Advanced brain age in community-dwelling population with combined physical and cognitive impairments

We investigated whether advanced brain biological age is associated with accelerated age-related physical and/or cognitive functional decline: mobility impairment no disability (MIND), cognitive impairment no dementia (CIND), and physio-cognitive decline syndrome (PCDS). We constructed a brain age prediction model using gray matter features from the magnetic resonance imaging of 1482 healthy individuals (aged 18-92 years). Predicted and chronological age differences were obtained (brain age gap [BAG]) and analyzed in another 1193 community-dwelling population aged ≥50 years. Among the 1193 participants, there were 501, 346, 148, and 198 in the robust, CIND, MIND, and PCDS groups, respectively. Participants with PCDS had significantly larger BAG (BAG = 2.99 ± 8.97) than the robust (BAG = -0.49 ± 9.27, p = 0.002; η2 = 0.014), CIND (BAG = 0.47 ± 9.16, p = 0.02; η2 = 0.01), and MIND (BAG = 0.36 ± 9.69, p = 0.036; η2 = 0.013) groups. Advanced brain aging is involved in the pathophysiology of the co-occurrence of physical and cognitive decline in the older people. The PCDS may be a clinical phenotype reflective of accelerated biological age in community-dwelling older individuals.


Introduction
Longevity is a phenomenon in many countries; according to World Health Organization's report, by 2030, 1 in 6 people will be aged ≥60 years worldwide (World Health Organization, 2021). Burden of morbidity due to dementia and physical disability would be increasing with aging population (World Health Organization, 2021). In response to such an unprecedented demographic shift, several syndromes have been proposed to identify older adults at risks of dementia or disability in the future Panza et al., 2018). In addition to being targets for prevention implementation, these syndromes may serve as clinical models to evaluate the pathophysiology of accelerated age-related functional decline at a preclinical stage. Among these syndromes, physio-cognitive decline syndrome (PCDS) is defined as concurrent mobility impairment no disability (MIND: slow gait and/or weak handgrip but without disable) and cognitive impairment no dementia (CIND: ≥1.5 standard deviation [SD] below the mean for age-, sex-, and education-matched norms in any cognitive domain but without dementia) . Several cohort studies have revealed the prevalence of PCDS in the older community, including 13.3% in Taiwan, 11.2% in Japan, and 18.8% in Singapore, which has multiethnic population (Liu et al., 2018;Merchant et al., 2021;Shimada et al., 2013). These studies also showed that PCDS predicted incident dementia (hazard ratio [HR], 95% confidence interval [CI] = 3.4, 2.4-5.0) and disability (HR, 95% CI = 3.9, 3.0-5.1) in the older people Tsutsumimoto et al., 2020). PCDS successfully comprises a considerable population of at-risk older people and potentially serves as a treatment target at an early preclinical stage of unhealthy aging (accelerated functional decline).
Brain age can be predicted in individuals based on neuroimaging data using machine learning approaches to model trajectories of brain aging (Cole and Franke, 2017). Model-predicted brain age in individuals can differ from one's chronological age, reflecting advanced or delayed brain aging depending on the difference between predicted brain age and chronological age, for example, the brain age gap (BAG) (Cole and Franke, 2017). The BAG has been used to examine the brain aging process in several neurological and psychiatric diseases (Cole and Franke, 2017;Kaufmann et al., 2019;.
In this study, we examined the BAGs in diverse categories of accelerated functional decline, such as robust, CIND, MIND, and PCDS groups, from a community-based functionally preserved population aged ≥50 years who had no dementia. It is hypothesized that advanced brain biological age is associated with PCDS, the preclinical phenotype of concurrent accelerated age-related cognitive and physical functional decline.

