Brain Age Difference at Baseline Predicts Clinical Dementia Rating Change in Approximately Two Years

Abstract

Background:

The Clinical Dementia Rating (CDR) has been widely used to assess dementia severity, but it is limited in predicting dementia progression, thus unable to advise preventive measures to those who are at high risk.

Objective:

Predicted age difference (PAD) was proposed to predict CDR change.

Methods:

All diffusion magnetic resonance imaging and CDR scores were obtained from the OASIS-3 databank. A brain age model was trained by a machine learning algorithm using the imaging data of 258 cognitively healthy adults. Two diffusion indices, i.e., mean diffusivity and fractional anisotropy, over the whole brain white matter were extracted to serve as the features for model training. The validated brain age model was applied to a longitudinal cohort of 217 participants who had CDR = 0 (CDR0), 0.5 (CDR0.5), and 1 (CDR1) at baseline. Participants were grouped according to different baseline CDR and their subsequent CDR in approximately 2 years of follow-up. PAD was compared between different groups with multiple comparison correction.

Results:

PADs were significantly different among participants with different baseline CDRs. PAD in participants with relatively stable CDR0.5 was significantly smaller than PAD in participants who had CDR0.5 at baseline but converted to CDR1 in the follow-up. Similarly, participants with relatively stable CDR0 had significantly smaller PAD than those who were CDR0 at baseline but converted to CDR0.5 in the follow-up.

Conclusion:

Our results imply that PAD might be a potential imaging biomarker for predicting CDR outcomes in patients with CDR0 or CDR0.5.

Keywords

Brain age dementia diffusion MRI predicted age difference

INTRODUCTION

The Clinical Dementia Rating (CDR) is a semi-structure interview used for staging the severity of dementia in Alzheimer’s disease (AD) [1] or non-AD dementia [2]. Compared with neuropsychological tests, the interview focuses on patients’ daily functioning and is more congruent with standard diagnostic criteria of dementia [3]. The CDR is conducted by certified physicians or nurses and administered to patients and informants. The interview consists of 6 different domains, i.e., memory, orientation, judgment and problem solving, community affairs, home and hobbies, and personal care. The functional impairment of each domain is graded by a five-point scale system: none = 0, questionable = 0.5, mild = 1, moderate = 2, severe = 3. The scores over the 6 domains are converted by an algorithm to a global score. The CDR score of 0, 0.5, 1, 2 and 3 corresponds to cognitively normal (CN), questionable, mild, moderate, and severe dementia, respectively. It has been validated to be a reliable tool to diagnose dementia and to grade the dementia severity [4] and is included by professional societies worldwide in the guideline of dementia care and diagnosis [5].

Although the CDR is reliable to grade the dementia severity, it cannot predict the progression of dementia. This limitation is clinically relevant especially in patients with very mild or questionable dementia, i.e., CDR = 0.5. The outcome of these patients is heterogeneous, some may progress to dementia (CDR = 1 or higher), some turn to normal cognition (CDR = 0), and some stay in the same CDR [6]. To address this limitation, several studies demonstrated that the sum scores of the six domains, i.e., sum of boxes [7, 8], or the subscale score in the orientation domain [9], or combination of sum of boxes CDR scores with additional memory test [10] could be used to predict dementia progression. However, all the above measures rely heavily on interview; they are affected by clinician’s judgement, informant’s reliability, and cultural differences [3]. Therefore, a brain image marker that can predict the progression of a given CDR is of clinical interest owing to its objectivity and avoidance of human errors involved in the interview process.

Brain age is a type of projected age derived from a brain image-based machine learning algorithm [11]. It entails training of a brain age model using a large amount of brain images from a group of healthy people. Typically, features of brain structure or function are extracted from the brain magnetic resonance imaging (MRI) data and the features are used as the input to the algorithm for model training. Once the model is trained and validated, it can be used to estimate brain age of an individual, and the predicted age difference (PAD), defined as the difference between brain age and chronological age of an individual, is calculated to indicate the brain aging status. PAD has been shown to be a potential cognitive aging marker. It is associated with cognitive impairment [12], fluid intelligence [13, 14], and various cognitive abilities [15, 16]. Beheshti et al. estimated PAD using individual’s morphometric features extracted from T1-weighted (T1w) MRI [17]. They found that PAD was associated with the scores of several clinical assessment scores of dementia including CDR. Beheshti’s work demonstrated the capability of PAD in predicting CDR at the time of MRI scanning; however, they did not have longitudinal data to investigate the prediction of PAD in the CDR change in the following years.

Among all MRI modalities, T1w data are used to train the brain age models more often than other modalities such as diffusion MRI, T2*, T2 FLAIR, task or resting functional MRI [11]. It has been shown that the models trained by T1w and diffusion-MRI brain data had comparable age prediction accuracy, outperforming the models derived from other modalities [13]. To date, most of the studies employed T1w brain age models to investigate the relationships of PAD with cognitive aging or dementia [15 , 17–19]. Multiple diffusion MRI studies have observed alterations of diffusion property in patients with dementia and the alterations were associated with the degree of cognitive impairment [20 –28]. Given that diffusion-MRI brain age is an index summarizing variabilities of microstructural integrity of white matter, it is plausible that it could be a surrogate marker of cognitive impairment such as the CDR. However, there are no studies on the associations of diffusion-MRI brain age and the CDR scores, only a few studies reported the cognitive or cardiometabolic associations in healthy population [16 , 29–31].

The Open Access Series of Imaging Studies (OASIS) is a databank which collects MRI and PET data, neuropsychological testing, and clinical data. It is publicly available to the scientific community with the aim of studying the effects of healthy aging and AD. The OASIS-3 is a dataset released in 2019 which comprises longitudinal data of 1,098 participants, aged from 42 to 95 years [32]. The participants include cognitively normal adults and patients with various CDR scores. The OASIS-3 dataset provides an excellent opportunity to investigate the capacity of PAD in predicting CDR change, especially in those with normal cognition or with very mild/questionable dementia.

Prediction of a person’s CDR change in the following years is of clinical interest because appropriate intervention can be prescribed individually. Therefore, in the present study, we used cognitively normal adults in the OASIS-3 data to build a brain age model and applied the model to participants with different CDRs encompassing 0, 0.5, and 1 at baseline, i.e., the time of MRI scanning, and investigated the association of PAD with the CDR change in several years of follow-up. Specifically, three hypotheses were tested. First, PADs were different among participants with different CDR scores of 0, 0.5, and 1 at baseline. Second, PAD in participants with stable CDR of 0.5 was different from PAD in participants with the CDR of 0.5 at baseline but turned to 0 or 1 in the follow-up. Third, PAD in participants with stable CDR of 0 was different from PAD in participants whose baseline CDR was 0 but became 0.5 in the follow-up.

