Cognitively defined Alzheimer's dementia subgroups have distinct atrophy patterns

Abstract INTRODUCTION We sought to determine structural magnetic resonance imaging (MRI) characteristics across subgroups defined based on relative cognitive domain impairments using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and to compare cognitively defined to imaging‐defined subgroups. METHODS We used data from 584 people with Alzheimer's disease (AD) (461 amyloid positive, 123 unknown amyloid status) and 118 amyloid‐negative controls. We used voxel‐based morphometry to compare gray matter volume (GMV) for each group compared to controls and to AD‐Memory. RESULTS There was pronounced bilateral lower medial temporal lobe atrophy with relative cortical sparing for AD‐Memory, lower left hemisphere GMV for AD‐Language, anterior lower GMV for AD‐Executive, and posterior lower GMV for AD‐Visuospatial. Formal asymmetry comparisons showed substantially more asymmetry in the AD‐Language group than any other group (p = 1.15 × 10−10). For overlap between imaging‐defined and cognitively defined subgroups, AD‐Memory matched up with an imaging‐defined limbic predominant group. DISCUSSION MRI findings differ across cognitively defined AD subgroups.


BACKGROUND
A personalized medicine approach has been recommended for Alzheimer's disease (AD). 1,2One success story for personalized medicine is breast cancer.Separating people with "breast cancer" into groups based on receptors is commonplace today; subtype-specific treatments have reduced mortality. 3similar strategy may be applicable to AD dementia.Heterogeneity among people with AD dementia may impede research progress into prevention and treatment.4 Identifying more homogeneous subgroups of people with AD dementia may be important in advancing research and may ultimately have therapeutic implications.
We previously developed an approach to subtyping people with typical late-onset AD dementia based on relative impairments across cognitive domains. 5Neuropsychologists use relative impairments to help determine dementia type. 6We extend this paradigm to further characterize relative impairments across domains within the AD dementia spectrum.
We have found differences between cognitively defined subgroups in terms of genetic markers (multiple cohorts), 7 associations with depression (three cohorts), 8 trajectories of glucose metabolism in Alzheimer's Disease Neuroimaging Initiative (ADNI) participants with FDG-PET scans, 9 and regional gray matter volumes (GMVs) from a tertiary memory clinic cohort. 10 this paper, we use ADNI's structural magnetic resonance imaging (MRI) data to compare GMV patterns across cognitively defined subgroups.We identified amyloid-positive or -unknown individuals who enrolled in ADNI with AD dementia or who developed AD dementia while enrolled.We compared GMVs across subgroups and with those from stable amyloid-negative cognitively normal (CN) older adults.We used an atlas-based approach to formally evaluate asymmetry.We stratified each group into quartiles based on overall atrophy to eval-uate pseudo-progression within subgroups.Finally, we compared our cognitively defined subgroups to anatomically defined subgroups 11,12 and to subgroups identified based on data-driven differences in structural MRI findings. 12

Participants
Data are from ADNI.ADNI detailed methods are published. 13,14ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD.ADNI's primary goal has been to test whether serial MRI, positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early AD dementia.For up-to-date information, see www.adni-info.org.
We included individuals who were enrolled with AD dementia (prevalent) or who progressed to AD dementia (incident) in any phase of ADNI up to ADNI3.A flow chart detailing exclusions and the final sample is in Figure S1.We included those with onset age ≥ 65 years.
We considered data from the first dementia visit: the first study visit (prevalent cases) and the visit AD dementia was diagnosed (incident cases).We used MRI data closest to the first dementia visit and excluded participants without a scan that passed quality control (see below) within 1.5 years of that visit (20 participants excluded).
All included AD dementia participants had a stable diagnosis following the first dementia visit (13 people excluded for MCI reversion).
CN controls were those who had normal cognition at enrollment and who continued without converting to MCI or to dementia during ADNI.
We used established cut-offs to determine amyloid positivity.We included people with AD dementia who were known to be amyloid positive (see below) or for whom amyloid status was unknown (n = 123); we excluded 56 people defined by ADNI as having AD dementia who were amyloid negative.
Amyloid levels change slowly, 15 19 ), and 166 using both CSF and PET; in all there were 461 amyloid-positive cases.There were 17 people with AD dementia who had discrepant amyloid status between CSF and PET; all were included as amyloid-positive individuals here.Of these, 14 were amyloid positive based on CSF but not PET, and three were amyoid positive based on PET but not CSF.
Amyloid negative CN controls used the same cut-offs.There were 30 based on CSF, 27 based on [ 18 F]florbetapir, and 61 with both CSF and PET, totaling 118 controls.