Participants-the community-dwelling I-Lan Longitudinal Aging Study cohort
Participants were selected from the I-Lan Longitudinal Aging Study (ILAS) cohort, which is a community-based aging cohort study that aims to investigate the interrelationships among aging, frailty, cognitive functions, and brain structural and functional changes . Community-dwelling adults, aged ≥50 years, inhabitants of the Yuanshan Township in I-Lan County, were invited to participate. Participants were not eligible if they met any of the following conditions: (1) unable to communicate and complete an interview; (2) unable to complete a simple physical task (e.g., a 6minute walk test) or cognitive assessments due to poor functional status; (3) presence of any major illness with associated decreased life expectancy (< 6 months); (4) contraindications to magnetic resonance imaging (MRI) scanning, including claustrophobia, ferromagnetic foreign bodies, or metal implants; and (5) history of neurobiological or neuropsychiatric diseases, such as stroke, dementia, brain tumor, major depression, or alcohol/substance abuse.
All included ILAS participants received a face-to-face neuropsychological assessment administered by trained interviewers, including the Mini-Mental State Examination (MMSE) and neuropsychological tests of different cognitive domains. Dementia was defined as MMSE score < 24 in well-educated participants (education years ≥6) or an MMSE score < 14 in less educated participants (education years < 6) (Liu et al., 1995;Sun et al., 2014). This criterion was established to accommodate the situation of prevalent low education profiles (< 6 years) in the Taiwanese older people community (Liu et al., 1995;Sun et al., 2014). We conducted a field survey on 2753 men and 2544 women from 4 urban and 4 rural participating communities in 1995. Their ages ranged from 41 to 88 years; 28% of them were aged at least 65 years. Their education ranged from 0 to 20 years, and 27% of them had less than a year of formal schooling. This criterion was validated based on the diagnosis of dementia according to the DSM-III-R (Diagnostic and Statistical Manual of Mental Disorders) criteria (Liu et al., 1995;Sun et al., 2014).
This study used demographic information, cognitive assessments, and brain anatomical MRI scans from the initial sampling wave of the ILAS (recruited between January 2011 and July 2014). The ILAS cohort included 1839 participants during the initial wave. Of those, 1281 participants with available T1-weighted anatomical scans were eligible for the present study. Participants with insufficient image quality (n = 41, image artifacts and head motion) and a diagnosis of dementia (n = 47) were excluded. The final population comprised of 1193 participants (age range = 50-91 years; 566 men and 627 women) was used for the following analysis (Fig. 1).

Participants-independent training data set for constructing a brain age model
To construct a brain age prediction model, we aggregated T1weighted anatomical scans of healthy participants from 5 sites (6 MRI scanners) across Taiwan . A total of 1482 healthy participants (age range = 18-92 years, 681 men and 801 women) were used as the training data set for constructing the brain age prediction model. None of the participants had a history of head trauma, brain lesions, other major medical conditions (malignancies, myocardial infarction, stroke, respiratory failure, and end-stage kidney disease), or neurological/neuropsychiatric diseases. The study was approved by the ethics committees of the respective study sites, and participants or their guardians provided signed informed consent prior to participation. The age distribution of each data set is shown in Fig. 2A. The site-specific MRI scanner manufacturer and  The data distribution in training and testing data set. (B) The brain image data were used to extract the GMV features by using the VBM preprocessing pipeline. (C) The GMV features were used to construct the brain age model and validate the performance in the training data set. (D) The constructed brain age model was applied to the ILAS data set and estimated the individual brain age gap (BAG). (E) The statistical analysis was further compared to the difference in BAG among the 4 groups. Abbreviations: BAG, brain age gap; GMV, gray matter volume; ILAS, I-Lan Longitudinal Aging Study; KSCGH, Kaohsiung Chang Gung Memorial Hospital; KLCGH, Keelung Chang Gung Memorial Hospital; NYCU, National Yang Ming Chiao Tung University; ROI, region of interest; TVGH, Taipei Veteran General Hospital; TSGH, Tri-Service General Hospital; VBM, voxel-based morphometry. T1-weighted imaging parameters have been summarized in our previous study .
To ensure that data of all sites could be used together and to lessen the bias from different scanners, we had adopted several measures: (1) The image acquisition protocols were optimized to match their respective MRI scanners.
(2) Prior to any image preprocessing, an experienced radiologist meticulously reviewed all scans to exclude participants with organic brain disorders (e.g., trauma, tumors, and hemorrhagic or infarct lesions) or scans exhibiting poor image quality (e.g., excessive motion or acquisition artifacts).
(3) A standardized image processing pipeline, including voxel-based morphometry analytical procedure and region of interest (ROI)based feature extraction, was implemented consistently across all data sets. (4) A comprehensive image quality control procedure was further conducted by an experienced neuroradiologist.
These rigorous efforts were undertaken to ensure data consistency and usability across different centers, thereby forming a representative training data set for the construction of the final brain age prediction model.