METHODS

Subjects

This is an observational study using longitudinal data of an elderly cohort. The participants in OASIS-3 were enrolled by several projects related to Knight ADRC. They were either generally healthy and cognitively normal (CDR = 0) individuals with or without a family history of AD, or generally healthy individuals with CDR = 0.5, 1, or 2. The exclusion criteria included medical conditions that precluded longitudinal participation and contraindications for the MRI study.

Standard protocol approvals and consents

All participants gave informed consent approved by the Institutional Review Board of Washington University School of Medicine in St. Louis.

Data availability statement

The data of OASIS-3 was obtained under the Data Use Agreement.

Data selection

The OASIS data comprised 2,168 MRI sessions collected from 3 scanners (Siemens, Erlangen, Germany), one Biograph mMR 3T PET/MR scanner and two TIM Trio 3T MR scanners. Among them, 1,246 MRI sessions had both T1w imaging and diffusion tensor imaging (DTI) data. We chose 630 MRI sessions whose DTI scans were acquired from the TIM Trio scanners to reduce inter-scanner variability. After excluding images with obvious image artefact (e.g., motion blurring or failed artefact correction), 603 MRI sessions remained. 82 MRI sessions were further excluded because they did not have CDR records around the time of MRI scanning or they did not meet the grouping criteria. Finally, 521 MRI sessions of 475 participants were enrolled in the study.

MRI data acquisition

T1w data were acquired using the 3-dimensional magnetization prepared rapid gradient echo sequence. The imaging parameters were repetition time (TR) = 2,400 ms, echo time (TE) = 3.16 ms, inversion time (TI) = 1,000 ms, field-of-view (FOV) = 256×256×176 mm³, and matrix size = 256×256×176. DTI data were acquired using diffusion-weighted 2-dimensional single shot spin-echo echo planar imaging using 24 different magnitudes of diffusion sensitivity (b-values) corresponding to 24 different directions (b-vectors), repeated twice (please see Supplementary Table 1). The imaging parameters were TR = 14,500 ms, TE = 112 ms, FOV = 224×224 mm², matrix size = 112×112, slice thickness = 2 mm. The two DTI acquisitions were stacked to form a single DTI dataset.

Image processing

The procedures of the image processing and brain age modeling are illustrated in Fig. 1. All MRI data were processed with the same pipeline as shown in Fig. 1A.

Fig. 1

The flow chart of image processing (A) and brain age modeling (B and C). A) Each individual’s MRI data were processed with the same pipeline as shown here. B) The data in the training set were used to construct the brain age model, which comprised a GPR model and a bias-correction model. C) The data in the testing set were fed to a GPR model followed by a bias-correction model to obtain the output of the brain age.

Brain segmentation on T1w image

The T1w image was processed using the Segment [33] toolbox in the SPM12 software (Wellcome Trust Centre for Neuroimaging, University College London, London, UK). The toolbox segmented T1w image into six components including gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), bone, scalp, and others, each was represented by a tissue probability map (TPM). The TPMs of GM and WM were used to conduct spatial normalization to the MNI space. The Segment toolbox also corrected the intensity inhomogeneity on the T1w image, and the corrected T1w image was used in the subsequent artefact correction on DTI data.

Artefact correction on DTI data

Raw DTI data of each individual were processed by the DACO algorithm [34] to correct various artefacts. The algorithm synthesized a pseudo b0 image by inverting the contrast of the inhomogeneity-corrected T1w image. We used the pseudo b0 image as the reference image because it had no distortion nor intensity inhomogeneity. By registering the DTI data to the pseudo b0 image, DACO corrected the artefacts of DTI data, including susceptibility-induced distortions, eddy current-induced distortions, head motions, and intensity inhomogeneity. Note that DACO also rigidly aligned the artefact-corrected DTI data with the space of T1w image.

Diffusion tensor estimation

The artefact-corrected DTI data were processed to estimate the diffusion tensor using the approach proposed by Koay et al. [35]. The weighted-linear-least-square method was first conducted, and the results were used as the initial estimation for the constrained-nonlinear-least-square method to obtain the diffusion tensors, for which positive-definite was ensured. The fractional anisotropy (FA) and mean diffusivity (MD) values were derived from the estimated diffusion tensor at each pixel using the standard formula [35].

Spatial normalization to MNI space

The TPMs of GM and WM were registered to the ICBM152 template in MNI space using a variant of the Large Deformation Diffeomorphic Metric Mapping (LDDMM) algorithm [36, 37]. The initial velocity situated in the ICBM152 space was iteratively estimated by shooting this velocity along the time dimension toward the individual’s native space. We used an isotropic Gaussian filter with 10-mm width to ensure the smoothness of the initial velocity and divided the course of registration into 10 uniform intervals. On convergence, the associated deformation map could transform the individual’s TPMs to match the ICBM152 TPM template.

Diffusion index extraction

The FA and MD maps in native space were normalized to MNI space through the deformation maps derived from the LDDMM registration. The FA and MD values in the WM pixels were extracted using the WM TPM mask in the ICBM152 template. Since FA and MD were DTI indices representing white matter microstructural property [38], they were used as the features for brain age modeling.

All the processing procedures were conducted using in-house programs in MATLAB (The MathWorks, Inc., Natick, MA, USA) except the brain segmentation that was performed using SPM12.

Grouping according to CDR track records

The participants enrolled in the study were assigned to 9 groups according to the track records of their CDR values (Table 1) and PADs were compared between groups. The 9 groups were: CN*-for-modeling, CN*-to-CN, CN-to-CN*, CN*-to-questionable, questionable-to-CN*, questionable*-to-CN, nominal questionable*, questionable*-to-mild, and nominal mild*. The asterisk (*) indicates the CDR state when MRI was scanned. All groups included different participants except CN*-to-CN and CN-to-CN*, they were the same participants undertaking the first and second MRI scanning, respectively. There were 258 participants in the CN*-for-modeling group and 217 participants in the rest of the groups.

Table 1

Demographics of patients

Group	Category	Elapsed Time (y)	Sample Size	Gender		Age (y)	Age Range (y)	MMSE	CDR (sum of box)	Education (y)	APOE ɛ4^‡	PAD (y)
					Female	Male
CN*-for-modeling	Training Data		258	57.8%	42.2%	68.54±8.75	42.66–95.20	29.24±1.17	0.01±0.06	16.22±2.47	31.9%
CN*-to-CN	Testing Data	2.91±0.67^†	46	65.2%	34.8%	66.19±8.13	49.63–77.55	29.17±0.93	0.00±0.00	16.13±2.60	34.8%	–0.86±5.48
CN-to-CN*	Statistical Analysis	2.91±0.67^†	46	65.2%	34.8%	69.10±7.86	52.47–79.64	29.09±1.15	0.00±0.00	16.13±2.60	34.8%	–0.75±5.53
CN*-to-questionable	Statistical Analysis	2.49±1.17	34	50.0%	50.0%	75.51±6.45	64.38–93.13	28.85±1.42	0.04±0.14	15.41±2.70	38.2%	3.10±7.72
questionable-to-CN*	Statistical Analysis	3.06±1.87	25	60.0%	40.0%	75.75±8.27	61.95–91.46	29.08±1.04	0.00±0.00	16.08±2.80	36.0%	2.90±7.19
questionable*-to-CN	Statistical Analysis	1.44±0.58	26	38.5%	61.5%	74.87±7.29	57.48–86.05	28.35±1.57	0.94±0.59	15.27±2.44	42.3%	2.99±7.16
nominal questionable*	Statistical Analysis		34	50.0%	50.0%	76.23±7.21	66.50–94.44	26.65±2.91	1.91±0.97	14.91±3.39	52.9%	5.48±6.68
questionable*-to-mild	Statistical Analysis	2.18±1.06	28	32.1%	67.9%	75.45±6.20	63.44–85.35	25.71±2.45	2.45±1.07	15.29±3.09	57.1%	9.21±6.14
nominal mild*	Statistical Analysis		24	33.3%	66.7%	76.40±10.31	50.98–95.58	21.13±3.89	5.46±1.11	13.71±2.93	54.2%	12.33±10.89

MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rating; CN, CDR = 0 (Cognitively Normal); questionable, CDR = 0.5; mild, CDR = 1; PAD, Predicted Age Difference. ^*The CDR state when MRI was scanned. ^†The elapsed time here represents the inter-scan interval between CN*-to-CN and CN-to-CN*. ^‡APOE ɛ4 indicates the percentage in the group population that poses either one or two ɛ4 alleles.

CN*-for-modeling

This group comprised 258 participants who were cognitively normal (CN, CDR = 0) and stable. The CDR scores were 0 in all clinical records and had at least one such record within 1 year before or after the MRI scanning. The data in this group served as the training data to build the diffusion-MRI brain age model.

CN*-to-CN

This group comprised 46 participants who were cognitively normal and stable. The CDR inclusion criteria were the same as the CN*-for-modeling group. They received two MRI scans with the inter-scan interval of 2.91±0.67 years. The data of this group comprised the data of the first MRI scans and served as the testing data of brain age modeling.

CN-to-CN*

Participants in this group (N = 46) were the same participants as those in CN*-to-CN. The data of this group comprised the data of the second MRI scans and served as the CN group for group comparison. It should be noted that participants in the CN-to-CN* group (same as CN*-to-CN) received two MRI scans and the participants were different from those in the CN*-for-modeling group.

questionable*-to-mild

Participants in this group (N = 28) were in questionable dementia (CDR = 0.5) within the interval of±180 days from the MRI scanning date but converted to mild symptomatic AD (CDR = 1) 180 days after the scanning. Most of the longitudinal studies on MCI or AD progression followed patients’ cognitive states at an interval of 1 year for several years [39, 40]. Our choice of±180 days (i.e., ±0.5 year) from the MRI scanning date took into consideration this standard follow-up interval of 1 year. When we took the CDR score within the±180 days from the MRI scanning, the CDR score in the next (or previous) visit would be 180 days after (or before) the MRI scan.

CN*-to-questionable

Participants in this group (N = 34) were cognitively normal (CDR = 0) within the interval of±180 days from the date of MRI scanning but converted to questionable dementia (CDR = 0.5) 180 days after the scanning.

questionable-to-CN*

Participants in this group (N = 25) were cognitively normal (CDR = 0) within the interval of ±180 days from the MRI scanning date but were in questionable dementia (CDR = 0.5) 180 days before the scanning.

questionable*-to-CN

Participants in this group (N = 26) were in questionable dementia (CDR = 0.5) within the interval of ±180 days from the MRI scanning date but turned to CN (CDR = 0) 180 days after the scanning.

nominal questionable*

Participants in this group (N = 34) were in questionable dementia (CDR = 0.5) within the interval of ±180 days from the MRI scanning date. After the interval, 19 participants had available clinical records showing that they stayed in the questionable dementia stage, while rest of the participants (N = 15) did not have any record of CDR available. Therefore, nominal questionable* was named for this group. It should be noted that participants whose CDR was 0.5 (i.e., questionable dementia) at the time of MRI scanning might have different CDR scores 180 days before MRI scanning. They might include, but not limited to, CN-to-questionable*, questionable-to-questionable*, or mild-to-questionable*. In this study, we included these patients and stratified them according to the CDR outcomes 180 days after MRI scanning. In this way, they were stratified into questionable*-to-mild, questionable*-to-CN, and nominal questionable*.

nominal mild*

Participants in this group (N = 24) were in mild dementia (CDR = 1) within the interval of±180 days from the MRI scanning date. After the interval, 13 participants converted to moderate symptomatic AD (CDR = 2), 6 participants had all clinical records available showing that they stayed unchanged, and 5 participants did not have any CDR record. Therefore, nominal mild* was named for this group.

Diffusion-MRI brain age

The Gaussian process regression (GPR) method was used to regress the chronological age of the participants against the FA and MD values inside the WM region. We used FA and MD of the voxels inside the white matter region to conduct the regression analysis. There were 23,722 voxels in the white matter region, and therefore, a total of 47,444 features were used in the Gaussian process regression. Gaussian process regression fits the chronological age (dependent variable) against the multi-dimensional features (independent variables) of each individual in the training group by estimating 4 unknown variables. The 4 variables include one for the basis function which is a constant, two for the kernel function in the form of a squared exponential function, and one for the noise variance. The number of unknown variables is substantially less than that of participants (N = 258), thus minimizing the overfitting issue. As described below, this diffusion-MRI brain age model also corrected the age- and sex-related bias, which was constantly observed in brain age literatures [13 –16].

We used the leave-one-out cross-validation approach to evaluate the age-related bias of the GPR model. Specifically, for a training set with N subjects, we used the image data (i.e., FA and MD inside WM) and chronological age of N-1 subjects to construct a GPR model, and then fed the image data of the left one (validation data) to the GPR model to obtain the ‘interim’ brain age. The process was repeated N times until every subject had been used as the validation data once. In this manner, every subject was associated with (x,y), where x was the ‘interim’ brain age and y was the chronological age. It has been shown that x and y have a bias such that x is over-estimated (i.e., x > y) in younger age and under-estimated (i.e., x < y) in older age. To mitigate the bias, we used a quadratic polynomial to establish a bias-correction model. Therefore, the corrected brain age x’ was $x^{'} = α x^{2} + bx + c .$

The coefficients a, b, and c were estimated by minimizing the absolute value of correlation coefficient between x’-y (i.e., corrected brain age minus chronological age) and y (i.e., chronological age). Here we constructed the bias-correction model for each sex separately to account for the sex difference. Therefore, a complete brain age model consisted of a GPR model and a bias-correction model.

Data in the CN*-for-modeling group were used to train the diffusion-MRI brain age model. After the training process, the model was applied to each MRI data in other groups to estimate the brain age and PAD. The performance of the model was tested in the CN*-to-CN group in terms of the mean absolute error (MAE) and Pearson’s correlation coefficient (r) between the brain age and chronological age. We also evaluated the model performance in the training set (CN*-for-modeling) using the leave-one-out cross validation approach.