Diversity, equity, and inclusion (DEI)
ADNI had no exclusion criteria for any group defined on the basis of ethnicity or race.ADNI enrollment was characterized by overrepresentation of people with European ancestry.Current ADNI funding focuses specifically on enhancing diversity in new enrollees.The present analyses are of data from the earlier parts of the ADNI study.

Cognitive measures and subgrouping
][9][10] As previously and as discussed in those prior publications, we did not include the attention domain.
ADNI assessed cognition with a full neuropsychological battery using measures of memory, executive function, language, and visuospatial abilities.An expert panel (ET, JM, AS, PC) considered each item administered and assigned it to one domain (memory, executive function, language, and visuospatial function) or "other."We used bifactor confirmatory factor analysis approaches using Mplus 20 to generate composite scores for each domain.We co-calibrated ADNI data with those from other studies.We used scores from 825 people with incident AD dementia from Adult Changes in Thought (ACT) to define the mean at 0 and standard deviation (SD) at 1, as described. 7We excluded eight cases due to insufficient cognitive data for all four domain scores (Figure S1).
Our approach to subgrouping is schematically illustrated in Figure S2.For each case, we determined the average of memory, executive functioning, language, and visuospatial scores.We then determined the difference between each domain score and that average.As pub- lished, we used a difference of 0.80 units to identify domains with scores substantially lower than the individual average.We considered the number of domains substantially lower than the individual average.
Those with no such domains (ie, all scores similar) were AD-No Domain.
Those with a single such domain were categorized as AD-Memory, AD-Language, AD-Visuospatial, or AD-Executive.Those with multiple such domains were categorized as AD-Multiple Domains.

MRI processing
Structural MRI data were downloaded from www.adni.loni.usc.edu.
Scans were corrected prior to download as described 21 22 for ADNI-1 and ADNIGO/2 scans.For ADNI-3 ADNI is no longer generating corrected scans due to improved scan quality.Scans were processed using voxel-based morphometry (VBM) in SPM12 with DARTEL.Briefly, using a standard DARTEL-based SPM12 processing pipeline, scans were segmented into CSF, white matter volumes, and GMVs.Segmented scans were rigidly aligned to a T1 template to ensure overlap in Montreal Neurological Institute (MNI) space.These aligned scans were then co-registered using non-linear and high-dimensional warping, smoothed with an 8-mm full-width-at-half-maximum isotropic Gaussian kernel, modulated to preserve tissue volume signal, and spatially normalized to MNI space. 10Quality control via visual inspection was done after every pre-processing stage; 36 people with AD dementia were excluded based on these checks.

W-scores
Regional and global atrophy is operationalized by W-scores, which represent covariate-adjusted Z-scores normalized against CN controls. 23r each voxel we determined gray matter (GM) density distribution controlling for age, sex, field strength, and intracranial volume.We used these findings to determine W-scores for each voxel for each included participant.We used mean W-scores to account for overall progression in subsequent models.