Defining subgroups in the ILAS cohort
All participants in the ILAS cohort were further divided into 4 distinct subgroups (i.e., robust, CIND, MIND, and PCDS) according to the consensus criteria proposed by the 2019 Asian Working Group for Sarcopenia (Chen et al., 2020a;Chung et al., 2021).
Participants whose cognitive performance was lower than 1.5 SD of the study population in any cognitive domain were defined as having CIND. The battery of neuropsychological tests that evaluated multiple domains of cognitive function was conducted via face-toface interviews, including verbal memory (10-minute delayed recall in the Chinese Version Verbal Learning Test) (Chang et al., 2010), language function (the Boston Naming Test) , visuospatial function (the Taylor Complex Figure test) (Taylor, 1969), and executive function (the Backward Digit Test and the Clock Drawing Test) (Agrell and Dehlin, 2012;Hester et al., 2004). Participants with MIND were defined as having handgrip weakness and/ or walking slowness without physical disability [2,13]. A handgrip strength dynamometer (Smedley's Dynamo Meter; TTM, Tokyo, Japan) was used to measure individual's muscle strength, which served as a proxy measure of the degree of weakness, and a timed 6m walk test was used to evaluate participants' walking speed. Finally, participants with PCDS were defined as meeting the criteria for both MIND and CIND (Chen et al., 2020a;Chung et al., 2021).

MRI data acquisition
T1-weighted anatomical scans of the ILAS cohort were acquired at the National Yang Ming Chiao Tung University, Taipei, Taiwan, on a Siemens 3T MRI scanner (Siemens Healthcare, Magnetom Tim Trio, Erlangen, Germany), using an identical acquisition protocol. All anatomical scans were acquired with a 12-channel phased-array head coil using a three-dimensional magnetization-prepared rapidacquisition gradient-echo sequence with the following parameters: repetition time/echo time/inversion time = 3500/3.5/1100 ms, flip angle = 7°, field of view = 256 × 256 mm 2 , matrix size = 256 × 256, number of excitations = 1, isotropic resolution (without interslice gap) = 1 mm 3 , and sagittal slices = 192. Before image preprocessing, all anatomical scans were visually inspected by an experienced radiologist to exclude participants with organic brain disorders (e.g., trauma, tumors, and hemorrhagic or infarct lesions) or poor image quality (e.g., excessive motion or acquisition artifacts).

Brain age prediction model establishment and analysis
An overview of the analytic flow of the present study, including image processing, feature extraction, brain age model construction, prediction performance evaluation, and application, is shown in Fig. 2.

T1-weighted image processing and feature extraction
We extracted regional gray matter volume (GMV) from T1weighed scans with a voxel-based morphometry analytical pipeline using Statistical Parametric Mapping 12 (version 7487; Wellcome Institute of Neurology, University College London, UK) running in a MATLAB environment (version R2015b; Mathworks, Natick, MA). T1weighted scans were processed as described previously and consisted of 5 steps (Baecker et al., 2021;Chou et al., 2021): (1) bias field correction, (2) tissue segmentation, (3) study-specific tissue template generation and spatial normalization to the standard Montreal Neurological Institute space using Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra approach (Ashburner, 2007), (4) tissue modulation to preserve actual tissue volume before and after spatial normalization, and (5) spatial smoothing with a 6 mm full width at half maximum Gaussian kernel. The quality of each image processing step was visually inspected by an experienced neuroradiologist, and no participants were excluded. To further reduce the high dimensionality of voxel-wise data, we utilized a fine-grained cortical parcellation scheme to extract the mean GMV for 400 ROIs in addition to the set of subcortical and cerebellar ROIs from the Harvard-Oxford subcortical structural atlas (14 ROIs) and spatially unbiased infratentorial atlas (28 ROIs) (Schaefer et al., 2018;Desikan et al., 2006;Diedrichsen, 2006;. This procedure yielded 442 regional mean GMV metrics as predictors in a multivariable regression model to characterize chronological age in the training data set.