Statistical inference

PADs were compared between any two of the seven groups, namely CN-to-CN*, questionable*-to-mild, CN*-to-questionable, questionable-to-CN*, questionable*-to-CN, nominal questionable*, and nominal mild*, amounting to 21 pairwise comparisons. For each pair of groups, two sample t-test was conducted. The age and sex were not considered as confounders in the tests because their influence had been mitigated by the bias correction conducted in the diffusion-MRI brain age model. The APOE ɛ4 status was not considered as a confounder because it has been reported that there is no significant difference in brain predicted age between ɛ4 carriers and ɛ4 non-carriers [41]. The Benjamini–Hochberg procedure with a false discovery rate of 0.05 was used to determine statistical significance of the tests.

RESULTS

Demographics

Table 1 summarizes the demographics of the 9 groups including CN*-for-modeling (training data), CN*-to-CN (testing data) and 7 comparison groups. In the 7 comparison groups, the CN-to-CN*, nominal questionable*, and nominal mild* groups were considered relatively stable because these 3 groups stayed in their CDR conditions beyond the interval of ±180 days from the MRI scanning date. In contrast, CN*-to-questionable, questionable-to-CN*, questionable*-to-CN, and questionable*-to-mild groups showed CDR change beyond the interval of ±180 days and were thus considered meta-stable. If the groups were ranked according to the severity of dementia, CN*-to-questionable, questionable-to-CN*, and questionable*-to-CN were positioned between CN-to-CN* and nominal questionable*, while questionable*-to-mild was placed between nominal questionable* and nominal mild*. With this order, a tendency was revealed showing the increase in sum of boxes CDR and PAD and decrease in MMSE from CN-to-CN* to nominal mild*. There was also a tendency of increased male proportion and APOE e4 percentage along the ranked order.

Performance of the brain age model

In the testing set prior to correcting age-related bias, the interim brain age showed MAE = 4.05 years, r = 0.77 (i.e., chronological age versus interim brain age), and r = –0.75 for age-related bias (i.e., chronological age versus interim brain age minus chronological age). After bias correction, the brain age had MAE = 4.35 years, r = 0.77 (p = 3.64×10^–10), and r = –0.15 for age-related bias. In the training set, the leave-one-out cross validation of the brain age performance gave MAE = 4.59 years, and R-squared = 0.69.

Group comparison

Table 2 summarizes the statistical results of the pairwise comparison among the seven groups. The comparison results were described in three parts as follow.

Table 2

Adjusted p-values of pairwise comparisons following two sample t-test, adjusted for multiple comparisons using the Benjamini–Hochberg method

Comparison #	Name of Group 1	PAD of Group 1 (y)	Name of Group 2	PAD of Group 2 (y)	Difference (95% CI)^†	Original p	Adjusted p	Effect Size^‡
1	CN-to-CN*	–0.75±5.53	CN*-to-questionable	3.10±7.72	–3.85 (–6.80, –0.90)	0.0112	0.0213⁺	–0.59
2	CN-to-CN*	–0.75±5.53	questionable-to-CN*	2.90±7.19	–3.65 (–6.71, –0.60)	0.0197	0.0318⁺	–0.59
3	CN-to-CN*	–0.75±5.53	questionable*-to-CN	2.99±7.16	–3.74 (–6.75, –0.72)	0.0158	0.0276⁺	–0.61
4	CN-to-CN*	–0.75±5.53	nominal questionable*	5.48±6.68	–6.23 (–8.96, –3.51)	0.0000	0.0001⁺	–1.03
5	CN-to-CN*	–0.75±5.53	questionable*-to-mild	9.21±6.14	–9.96 (–12.71, –7.21)	0.0000	0.0000⁺	–1.73
6	CN-to-CN*	–0.75±5.53	nominal mild*	12.33±10.89	–13.08 (–16.99, –9.18)	0.0000	0.0000⁺	–1.68
7	CN*-to-questionable	3.10±7.72	questionable-to-CN*	2.90±7.19	0.19 (–3.76, 4.15)	0.9219	1.0000	0.03
8	CN*-to-questionable	3.10±7.72	questionable*-to-CN	2.99±7.16	0.11 (–3.79, 4.01)	0.9555	1.0000	0.01
9	CN*-to-questionable	3.10±7.72	nominal questionable*	5.48±6.68	–2.39 (–5.88, 1.11)	0.1774	0.2192	–0.33
10	CN*-to-questionable	3.10±7.72	questionable*-to-mild	9.21±6.14	–6.11 (–9.71, –2.51)	0.0012	0.0028⁺	–0.87
11	CN*-to-questionable	3.10±7.72	nominal mild*	12.33±10.89	–9.23 (–14.12, –4.34)	0.0004	0.0020⁺	–1.01
12	questionable-to-CN*	2.90±7.19	questionable*-to-CN	2.99±7.16	–0.09 (–4.12, 3.95)	0.9663	0.9663	–0.01
13	questionable-to-CN*	2.90±7.19	nominal questionable*	5.48±6.68	–2.58 (–6.22, 1.06)	0.1610	0.2255	–0.37
14	questionable-to-CN*	2.90±7.19	questionable*-to-mild	9.21±6.14	–6.31 (–9.98, –2.63)	0.0012	0.0035⁺	–0.95
15	questionable-to-CN*	2.90±7.19	nominal mild*	12.33±10.89	–9.43 (–14.71, –4.15)	0.0008	0.0028⁺	–1.03
16	questionable*-to-CN	2.99±7.16	nominal questionable*	5.48±6.68	–2.50 (–6.09, 1.10)	0.1696	0.2227	–0.36
17	questionable*-to-CN	2.99±7.16	questionable*-to-mild	9.21±6.14	–6.22 (–9.85, –2.59)	0.0012	0.0031⁺	–0.94
18	questionable*-to-CN	2.99±7.16	nominal mild*	12.33±10.89	–9.34 (–14.54, –4.14)	0.0007	0.0030⁺	–1.02
19	nominal questionable*	5.48±6.68	questionable*-to-mild	9.21±6.14	–3.72 (–7.01, –0.44)	0.0271	0.0406⁺	–0.58
20	nominal questionable*	5.48±6.68	nominal mild*	12.33±10.89	–6.85 (–11.47, –2.22)	0.0044	0.0093⁺	–0.79
21	questionable*-to-mild	9.21±6.14	nominal mild*	12.33±10.89	–3.12 (–7.96, 1.71)	0.2006	0.2340	–0.36

CDR, Clinical Dementia Rating; CN, CDR = 0 (Cognitively Normal); questionable, CDR = 0.5; mild, CDR = 1; PAD, Predicted Age Difference; 95% CI, 95% Confidence Interval. ^*The CDR state when MRI was scanned. ^†Difference is defined as Group 1 minus Group 2. ^‡Sign of effect size results from Group 1 minus Group 2. ⁺Statistically significant according to p < 0.05.