Voxel-based morphometry
We compared demographic, neuropsychological, and clinical variables with linear regression and chi-squared tests.We compared normalized GM volume images on a voxel-by-voxel basis across groups using a one-way analysis of covariance covaried for age at scan, sex, years of education, total intracranial volume, field strength (1.5T vs 3T), and global mean W-score. 23Statistical maps were generated at a p < 0.05 threshold with cluster-wise multiple comparison correction (voxel-wise threshold p < 0.001, minimum cluster size 840 voxels).We displayed beta maps using MRIcronGL (https://www.nitrc.org/plugins/mwiki/index.php/mricrogl:MainPage). Spatial maps representing CN control/AD subgroup differences were displayed at the same minimum and maximum threshold (β = 0 to 0.063); all AD subgroup differences used β = 0 to 0.035.

Asymmetry analyses
We used FreeSurfer version 5.1 to create regions of interest (ROIs) based on the Desikan-Killiany atlas. 24We extracted ROI GMVs for each subgroup and for 30 randomly selected CN controls.We determined adjusted mean ROI GMV with linear regression, controlling for age, sex, handedness, total intracranial volume, field strength, and global mean W-score.We used adjusted mean ROI GMVs to calculate asymmetry metrics for each region: (left GMV − right GMV)/(left GMV + right GMV).This formula is negative when the left-side volume is smaller than the right, and positive when the left is larger.For each region, we fit regression models with robust standard errors using a subgroup indicator, with CN controls as reference.Tabulated values are standardized coefficients; bold indicates p < 0.05.All the findings reported represent differences compared to the amount of asymmetry found in the CN controls.Any differences in the number of voxels in a region between the right and left sides would be reflected by the finding for the CN controls and would not explain differences across subgroups with respect to that reference category.We compared the number of statistically significantly regional differences by subgroup using Fisher's exact test.

Pseudo-progression
Within each group we determined quartiles of overall atrophy based on W-score voxel count, the number of voxels with W ← 1.5.We plotted mean GMV maps for each quartile of each subgroup using MRIcroGL.

2.9
Comparison of cognitively defined subgroups to other systems of subgrouping Murray et al. proposed to differentiate people with AD based on neuropathology data. 25They quantified tau tangles in hippocampus and neocortex and defined a "limbic-predominant" subtype with high hippocampal and low neocortical tau loads, a "hippocampal sparing" subtype with low hippocampal and high neocortical tau loads, and a "typical" group with similar hippocampal and neocortical tau loads. 25[27][28][29][30][31][32][33][34][35][36][37] We used the Risacher et al. approach to categorize people into anatomically defined subgroups. 11We considered the same scanning occasion discussed previously.We used multinomial logistic regression models with AD-No Domain as reference.We excluded AD-Executive (the smallest cognitively defined subgroup); no individual with AD-Executive was categorized as limbic predominant.
We were curious as to the stability of anatomically defined subtypes over time in ADNI.We evaluated first and most recent ADNI scans and considered subgroup stability from enrollment to the first dementia visit and from first dementia visit to the most recent study visit.
Finally, we obtained group assignments from Poulakis et al. 12 They used a Bayesian clustering approach with longitudinal structural imaging data.We used multinomial logistic regression to compare cognitively defined subgroups to the approach used by Poulakis et al.

Standard protocol approvals and patient consents
All data are from ADNI.All ADNI participants signed informed consent forms.University of Washington Institutional Review Board approval is STUDY00008205.

Demographic and clinical characteristics
There were 584 people with AD dementia and 118 CN controls included.

Subgroup voxel-based morphometry analyses compared with cognitively normal controls
Several subgroups' VBM findings compared with CN controls appeared similar.Figure 1   The AD-Language subgroup differed from this pattern (Figure 2).
We noted asymmetrical GMV differences.P values are in Figure S12.
β coefficient comparisons for all groups are shown in Figure S13, and p values for all groups are shown in Figure S14.