Brain age model and computation of the BAG
Using the Scikit-learn library (version 0.23.2; https://scikit-learn. org/stable/), support vector regression (SVR) with the radial basis function (RBF) kernel was used to construct the brain age prediction model (Baecker et al., 2021;Pedregosa et al., 2011). Previous studies have shown that including uncertain confounding variables in regression analyses may potentially introduce significant bias in the predictive performance of established models and therefore is less likely to improve models' prediction accuracy (Snoek et al., 2019;Rao et al., 2017;Chyzhyk et al., 2022). Thus, we did not adjust for potential covariates either before or during the construction of the brain age model. A nested 5-fold cross-validation scheme was applied to the training data set, which was randomly split into 5 subsamples with equal sample size . Specifically, the optimal hyperparameter C and gamma of the SVR algorithm were determined using the Scikit-learn GridSearch CV function over the search space 0.001, 0.01, 0.1, 1, 10, 100, and 1000 in the inner cross-validation loop. All other parameters of the RBF-SVR algorithm were set to their default values. In the outer cross-validation loop, model performance was evaluated by comparing the predicted brain age with chronological age via the mean absolute error (MAE), Pearson correlation coefficient (r), and adjusted coefficient of determination (R 2 ) at the individual level. All performance metrics are reported as the mean across iterations of cross-validation. Subsequently, we recomputed the brain age prediction model using the entire training data set, and the corresponding hyperparameters were determined using 5-fold cross-validation scheme. This well-trained brain age model was then applied to estimate the individual brain ages of the participants in the entire ILAS cohort. Finally, to identify degrees of deviation from normative aging trajectories, the individual BAG was calculated by subtracting the chronological age of each individual from their predicted brain age (Franke et al., 2010). These BAG metrics were then used for further statistical analyses.

Statistical analyses
All statistical analyses were conducted using Statistical Package for the Social Sciences version 25 (SPSS; IBM Corp., Armonk, NY, USA).

Demographic and clinical characteristics
Continuous variables are expressed as mean and SD, and categorical variables are expressed as proportions. We used the analysis of variance and analysis of covariance (ANCOVA) tests for continuous variables and the χ 2 test for categorical variables to compare statistical differences among the 4 study groups. The detailed settings of the nuisance variables for each statistical model are presented in Table 1. Two-tailed p value < 0.05 was considered statistically significant in all analyses.

Comparisons of BAG among robust, CIND, MIND, and PCDS
As there was no evidence of nonnormal distribution in the data (p > 0.05), we used a 1-way ANCOVA test to examine differences in biological aging as reflected by the BAG among the 4 study groups. Individual's chronological age, square of chronological age, sex, education level (years), entropy focus criterion index, and total intracranial volume were included as nuisance variables in the ANCOVA model. Using individual's chronological age as a nuisance variable, a systemic age-related bias could be adjusted in predicting brain age (overestimated predictions in younger participants and underestimated predictions in older participants) . The entropy focus criterion index is a quantitative index for characterizing image blurring and ghosting results from participants' head movement and was calculated using the MRI Quality Control tool (https://github.com/poldracklab/mriqc) in the native T1 space. In addition, we further explored the potential effects of chronological age on analytic brain biological aging results by stratifying the analytic cohort into 50-64 years (middle aged) and ≥65 years old (old aged) groups. Finally, effect sizes were calculated using the partial eta squared (η 2 ) approach.

Standard protocol approvals and patient consents
The study was approved by the Institutional Review Board of National Yang Ming Chiao Tung University, Taipei, Taiwan. All participants provided signed informed consent before participation.

