements as predictors, these previous studies could provide complementary understanding to characterize individual brain health in terms of brain age.

From the perspective of large-scale SCNs, we observed accelerated brain aging in the AD and SCZ groups, but not in MDD groups. Previous large-scale multisite neuroimaging studies have demonstrated a more widespread and severe cortical and subcortical gray matter abnormalities in patients with SCZ, especially in the hippocampus, amygdala, thalamus, and nucleus accumbens, than in patients with MDD (Schmaal et al. 2016, 2017; van Erp et al. 2016, 2018). In contrast, patients with AD also demonstrated widespread gray matter atrophy in regions including the frontal lobe, temporal lobe, hippocampus, parahippocampal gyrus, and cerebellum (Wang et al. 2015; Jacobs et al. 2018). The differences in the degree of GMV and structural network deficits in the cortical, subcortical, and cerebellar areas among the three disease groups may be related to our observation regarding brain age prediction. In addition, although chronologically our patient groups were older than those in previous studies, our result further suggested that the pattern of accelerated brain aging may not only be present in early adulthood but also be extended into the middle-to-late adulthood in neurodegenerative and neuropsychiatric disorders. In summary, our result further implies that the brain aging trajectories may be different among different neuropsychiatric and neurodegenerative diseases. However, we have a limited sample size for providing the initial proof-of-concept evidence of clinical application using SCN-based brain age estimation approach; future studies with larger sample sizes may be needed to explore the potential brain abnormalities in terms of brain age across the entire lifespan in different disease groups.

Selective Vulnerability of Large-Scale SCNs in Neuropsychiatric and Neurodegenerative Diseases

Although larger estimated brain age gaps were observed in both SCZ and AD groups, using additional individual SCN integrity analyses, we further demonstrated the potential differential contributions of various SCNs in different diseases. Consistent with previous large-scale brain network studies, the altered SCN integrity of the subcortical regions, including the hippocampus and caudate nucleus, were found and may be associated with memory-related dysfunction and worsened speech recognition performance in patients with SCZ (Avery et al. 2018; Zheng et al. 2018). Beyond connectivity disturbance of the subcortical regions, multiple large-scale cortical SCNs, which included the fronto-parietal control, salience, and primary visual networks were also changed in patients with SCZ. Previous studies have demonstrated functional and structural connectivity disruptions in the associative fronto-parietal control, salience, and primary visual networks in patients with chronic SCZ (Spreng and Turner 2013; van de Ven et al. 2017) and demonstrated the potential association with various clinical symptoms including delusions, visual hallucination, and perceptual symptoms (Oertel et al. 2007; Palaniyappan and Liddle 2012; Tu et al. 2013). These findings suggest selective alterations in large-scale brain networks may be associated with clinical representations and could be a potential predictive network in patients with SCZ. We also observed the widespread cortical and subcortical SCN alterations in patients with AD. For cortical SCNs, the altered network integrity of the associative fronto-parietal control, visual, and motor networks, which is consistent with previous brain connectomics studies (Dipasquale et al. 2015; Hafkemeijer et al. 2016; Li et al. 2016). These multiple cortical networks have been suggested to be associated with attention deficit and visual motion perception impairments and grip strength decline in patients with AD (Rizzo and Nawrot 1998; Perry and Hodges 1999; Buchman et al. 2007). Furthermore, we identified the altered posterior DMN and hippocampus network in patients with AD. The DMN and hippocampus network was the most consistently reported network involved in AD (Zhang et al. 2010; Badhwar et al. 2017). Previous studies also indicated these two large-scale brain networks may be associated with multiple domains of cognitive functions, including autobiographical/episodic memory retrieval, attention, and emotion regulation (Andrews-Hanna et al. 2010; Leech and Sharp 2014). These multiple domains of cognitive function were also vulnerable in AD progression. In addition to these cortical SCNs, the altered thalamus network, which was involved in the sensory relay and information process, was identified in the present study (Zhou et al. 2013). In conjunction, these different cortical–subcortical alternation patterns of the large-scale SCN in patients with SCZ and AD could further help to uncover the possible differential contributions of SCNs in accelerated brain aging phenomena. Alternatively, although the estimated brain age gap did not differ between healthy participants and patients with MDD, several SCNs, including the hippocampus, salience, and fronto-parietal control networks also demonstrated lower network integrity in patients with MDD. These large-scale brain networks have also been found to be important in the pathophysiology of MDD (Kaiser et al. 2015; Schultz et al. 2019). However, compared with the SCN alternation pattern of patients with SCZ, the between-group differences of the network integrity in the caudate network, which was considered a major contributor in individual brain age estimation was less evident in the MDD group. Several previous brain volumetric studies also demonstrated that greater age-related changes in caudate nucleus volume were more evident in patients with SCZ than MDD (Cropley et al. 2017; Sacchet et al. 2017). This unique selective vulnerability profile of SCN integrity in patients with MDD may help us explain why the accelerating brain aging hypothesis was not relevant in the present study. Further studies integrating individual SCNs analysis and SCN-based brain age estimation could provide a more complete view of brain biological aging in various clinical populations from a large-scale structural network perspective.