Voxel-based morphometry analyses of subgroups compared with AD-Memory subgroup
Contrasts between subgroups were apparent.Figure 3 shows AD-No Domain versus AD-Memory groups.Figure 3A shows voxels where the AD-Memory group has lower GMV than AD-No Domain; Figure 3B shows voxels where the AD-No Domain group has lower GMV than the AD-Memory group.
Medial temporal involvement compared to controls was evident for both these groups (Figures 1 and S4), and Figure 3A shows even greater involvement for the AD-Memory group.Figure 3B shows broad cortical involvement outside medial temporal lobes where AD-No Domain has lower GMV than AD-Memory.Figure S15 shows corresponding p value comparisons.
Figure 4 shows comparisons between AD-Memory and AD-Language.We observed greater hippocampal involvement for AD-Memory bilaterally, but more pronounced on the right.Large portions of the right hemisphere have lower GMV in AD-Memory, while there are left temporal cortical areas with lower GMV in AD-Language.

F I G U R E 4
Comparison of gray matter volume for AD-Memory group with AD-Language group.*Marginal totals for each subgroup are shown at bottom; for example, 17% of the people with Alzheimer's disease (AD) dementia who met the inclusion criteria were defined as having limbic-predominant AD.We used the scan at the first AD dementia visit to determine subtypes.These data are also tabulated in Table S2.
Figure S19 shows comparisons of AD-Memory and AD-Executive.
There was lower GMV in bilateral hippocampi in AD-Memory and lower cortical GMV anteriorly in AD-Executive.Figure S20 shows comparisons of AD-Memory and AD-Multiple Domains.While not as stark as contrasts with other subgroups, there was greater hippocampal involvement in AD-Memory.There were scattered areas of cortex with greater involvement in AD-Multiple Domains.

Comparison with previously published findings
All these findings closely replicate those previously reported from a distinct study population. 10

Regional asymmetry comparisons
Distinct asymmetry for the AD-Language group compared with CN controls (Figure 2) and compared with AD-Memory (Figure 4) led us to consider formal asymmetry tests.Selected results from these analyses for right-handed individuals are shown in Table 2.The full results of all of the asymmetry analyses performed are in Tables S1 and S2.
AD-Language had four regions with statistically significant differences when limited to right-handed people, and there were no such regions in any of the other groups (Fisher's exact test p = 0.004).

TA B L E 2
Asymmetry results for right-handed people in Z-score units.Negative numbers occur with lower volume on left compared with right.

AD-Multiple Domains
Deep

Mean β coefficient maps for quartiles of each subgroup
We sorted each group by quartiles of voxels with W-scores < −1.5.
Table S3 shows thresholds between quartiles for each group.There was considerable overlap between the bottom quartiles in each group.In each AD subgroup, the 25th percentile ranged from 19,000 to 34,000 voxels, comparable to the controls' 75th percentile (21,000 voxels).AD-Memory had a restricted range of atrophy compared to other groups; the median (35,000 voxels) and 75th percentile (65,000 voxels) were lower than those of any other subgroup (medians 44,000-77,000, and 75th percentiles 78,000-122,000).Figure 5 shows β coefficient maps for each quartile for each subgroup compared to controls.In Figure 5, the left side of the coronal slices is to the left.