Demographic characteristics
Among the 1193 ILAS participants, there were 501, 346, 148, and 198 in the robust, CIND, MIND, and PCDS groups, respectively. Table 1 presents the detailed demographics of the 4 groups. Compared to the robust, CIND, and MIND groups, the PCDS group was older and less educated. Additionally, the PCDS group also had lower scores in all neuropsychological tests, including the MMSE, 10minute Chinese Version Verbal Learning Test, Clock Drawing Test, Taylor Complex Figure test, Boston Naming Test, verbal fluency test, and Backward Digit Test, than the robust, CIND, and MIND groups.
To evaluate the potential age effect, we stratified the population into 2 age groups: < 65 and ≥65 years. Table 2 presents the comparisons between the robust, CIND, MIND, and PCDS groups in each age-stratified population. In the middle-aged group (50-64 years), the PCDS group had fewer years of education than the robust and CIND groups. Regarding cognitive performance in this age group, the PCDS group was similar to the CIND group but worse than the robust and MIND groups. In the old-aged group (≥65 years), the PCDS group had the lowest number of years of education but only showed statistically significant differences in comparison with the robust group. Similar to the middle-aged group, the PCDS group had worse cognitive performance in all domains than the robust and MIND groups in the old-aged group.

Prediction performance of the established brain age model
The brain age prediction model constructed with the training data set using the proposed analytical framework and SVR-RBF multivariate regression algorithm showed good performance (MAE = 6.456 years;  Fig. 3A). Fig. 3B showed the correlation between chronological age and brain age in the ILAS data set, while applying the brain age model to unseen testing data set (MAE = 7.432 years; r = 0.744, p < 0.001; R 2 = −0.841).