Comparison between different CDR levels

Figure 2 shows three relatively stable groups with different CDR levels. The PADs of CN-to-CN*, nominal questionable*, and nominal mild* were –0.75±5.53, 5.48±6.68, and 12.33±10.89 years, respectively. All comparisons on the three pairs of groups were statistically significant (Table 2, #4, #6, #20).

Fig. 2

Comparisons between CN-to-CN*, nominal questionable*, and nominal mild*. The distribution plots of the CN-to-CN*, nominal questionable*, and nominal mild* groups show an increase of the mean PADs from CN-to-CN* (–0.75 years), nominal questionable* (5.48 years), to nominal mild* (12.33 years). Pairwise t-test shows significant differences between any two groups.

Comparison within participants with baseline CDR of 0.5

Figure 3 shows the participants whose CDR was 0.5 at baseline but reverted to CN (questionable*-to-CN), stayed relatively stable (nominal questionable*), or converted to mild dementia (questionable*-to-mild). PADs of questionable*-to-CN, nominal questionable*, and questionable*-to-mild were 2.99±7.16, 5.48±6.68, 9.21±6.14 years, respectively. The PAD value in questionable*-to-mild was significantly higher than that in questionable*-to-CN (adjusted p = 0.0031) and in nominal questionable* (adjusted p = 0.0406) (Table 2, #17, #19). Although PAD in questionable*-to-CN was lower than that in nominal questionable*, it did not reach the level of statistical significance (adjusted p = 0.2227) (Table 2, #16).

Fig. 3

Comparisons between questionable*-to-CN, questionable*-stable, and questionable*-to-mild. The distribution plots of the questionable*-to-CN, questionable*-stable, and questionable*-to-mild groups show an increase of the mean PADs from questionable*-to-CN (2.99 years), questionable*-stable (5.48 years), to questionable*-to-mild (9.21 years). Pairwise t-test shows significant differences between the questionable*-to-CN and questionable*-to-mild groups and between the questionable*-stable and questionable*-to-mild groups.

Comparison within participants with baseline CDR of 0

Figure 4 shows the participants who were in CN at baseline but presented relatively stable (CN-to-CN*) or unstable states (CN*-to-questionable and questionable-to-CN*). PADs of CN-to-CN*, CN*-to-questionable, and questionable-to-CN* were –0.75±5.53, 3.10±7.72, and 2.90±7.19 years, respectively. Both CN*-to-questionable (adjusted p = 0.0213) and questionable-to-CN* (adjusted p = 0.0318) had PADs significantly higher than the CN-to-CN* group (Table 2, #1, #2). There was no significant difference between the CN*-to-questionable and questionable-to-CN* groups (adjusted p = 1.0000) (Table 2, #7).

Fig. 4

Comparisons between CN-to-CN*, CN*-to-questionable, and questionable-to-CN*. The distribution plots of the CN-to-CN*, CN*-to-questionable, and questionable-to-CN* groups show an increase of the mean PADs in the questionable-to-CN* (2.90 years), and CN*-to-questionable (3.10 years) as compared with CN-to-CN* (–0.75 years). Pairwise t-test shows significant differences between the CN-to-CN* and CN*-to-questionable groups and between the CN-to-CN* and CN*-to-questionable groups.

Spectrum of PAD

Figure 5 shows the seven comparison groups sorted by their mean PADs, and we see five clusters emerged. The first cluster comprised the participants with relatively stable normal cognition (CN-to-CN*), the mean PAD of which was approximately 0 years. The second cluster included those exhibiting transitions between CN and questionable dementia (i.e., CN*-to-questionable, questionable-to-CN*, and questionable*-to-CN), the mean PADs were approximately 3 years. The third, fourth, and fifth clusters were the participants with nominal questionable*, questionable*-to-mild, and nominal mild*, respectively, with the mean PADs of 5.48, 9.21 and 12.33 years.

Fig. 5

Spectrum of PAD. The seven comparison groups sorted by the mean PADs reveal five clusters. The first cluster is CN-to-CN* with the mean PAD of –0.75 years. The second cluster consists of CN*-to-questionable, questionable-to-CN*, and questionable*-to-CN with the mean PAD around 3 years. The third, fourth and fifth clusters correspond to questionable*-stable, questionable*-to-mild, and nominal mild*, and the mean PADs are 5.48, 9.21 and 12.33 years, respectively.

DISCUSSION

The most significant finding of the present study is to demonstrate the potential of diffusion-MRI PAD in associating with the CDR change in participants with the baseline CDR of 0 or 0.5. We demonstrated that PADs were significantly different between CDR scores of 0, 0.5, and 1, supporting the first hypothesis. For the second hypothesis, we demonstrated that PAD of those with relatively stable CDR of 0.5 was significantly different from PAD of those whose CDR was 0.5 at baseline but converted to 1 in approximately 2 years. Finally, we showed that PAD of those with relatively stable CDR of 0 was significantly different from PAD with the baseline CDR of 0 but converted to 0.5 in approximately 2.5 years, supporting the third hypothesis. Our results imply that PAD might be a potential biomarker not only for grading the dementia severity, but also for predicting the change of dementia severity over time.

In this paper, we demonstrated that PAD derived from white matter microstructural property, as indicated by FA and MD of DTI, was associated with CDR of 0, 0.5, and 1 (Fig. 2). Beheshti et al. developed a PAD metric, called Brain-Age Score (BAS), using voxel-based volumetric data of cerebral gray matter [17]. When they applied BAS to a cohort of J-ADNI dataset, BAS was found to be significantly associated with CDR and other cognitive and functional scores. To the best of our knowledge, our results are the first report demonstrating that DTI-derived PAD was also associated with CDR. This report complements the capacity of PAD in predicting the CDR, with the features extracted from either gray matter volume or white matter microstructure.

PAD has been considered to be a brain health marker of an individual [11] and has been found to be associated with cognitive functions in healthy individuals [13 –16]. On the other hand, the CDR is designed to assess AD-related dementia from a patient’s daily manifestations which require key cognitive functions to support the activities. The association of PAD with the CDR in this study implies that DTI-derived PAD is not only a cognitive marker in healthy individuals; it is also a potential imaging biomarker to differentiate low to moderate dementia.

In patients with the CDR of 0.5, they presented different outcomes in approximately 1 to 2 years (Fig. 3). Our results demonstrated that PADs among subgroups of these patients were significantly different. Patients whose CDR changed from 0.5 to 1 in approximately 2 years had mean PAD of 9.21 years, significantly higher than patients with stable CDR of 0.5 (mean PAD = 5.48 years, adjusted p = 0.0406) or those with the baseline CDR of 0.5 but turned to 0 in approximately 1.5 years (mean PAD = 2.99 years, adjusted p = 0.0031). The results have clinical implications. Currently exercise and cognitive training are the interventions recommended for patients with MCI, whose CDR is mostly 0.5. However, no information is available to predict patient’s cognitive outcome, and therefore, the intervention cannot be tailored to each individual patient according to the risk of cognitive decline. Our finding implies that PAD may be useful to stratify patients with MCI into high risk group (CDR changed from 0.5 to 1) and low risk group (stable CDR of 0.5 or CDR changed from 0.5 to 0). Appropriate intervention can then be designed for different groups.