Comparison of cognitively defined and atrophy-defined subgroups
We used MRI scans at the dementia visit to determine anatomically defined subgroups, as published. 11Table S4 and Figure S26 show comparisons with cognitively defined subgroups.
We used multinomial logistic regression to evaluate associations across groupings.We used relationships of anatomically defined subgroups for AD-No Domains as reference.For each cognitively defined subgroup there are comparisons of the ratio of people with hippocampal sparing subtype (and limbic predominant subtype) to those for the typical AD subtype to those for the AD-No Domain subgroup (Table S5).Compared to AD-No Domain, the limbic-predominant pattern was associated with higher AD-Memory risk (relative risk ratio [RRR] 2.6, p = 4.6 × 10 −4 ), and the hippocampal sparing pattern was associated with lower AD-Memory risk (RRR 0.34, p = 0.0034).
Regarding the stability of anatomically defined subgroups over time, for people with incident dementia, we considered the enrollment and dementia visit scans (Figure S27).In all, 21% had a different anatomically defined subgroup at enrollment than at the dementia visit.In this subset, 18 of 47 (37%) with the hippocampal sparing subtype at enrollment had the typical AD subtype at dementia diagnosis, and 17 of 46 (37%) with the hippocampal sparing subtype at dementia diagnosis had the typical AD subtype at the first study visit (Figure S27).
Similarly, 11 of 41 with the limbic-predominant subtype at enrollment had the typical AD subtype at dementia diagnosis (27%), and 13 of 43 with the limbic-predominant subtype at dementia diagnosis had the typical Alzheimer's subtype at enrollment (30%; Figure S27).Similarly, considering subgroups beginning at dementia diagnosis, there was considerable movement across subtypes from the dementia visit to the most recent follow-up (Figure S28).In all, 19% had a different anatomically defined subgroup at the dementia visit and the most recent visit.
There were 248 people from our sample who had a subgroup defined by Poulakis et al.'s hierarchical Bayesian clustering approach applied to longitudinal imaging data. 12Tables S6 and S7 show that there were no significant associations between Poulakis et al.'s subgroups and cognitively defined subgroups.

DISCUSSION
We found different GMV patterns across cognitively defined subgroups.AD-Memory had relative cortical sparing and severe medial temporal atrophy compared to other subgroups.AD-Language had notable left-greater-than-right atrophy, with statistically greater atrophy across much of the brain.These findings strongly replicate findings from a distinct study population.
Our prior study from Vrije Universiteit Medical Center (VUMC) Amsterdam also evaluated people with posterior cortical atrophy and with logopenic variant primary progressive aphasia.AD-Language's regional findings resembled those of people with logopenic primary progressive aphasia, though there was more medial temporal lobe involvement in AD-Language.In a formal voxel-level comparison, AD-Language's atrophy patterns were more highly correlated with logopenic primary progressive aphasia than with any other AD dementia subgroup.Similarly, AD-Visuospatial's findings resembled those of posterior cortical atrophy, though there was more medial temporal involvement in AD-Visuospatial.At the voxel level, AD-Visuospatial had high correlations with posterior cortical atrophy. 10-Language's left-predominant atrophy pattern was consistent between ADNI and VUMC Amsterdam (Figure S23).We performed formal asymmetry analyses using average values over atlas-defined ROIs.We found strong statistical support for asymmetry (Table 2).
Identifying ways to subdivide AD is an important recent area of research; the theory-based cognitive approach we use is one of several candidate approaches.Notably, several anatomy-based approaches -including the Poulakis et al. approach 12 -in a first step average data from the left and right sides.While this reduces the number of regions considered, it makes it impossible to discern a subgroup with pronounced asymmetry.
Sarica et al. 38 reviewed studies addressing asymmetry in people with AD compared with controls, including cortical thickness, cortical volumes, and cortical surface area, as well as white matter properties and functional connectivity.The details of these studies along with citations are provided in Supplementary Text 1.Most of these studies showed that structures on the left side had more atrophy than the right.Almost all of these studies have considered people with AD as a single group in their analyses.[41] While this could explain differences between people with AD dementia and controls, it cannot explain differences across cognitively defined subgroups of people with AD dementia.
A few studies have considered asymmetry and particular cognitive domains.Keilp et al. evaluated perfusion deficits and performance in specific cognitive domains. 42Derflinger et al. found that faster left hemisphere degeneration was associated with worse performance in language-based cognitive tests across MCI and AD dementia. 39Frings et al. evaluated a sample of referred patients at a German specialty center and found that people with AD dementia with predominant language deficits exhibited more left-lateralized Aβ burden based on [ 11 C]PiB PET scans and hypometabolism based on [ 18 F]FDG PET scans compared to people with AD dementia with predominant visuospatial impairment. 43Frings et al. did not have MRI scans on most of the people in their sample and highly recommended similar analyses in the ADNI dataset. 43The approaches to categorizing people with AD in Frings et al. is similar to that performed here and in our previous analyses of the Amsterdam University Medical Center cohort.The pattern of findings for those with relative language impairments reported in Frings et al. is similar to our findings.Intriguingly, in their analyses of tau deposition using tau PET scans, Vogel et al. identified one group with prominent left greater than right tau involvement in their discovery sample but not in a replication analysis. 36e implication of prominent asymmetric findings in a subset of people with typical late-onset AD dementia is the possibility that neuropathology findings at autopsy may differ on the right and left sides.
This has been evaluated with autopsy data.King et al. found asymmetric pTDP-43 and plaque and tangle pathology in some people with clinical diagnoses of typical late-onset AD dementia. 44Similarly, Stefanitis et al. found some left/right asymmetry in tau staining in some cases of AD dementia. 45There were insufficient cognitive