Discussion
The present study discovered an association between neuroimaging estimated brain age and accelerated age-related functional decline in a community-based population aged 50-91 years. Participants with PCDS (concurrent physical and cognitive functional impairments) showed advanced brain age compared with robust participants. Participants with isolated physical (MIND) or cognitive (CIND) functional impairments did not show significantly different BAG compared to robust participants.
Both physical and cognitive functions decline with aging since the fifth decade of life (Bohannon and Williams, 2011;Dodds et al., 2014;Hedden and Gabrieli, 2004). Several studies have shown interactions between physical and cognitive functional decline in the aging process, particularly a synergistic detrimental effect on health status of the older people (Kojima et al., 2016;Panza et al., 2018;Tian et al., 2023;Zheng et al., 2020). A recent meta-analysis study has shown that the risk for dementia was higher among those with co-occurrent physical frailty and cognitive impairment compared to those with cognitive impairment alone (Zheng et al., 2020). Previous studies have also showed that PCDS predicted incident dementia (HR, 95% CI = 3.4, 2.4-5.0) and disability (HR, 95% CI = 3.9, 3.0-5.1) in the older people Tsutsumimoto et al., 2020). BAG has been proposed as a biomarker of age-related deterioration of the brain; having an older brain age has been linked to poorer cognitive performance, dementia, and mortality in the older people (Cole and Franke, 2017). A recent study also revealed that those with older brain age at midlife had an accelerated pace of biological aging, older facial appearance, and cognitive decline (Elliott et al., 2021). Advanced brain age in PCDS, first revealed by the present study, might be involved in the patho-etiology of higher risks of dementia and disability in the older people with concurrent physical and cognitive functional impairments. Meanwhile, our study results showing advanced brain age in participants with PCDS but not CIND or MIND support the synergistic detrimental effects of physical and cognitive decline.
In another study evaluating the pathophysiology of PCDS, agerelated covert cerebral small vessel disease (SVD) was found to be associated with PCDS independent of age, sex, and vascular risk factors . Community-dwelling older people with PCDS had more severe white matter hyperintensities (> 50th percentile of white matter hyperintensity volume ratio) and a higher prevalence of multiple lacunes and cerebral microbleeds. The analyzed results showed that 21.1% of PCDS might be attributable to covert SVD . In another previous neuroimaging study, SVD brain lesions were associated with disrupted frontalsubcortical circuit and diffuse cortical damages in all cerebral lobes and cerebellum (Chou et al., 2021). Thus, SVD might mediate, at least partially, the association between accelerated brain age, measured using neuroimaging-yielded GMV-based model, and PCDS.  Many epidemiological evidences support PCDS as the phenotype for early identification of older adults at risk for dementia and disability; therefore, PCDS might serve as a treatment target to prevent or even reverse the accelerated functional decline . Several well-designed randomized controlled trials have demonstrated the efficacy of multidomain interventions in preventing physical and cognitive decline in older adults (Chen et al., 2020b;Ngandu et al., 2015). Our latest trials, Taiwan Integrated Geriatric Care study and Taiwan Health Promotion Intervention Study for Community Elders, also showed significant improvements in physical and cognitive functions and quality of life in older people with (Taiwan Integrated Geriatric Care study) or without (Taiwan Health promotion Intervention Study for Community Elders) multimorbidity after multidomain interventions (Chen et al., 2020b;. The secondary analyses of these trials stratified the study population by status of physio-cognitive decline, and the results showed significant improvement in cognitive functions of several domains and mobility frailty status in community-dwelling older people with PCDS (Chen et al., 2020b;. These results make us wonder whether the underlying mechanism of multidomain interventions is to reverse the accelerated brain structural aging or gain resilience ability resisting functional decline caused by accelerated brain age in PCDS. In addition, multidomain interventions should be considered in future clinical trials of other age-related diseases with advanced brain age. Age has an effect on the association between BAG and PCDS. The BAG of PCDS was greater in the older group (≥65 years: 3.99 ± 7.54 years) than in the younger group (50-64 years: 1.92 ± 10.24 years). This result is consistent with our previous findings that the brain regions with reduced GMV in PCDS were greater and larger in the older group than in the younger group (Liu et al., 2020). Therefore, participants identified as having PCDS at an older age are presumed to have more severe age-related brain damage (Liu et al., 2020). Notably, in participants aged ≥65 years, MIND had the largest BAG (4.03 ± 8.14 years) among the 4 groups, although it showed no statistical significance compared with the other groups. Since the population size of the older MIND group was relatively small (n = 47), further investigation with a larger study population is needed to determine whether age also has an effect between BAG and MIND; for example, MIND only at an older age is associated with advanced brain age.
We would like to point out our efforts to mitigate the potential bias from our machine learning-yielded model for brain age prediction. In the testing data set (ILAS data set), this potential bias was presumed to equally affect the 4 groups (robust, CIND, MIND, and PCDS). We used BAG but not the direct measured brain age to compare among groups. And most importantly, robust group was used as the reference group to evaluate whether groups with accelerated functional declines had advanced brain age. The following methodological considerations should also be considered when interpreting the results. First, the present model of brain age prediction only used GM features (especially GMV) from T1w MRI. Different MRI modalities capture not only shared but also unique information about brain aging (Groves et al., 2012); prediction accuracy may benefit by incorporating these additional sources of information. For instance, Brown et al. (2012) and Erus et al. (2015) have shown that combining information from gray and white matter anatomy increases prediction accuracy in young subjects. Liem et al. (2017) have further investigated how multimodal brain imaging data would improve age prediction and found that multimodal data did improve brain-based age prediction accuracy (decreased MAE) and decrease the SDs compared with any single structural or functional modality. Future research with other brain features, such as white matter volume, microstructural integrity, and functional capacity, is needed to improve our understanding of the association between advanced brain aging and accelerated age-related functional decline in the older people. Second, the present study did not analyze the spatial characteristics and temporal relationships of brain age in PCDS. Longitudinal studies or other specific computational analyses are required to elucidate the regional order of abnormal brain aging in PCDS.
In conclusion, this study revealed evidence of increased BAG levels in patients with PCDS. Again, we have provided evidence supporting PCDS as a clinical phenotype of accelerated biological age for community-dwelling older people. Our results also indicate that advanced brain aging may be involved in the pathophysiology of the co-occurrence of physical and cognitive decline in the older people.

Data availability
Data supporting the present findings are available upon reasonable request.

Disclosure statement
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and /or publication of this article.