Pasternak et al. proposed a bi-tensor model to describe the DTI signal in terms of two components, i.e., free water (FW) and tissue components, with the aim of correcting the partial volume effect of the CSF [42]. Maillard et al. applied the model to a large cohort of cognitively diverse individuals [43]. They found that higher baseline FW, but not FW-corrected FA and MD, was significantly associated with higher baseline CDR, and higher probability to change to a higher CDR score in later interviews. Since our PAD was derived from FA and MD, our results seem to contradict theirs. However, the FA and MD values used in our study were not corrected for FW, so these values may be possibly driven by FW, as alluded by Maillard et al. Further studies are needed to demonstrate the associations between FW and DTI-derived PAD to clarify their relationships.

Our results also showed that participants with constant CDR of 0 had significantly different PAD from those with baseline CDR of 0 but turned to 0.5 in approximately 2.5 year (mean PAD: –0.75 years versus 3.10 years, adjusted p = 0.0213; Fig. 4). The results indicate that there is a variation of PAD even in cognitively normal people. Multiple brain age studies on cognitively normal people have demonstrated that PAD is associated with lifestyle factors, such as physical exercise, alcohol and tobacco consumption, and interpersonal relationship, and cardiovascular risks, such as blood pressure, diabetes, and body mass index [44 –48]. These studies suggest that the variation of PAD arises from different health-related risk factors among individuals, the more the risk factors, the higher the PAD. Moreover, multiple epidemiological studies have shown that poor health-related risk factors have higher risks of afflicting dementia later in life [49 –54]. Taken together, previous studies reported the association between PAD and lifestyle risk factors, and the associations between lifestyle factors and dementia. Our data suggest that for the cognitively healthy people above 60 years old (approximately mean –SD = 69.10 –7.86 years in Table 1, CN-to-CN*), those with PAD higher than 10 years (approximately mean + SD = 3.10 + 7.72 years in Table 2, CN*-to-questionable) are more likely to become cognitively impaired (CDR = 0.5) in 3 years than those with PAD lower than –6 years (approximately mean –SD = –0.75 –5.53 years in Table 2, CN-to-CN*).

We found that some of the groups in the study population exhibited relatively stable CDR scores in 2 to 3 years, i.e., CN-to-CN*, nominal questionable* and nominal mild*, while some groups presented relatively rapid transitions of the CDR, i.e., CN*-to-questionable, questionable-to-CN*, questionable*-to-CN, and questionable*-to-mild (Table 2). When labeling each group with PAD, a continuum of PAD revealed across the relatively stable and meta-stable groups (Fig. 5). PAD was smallest in the cognitively normal group (CN, –0.75 years), followed by 3 meta-stable groups (CN*-to-questionable, questionable-to-CN*, and questionable*-to-CN, presenting 3.10, 2.90, and 2.99 years, respectively), the relatively stable group of CDR = 0.5 (nominal questionable*, 5.48 years), another meta-stable group (questionable*-to-mild, 9.21 years), and the largest in the relatively stable group of CDR = 1 or beyond (nominal mild*, 12.33 years). The spectrum of PAD corresponding to the stable and meta-stable states of the CDR provides a new insight to the CDR in elderly people. Given a CDR score of 0, 0.5, or 1 of an individual, we could use PAD to predict the probability of staying cognitively normal (CDR = 0) or transitioning to a different stage (CDR = 0.5 or 1) in a few years.

Limitations

In fact, each participant presented a unique CDR trajectory. Our study cannot compare all the groups with different trajectories of the CDR. To increase the sample size in the meta-stable groups, we grouped the participants only according to the first record of change in CDR and neglected subsequent changes after the first change. For the relatively stable groups, not all participants had complete follow-up CDR records; some had records only within the predefined interval of MRI scanning and did not have records afterwards. A more fine-grained grouping would require a larger sample size. Second, our study was a retrospective study. The grouping was made based on the CDR records, and PADs between different groups were compared. Although the results were significant following rigorous statistical tests, our findings await validation with a prospective study. Moreover, the distributions of PAD still overlapped substantially between the neighboring stages (Fig. 2). A larger sample size is needed to improve the accuracy of PAD and to achieve clinically acceptable performance of prediction. Recent studies showed that the accuracy of brain predicted age can be improved by combining the image features from multiple modalities and found that the DTI and T1 w features were the most significant contributors to model performance [13, 55]. Using the data from approximately seventeen thousand participants in the UK Biobank, the best MAE for single-modality brain age model was achieved by diffusion MRI, 3.897 years in [13], and 3.733 years in [55], slightly worse than the MAE using image features across all imaging modalities, 3.515 years in [13], and 3.482 years in [55].

Other than the sample size, the accuracy of PAD is also affected by data quality which involves SNR, image processing, machine learning algorithm, and training data size. Optimization of brain age modeling is required to improve the accuracy of PAD and enhance the associations between PAD and CDR. Third, this study focused on PAD in relatively stable CDR of 0, 0.5, and 1, and the meta-stable states between neighboring scores. Whether the association still holds in patients with the CDR higher than 2 and transitional states between CDR of 2 and 3 are not clear. In addition, the number of participants in the groups of mild-to-questionable*, mild*-to-questionable, and questionable*-to-moderate, were 1, 2, and 2, respectively. PAD and following statistical analysis were not performed for these groups due to small sample size. Fourth, PAD was derived from diffusion indices of white matter. We did not estimate PAD using other imaging modalities such as T1-weighted imaging or the combined data of diffusion and T1-weighted imaging. The comparison among PADs based on different imaging modalities are beyond the scope of this study.

Conclusions

For the elderly with the CDR of 0, 0.5, or 1, DTI-derived PAD corresponds to CDR scores, and PAD differs between people with relatively stable CDR and those with CDR changed to higher scores in approximately 2 years. Our study suggest that PAD might be a potential imaging biomarker for grading the dementia severity or predicting the change of severity in the following years. Prospective studies using a longitudinal design is warranted to validate the capability of PAD as demonstrated in this study.

Footnotes

ACKNOWLEDGMENTS

Data were provided by OASIS-3: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P50 AG00561, P30 NS09857781, P01 AG026276, P01 AG003991, R01 AG043434, UL1 TR000448, R01 EB009352.

Authors’ disclosures available online ().

The supplementary material is available in the electronic version of this article: .

References

Hughes

, Berg

, Danziger

, Coben

, Martin

(1982) A new clinical scale for the staging of dementia. Br J Psychiatry 140, 566–572.