P
values are from t tests or F tests except for those marked with an asterisk, which are from Fisher's exact tests.F I G U R E 1 Comparison of gray matter volume β coefficients for AD-No Domains group and cognitively normal elderly controls.was substantial overlap in age.Educational attainment was lower for people with AD dementia (p = 0.016); educational differences across subgroups were not statistically significant (p = 0.29).Almost half of those with AD dementia had prevalent dementia at ADNI enrollment, and the other half developed incident dementia.Proportions of prevalent versus incident dementia were similar across subgroups (p = 0.29).The apolipoprotein E (APOE) ε4 allele was more common among people with AD dementia than CN but did not vary significantly across subgroups (p = 0.85).
shows β coefficient findings for AD-No Domains.Colors indicate voxels where the AD-No Domain group had lower GMV than CN controls.Medial temporal lobe and symmetrical bilateral temporal cortex involvement is apparent.Figure S3 shows p value results.Findings were similar for AD-Memory, AD-Visuospatial, AD-Executive, and AD-Multiple Domains subgroups; β coefficient and p value findings are in Figures S4 to S11 .

Figure
Figure S16 shows enlargements of left temporal cortex, medial temporal lobe, and hippocampus.This figure shows more atrophy in the lateral temporal cortex in AD-Language and more atrophy in the medial temporal lobe and hippocampus in AD-Memory.

Figure
Figure S17 shows comparisons of AD-Memory and AD-Visuospatial.There was lower GMV in the bilateral hippocampi and medial temporal lobes in AD-Memory and lower GMV in bilateral cortical regions for AD-Visuospatial.FigureS18shows the corresponding p values.

F I G U R E 2
Comparison of gray matter volume for AD-Language group and cognitively normal elderly controls.F I G U R E 3 Comparison of gray matter volume for AD-Memory group with AD-No Domain group.
Figures S21 to S25 show prior and current findings side by side.VBM uses proximity-based smoothing proce-dures but is agnostic in terms of anatomical relationships beyond distance.Brain regions that differ most from controls show clear replication.

F I G U R E 5
Bold font indicates observations with p < 0.05.Mean gray matter findings compared to controls for each quartile defined by W-score voxel counts for AD-No Domain, AD-Memory, AD-Visuospatial, AD-Language, and AD-Executive.

Table 1
Demographic and clinical characteristics.
summarizes demographic and clinical characteristics.People with AD dementia were older than CN controls, though there TA B L E 1

Table 1
also summarizes cognitive data.People with AD dementia had lower Mini-Mental State Examination (MMSE), Clinical Dementia Rating (CDR) sum of boxes, ADNI-Memory, ADNI-Executive, ADNI-Language, and ADNI-Visuospatial scores than CN controls (all p < 0.0001).MMSE (p = 0.12) and CDR sum of boxes (p = 0.95) did not differ across subgroups.As expected, mean domain scores were closely matched for people in the AD-No Domain and AD-Multiple Domains groups, while individual domain scores were substantially lower for the index domain (the single domain with relative impairment) for the other groups (all p < 0.0001).