Strengths and Limitations

To the best of our knowledge, this is the first study to investigate the properties of large-scale SCNs in predicting individual brain age of a large population in middle-to-late adulthood. However, these results should be interpreted with caution. First, the predicted brain age generated by the SCN-based estimator with an L1-regularized linear regression algorithm could only examine the linear relationships between the integrity of large-scale SCNs and individual chronological age. However, age-related anatomical changes may also demonstrate complex nonlinear patterns. In the future, approaches such as nonlinear kernel method and deep neural network approach might be used to evaluate whether significantly improved accuracy in predicting individual brain age can be achieved. Second, because we only used anatomical MRI scans from a single MRI scanner with identical imaging acquisition protocol, our sample size is relatively modest and data distribution is homogeneous, which might limit the generalizability of our results to larger dataset acquired using multiscanner study design. Third, the present study could not identify any potential associations between the individual brain age gap and clinical profiles in each disease group. The relatively small sample size and lack of a wide range of behavioral and cognitive assessments in the clinical groups may limit further potential clinical implications of brain age estimation. Hence, further study, including a relatively larger sample size and multiple-domain cognitive assessments, are necessary. In conjunction, our exploratory findings and constructed predictive model can guide future large-scale multisite studies, which may further validate and extend the current results.

Conclusions

Our conceptual framework includes automated processing of VBM data, network identification via ICA, and an exploratory statistical linear regression model to estimate brain age in large middle-to-late adulthood populations. This SCN-based protocol performed well in both training and testing dataset and demonstrated the clinical feasibility of detecting “accelerated brain aging” in neurodegenerative and neuropsychiatric disorders. Our brain age prediction for each individual suggests that aging information encoded in inter-relationships between large-scale SCNs might help elucidate the structural organization of the aging brain and have clinical applications for diseases associated with accelerated brain aging.

Funding

Aging and Health Research Center at National Yang Ming University, Taiwan (MOST 108-2634-F-010-001 to L.K.C.); Center for Geriatrics and Gerontology of Taipei Veterans General Hospital of Taiwan (MOST 108-2321-B-010-013-MY2 to P.N.W.); Ministry of Science and Technology, Taiwan (MOST 106-2221-E-010-011, MOST 107-2221-E-010-010-MY3 to K.H.C.; MOST 108-2420-H-010-001, MOST 108-2321-B-010-010-MY2 to C.P.L.); National Health Research Institutes (NHRI-EX108-10611EI to C.P.L.); The Brain Research Center, National Yang-Ming University from The Featured Areas Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education (MOE), Taipei, Taiwan.

Notes

Kun-Hsien Chou had full access to all of the data in the study and takes responsibility for the integrity of the data. Chen-Yuan Kuo, Pei-Lin Lee, Kun-Hsien Chou, Li-Kuo Liu, Chih-Ping Chung, Sheng-Che Hung, Ching-Po Lin, Wei-Ju Lee, Albert C. Yang, Shih-Jen Tsai, Pei-Ning Wang and Liang-Kung Chen contributed to the study concept and design, acquisition, and interpretation of data. Kun-Hsien Chou and Ching-Po Lin contributed to MRI technical and material support. Chen-Yuan Kuo and Pei-Lin Lee contributed to the analysis of the data and the creation of the figures. Chen-Yuan Kuo and Kun-Hsien Chou participated in drafting the manuscript. Li-Kuo Liu, Wei-Ju Lee, Albert C. Yang, Shih-Jen Tsai, and Pei-Ning Wang contributed to the collection and execution of the study. All authors have read and approve of the final version of the manuscript. Conflict of Interest: None declared.

References