Kazui

, Yoshiyama

, Kanemoto

, Suzuki

, Sato

, Hashimoto

, Ikeda

, Tanaka

, Hatada

, Matsushita

, Nishio

, Mori

, Tanimukai

, Komori

, Yoshida

, Shimizu

, Matsumoto

, Mori

, Kashibayashi

, Yokoyama

, Shimomura

, Kabeshita

, Adachi

, Tanaka

(2016) Differences of behavioral and psychological symptomsof dementia in disease severity in four major dementias. PLoSOne 11, e0161092.

Lim

, Chong

, Sahadevan

(2007) Utility of the clinical dementia rating in Asian populations. Clin Med Res 5, 61–70.

Morris

(1993) The Clinical Dementia Rating (CDR): Current version and scoring rules. Neurology 43, 2412–2414.

Petersen

, Lopez

, Armstrong

, Getchius

TSD

, Ganguli

, Gloss

, Gronseth

, Marson

, Pringsheim

, Day

, Sager

, Stevens

, Rae-Grant

(2018) Practice guideline update summary: Mild cognitive impairment: Report of the Guideline Development, Dissemination, and Implementation Subcommittee of the American Academy of Neurology. Neurology 90, 126–135.

Daly

, Zaitchik

, Copeland

, Schmahmann

, Gunther

, Albert

(2000) Predicting conversion to Alzheimer disease using standardized clinical information. Arch Neurol 57, 675–680.

O’Bryant

, Waring

, Cullum

, Hall

, Lacritz

, Massman

, Lupo

, Reisch

, Doody

(2008) Staging dementia using Clinical Dementia Rating Scale Sum of Boxes scores: A Texas Alzheimer’s research consortium study. Arch Neurol 65, 1091–1095.

Woolf

, Slavin

, Draper

, Thomassen

, Kochan

, Reppermund

, Crawford

, Trollor

, Brodaty

, Sachdev

(2016) Can the clinical dementia rating scale identify mild cognitive impairment and predict cognitive and functional decline? Dement Geriatr Cogn Disord 41, 292–302.

Kim

, Byun

, Sohn

, Yi

, Seo

, Choe

, Kim

, Choi

, Lee

, Chee

, Woo

, Lee

(2017) Clinical Dementia Rating orientation score as an excellent predictor of the progression to Alzheimer’s disease in mild cognitive impairment. Psychiatry Investig 14, 420–426.

10.

Lee

, Youn

, Choo

, Kim

, Jhoo

, Pak

, Suh

, Woo

(2006) Combination of clinical and neuropsychologic information as a better predictor of the progression to Alzheimer disease in questionable dementia individuals. Am J Geriatr Psychiatry 14, 130–138.

11.

Cole

, Franke

(2017) Predicting age using neuroimaging: Innovative brain ageing biomarkers. Trends Neurosci 40, 681–690.

12.

Liem

, Varoquaux

, Kynast

, Beyer

, Kharabian Masouleh

, Huntenburg

, Lampe

, Rahim

, Abraham

, Craddock

, Riedel-Heller

, Luck

, Loeffler

, Schroeter

, Witte

, Villringer

, Margulies

(2017) Predicting brain-age from multimodal imaging data captures cognitive impairment. Neuroimage 148, 179–188.

13.

Cole

(2020) Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical, lifestyle, and cognitive factors. Neurobiol Aging 92, 34–42.

14.

Cole

, Ritchie

, Bastin

, Valdés Hernández

, Muñoz Maniega

, Royle

, Corley

, Pattie

, Harris

, Zhang

, Wray

, Redmond

, Marioni

, Starr

, Cox

, Wardlaw

, Sharp

, Deary

(2018) Brain age predicts mortality. Mol Psychiatry 23, 1385–1392.

15.

Boyle

, Jollans

, Rueda-Delgado

, Rizzo

, Yener

, McMorrow

, Knight

, Carey

, Robertson

, Emek-Savaş

, Stern

, Kenny

, Whelan

(2020) Brain-predicted age difference score is related to specific cognitive functions: A multi-site replication analysis. Brain Imaging Behav 15, 327–345.

16.

Richard

, Kolskår

, Sanders

, Kaufmann

, Petersen

, Doan

, Monereo Sánchez

, Alnæs

, Ulrichsen

, Dørum

, Andreassen

, Nordvik

, Westlye

(2018) Assessing distinctpatterns of cognitive aging using tissue-specific brain ageprediction based on diffusion tensor imaging and brain morphometry. Peer J 6, e5908.

17.

Beheshti

, Maikusa

, Matsuda

(2018) The association between “Brain-Age Score” (BAS) and traditional neuropsychological screening tools in Alzheimer’s disease. Brain Behav 8, e01020.

18.

Franke

, Gaser

(2019) Ten years of brainage as a neuroimaging biomarker of brain aging: What insights have we gained? Front Neurol 10, 789.

19.

Gaser

, Franke

, Klöppel

, Koutsouleris

, Sauer

(2013) Brainage in mild cognitive impaired patients: Predicting the conversion to Alzheimer’s disease. PLoS One 8, e67346.

20.

Chen

, Chen

, Tseng

, Hsu

, Huang

, Chan

, Tseng

(2021) Differential associations of white matter brain age with language-related mechanisms in word-finding ability across the adult lifespan. Front Aging Neurosci 13, 701565.

21.

Beck

, de Lange

, Maximov

, Richard

, Andreassen

, Nordvik

, Westlye

(2021) White matter microstructure across the adult lifespan: A mixed longitudinal and cross-sectional study using advanced diffusion models and brain-age prediction. Neuroimage 224, 117441.

22.

de Lange

, Anatürk

, Suri

, Kaufmann

, Cole

, Griffanti

, Zsoldos

, Jensen

DEA

, Filippini

, Singh-Manoux

, Kivimäki

, Westlye

, Ebmeier

(2020) Multimodal brain-age prediction and cardiovascular risk: The Whitehall II MRI sub-study. Neuroimage 222, 117292.

23.

Liao

, Dang

, Li

, Bu

, Han

, Jiang

(2019) Microstructural damage of normal-appearing white matter in subcortical ischemic vascular dementia is associated with Montreal cognitive assessment scores. J Int Med Res 47, 5723–5731.

24.

Zavaliangos-Petropulu

, Nir

, Thomopoulos

, Reid

, Bernstein

, Borowski

, Jack

Jr , Weiner

, Jahanshad

, Thompson

(2019) Diffusion MRI indices and their relation to cognitive impairment in brain aging: The updated multi-protocol approach in ADNI3. Front Neuroinform 13, 2.

25.

Fan

, Lai

, Chen

, Hsu

, Chen

, Huang

, Cheng

, Tseng

, Hua

, Chen

, Chiu

(2018) Diminution of context association memory structure in subjects with subjective cognitive decline. Hum Brain Mapp 39, 2549–2562.

26.

Chang

, Yen

, Chen

, Yan

, Tseng

(2016) Clinical dementia rating scale detects white matter changes in older adults at risk for Alzheimer’s disease. J Alzheimers Dis 50, 411–423.

27.

Chang

, Chen

, Shih

, Chiu

, Yan

, Tseng

(2015) Regional cingulum disruption, not gray matter atrophy, detects cognitive changes in amnestic mild cognitive impairment subtypes. J Alzheimers Dis 44, 125–138.

28.

Zhao

, Wu

, Shi

, Xia

, Sun

(2014) Relationship between cognitive impairment and apparent diffusion coefficient values from magnetic resonance-diffusion weighted imaging in elderly hypertensive patients. Clin Interv Aging 9, 1223–1231.

29.

Chang

, Chao

, Hsu

, Chen

, Tseng

(2021) White matter network disruption and cognitive correlates underlying impaired memory awareness in mild cognitive impairment. Neuroimage Clin 30, 102626.

30.

Chang

, Chen

, Tseng

(2020) White matter pathways underlying Chinese semantic and phonological fluency in mild cognitive impairment. Neuropsychologia 149, 107671.

31.

Luo

, Tseng

, Chang

(2019) White matter microstructure disruptions mediate the adverse relationships between hypertension and multiple cognitive functions in cognitively intact older adult. Neuroimage 197, 109–119.

32.

LaMontagne

, Benzinger

TLS

, Morris

, Keefe

, Hornbeck

, Xiong

, Grant

, Hassenstab

, Moulder

, Vlassenko

, Raichle

, Cruchaga

, Marcus

(2019) OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medRxiv, doi: 10.1101/2019.12.13.19014902.

33.

Ashburner

, Friston

(2005) Unified segmentation. Neuroimage 26, 839–851.

34.

Hsu

, Tseng

WYI

(2021)DACO: Distortion/artefact correction for diffusion MRI data in an integrated framework. bioRxiv, doi: 10.1101/2021.07.06.450481.

35.

Koay

, Chang

, Pierpaoli

, Basser

(2007) Error propagationframework for diffusion tensor imaging via diffusion tensorrepresentations. IEEE Trans Med Imaging 26, 1017–1034.

36.

Beg

, Miller

, Trouvé

, Younes

(2005) Computing largedeformation metric mappings via geodesic flows of diffeomorphisms. Int J Comput Vis 61, 139–157.

37.

Miller

, Trouvé

, Younes

(2006) Geodesic shooting forcomputational anatomy. J Math Imaging Vis 24, 209–228.

38.

Alexander

, Hurley

, Samsonov

, Adluru

, Hosseinbor

, Mossahebi

, Tromp do

, Zakszewski

, Field

(2011) Characterization of cerebral white matter properties using quantitative magnetic resonance imaging stains. Brain Connect 6, 423–446.

39.

Franzmeier

, Koutsouleris

, Benzinger

, Goate

, Karch

, Fagan

, McDade

, Duering

, Dichgans

, Levin

, Gordon

, Lim

, Masters

, Rossor

, Fox

, O’Connor

, Chhatwal

, Salloway

, Danek

, Hassenstab

, Schofield

, Morris

, Bateman

, Alzheimer’s Disease Neuroimaging Initiative, Dominantly Inherited Alzheimer Network, Ewers

(2020) Predicting sporadic Alzheimer’s disease progression via inherited Alzheimer’s disease-informed machine-learning. Dement 16, 501–511.

40.

Zandifar

, Fonov

, Ducharme

, Belleville

, Collins

, Alzheimer’s Disease Neuroimaging Initiative (2020) MRI and cognitive scores complement each other to accurately predict Alzheimer’s dementia 2 to 7 years before clinical onset. Neuroimage Clin 25, 102121.

41.

Löwe

, Gaser

, Franke

, Alzheimer’s Disease Neuroimaging Initiative (2016) The effect of the APOE genotype on individual BrainAGE in normal aging, mild cognitive impairment, and Alzheimer’s disease. PLoS One 11, e0157514.

42.

Pasternak

, Sochen

, Gur

, Intrator

, Assaf

(2009) Free water elimination and mapping from diffusion MRI. Magn Reson Med 62, 717–730.

43.

Maillard

, Fletcher

, Singh

, Martinez

, Johnson

, Olichney

, Farias

, DeCarli

(2019) Cerebral white matter free water: A sensitive biomarker of cognition and function. Neurology 92, e2221–e2231.

44.

Hatton

, Franz

, Elman

, Panizzon

, Hagler

Jr , Fennema-Notestine

, Eyler

, McEvoy

, Lyons

, Dale

, Kremen

, Negative fateful life events in midlife and advanced predicted brain aging. Neurobiol Aging 67, 1–9.

45.

Kolbeinsson

, Filippi

, Panagakis

, Matthews

, Elliott

, Dehghan

, Tzoulaki

(2020) Accelerated MRI-predicted brain ageing and its associations with cardiometabolic and brain disorders. Sci Rep 10, 19940.

46.

Ning

, Zhao

, Matloff

, Sun

, Toga

(2020) Association of relative brain age with tobacco smoking, alcohol consumption, and genetic variants. Sci Rep 10, 10.

47.

Ronan

, Alexander-Bloch

, Wagstyl

, Farooqi

, Brayne

, Tyler

; Cam-CAN, Fletcher

(2016) Obesity associated with increased brain age from midlife. Neurobiol Aging 47, 63–70.

48.

Steffener

, Habeck

, O’Shea

, Razlighi

, Bherer

, Stern

(2016) Differences between chronological and brain age are related to education and self-reported physical activity. Neurobiol Aging 40, 138–144.

49.

Akbaraly

, Singh-Manoux

, Dugravot

, Brunner

, Kivimaki

, Sabia

(2019) Association of midlife diet with subsequent risk for dementia. JAMA 321, 957–968.

50.

Cations

, Withall

, Low

, Draper

(2016) What is the role of modifiable environmental and lifestyle risk factors in young onset dementia? Eur J Epidemiol 31, 107–124.

51.

Fayosse

, Nguyen

, Dugravot

, Dumurgier

, Tabak

, Kivimäki

, Sabia

, Singh-Manoux

(2020) Risk prediction models for dementia: Role of age and cardiometabolic risk factors. BMC Med 18, 107.

52.

Kivipelto

, Ngandu

, Laatikainen

, Winblad

, Soininen

, Tuomilehto

(2006) Risk score for the prediction of dementia risk in 20 years among middle aged people: A longitudinal, population-based study. Lancet Neurol 5, 735–741.

53.

Mukadam

, Sommerlad

, Huntley

, Livingston

(2019) Population attributable fractions for risk factors for dementia in low-income and middle-income countries: An analysis using cross-sectional survey data. Lancet Global Health 7, e596–e603.

54.

WHO (2019) Risk reduction of cognitive decline and dementia: WHO guidelines.

55.

Salih

, Galazzo

, Jaggi

, Estabragh

, Petersen

, Lekadir

, Radeva

, Menegaz

(2020) Multi-modal brain age estimation: A comparative study confirms the importance of microstructure. In Computational Diffusion MRI. Mathematics and Visualization, Gyori N, Hutter J, Nath V, Palombo M, Pizzolato M, Zhang F, eds. Springer, Cham, pp. 239–250.