Characterizing Heterogeneity in Neuroimaging, Cognition, Clinical Symptoms, and Genetics Among Patients With Late-Life Depression

Key Points

Question  Is late-life depression (LLD) associated with structural neuroimaging patterns?

Findings  In this case-control study, 2 dimensions best represented neuroanatomical heterogeneity in patients with LLD: one was associated with preserved brain structure and the other demonstrated diffuse structural abnormalities and greater cognitive impairment. One de novo independent genetic variant was significantly associated with dimension 1 but not with dimension 2, and dimension 2 was longitudinally associated with Alzheimer disease and brain aging more than dimension 1.

Meaning  The 2 dimensions representing heterogeneity in LLD in this study may offer the potential for clinical precision in diagnosis and prognosis.

Importance  Late-life depression (LLD) is characterized by considerable heterogeneity in clinical manifestation. Unraveling such heterogeneity might aid in elucidating etiological mechanisms and support precision and individualized medicine.

Objective  To cross-sectionally and longitudinally delineate disease-related heterogeneity in LLD associated with neuroanatomy, cognitive functioning, clinical symptoms, and genetic profiles.

Design, Setting, and Participants  The Imaging-Based Coordinate System for Aging and Neurodegenerative Diseases (iSTAGING) study is an international multicenter consortium investigating brain aging in pooled and harmonized data from 13 studies with more than 35 000 participants, including a subset of individuals with major depressive disorder. Multimodal data from a multicenter sample (N = 996), including neuroimaging, neurocognitive assessments, and genetics, were analyzed in this study. A semisupervised clustering method (heterogeneity through discriminative analysis) was applied to regional gray matter (GM) brain volumes to derive dimensional representations. Data were collected from July 2017 to July 2020 and analyzed from July 2020 to December 2021.

Main Outcomes and Measures  Two dimensions were identified to delineate LLD-associated heterogeneity in voxelwise GM maps, white matter (WM) fractional anisotropy, neurocognitive functioning, clinical phenotype, and genetics.

Results  A total of 501 participants with LLD (mean [SD] age, 67.39 [5.56] years; 332 women) and 495 healthy control individuals (mean [SD] age, 66.53 [5.16] years; 333 women) were included. Patients in dimension 1 demonstrated relatively preserved brain anatomy without WM disruptions relative to healthy control individuals. In contrast, patients in dimension 2 showed widespread brain atrophy and WM integrity disruptions, along with cognitive impairment and higher depression severity. Moreover, 1 de novo independent genetic variant (rs13120336; chromosome: 4, 186387714; minor allele, G) was significantly associated with dimension 1 (odds ratio, 2.35; SE, 0.15; P = 3.14 ×10⁸) but not with dimension 2. The 2 dimensions demonstrated significant single-nucleotide variant–based heritability of 18% to 27% within the general population (N = 12 518 in UK Biobank). In a subset of individuals having longitudinal measurements, those in dimension 2 experienced a more rapid longitudinal change in GM and brain age (Cohen f² = 0.03; P = .02) and were more likely to progress to Alzheimer disease (Cohen f² = 0.03; P = .03) compared with those in dimension 1 (N = 1431 participants and 7224 scans from the Alzheimer’s Disease Neuroimaging Initiative [ADNI], Baltimore Longitudinal Study of Aging [BLSA], and Biomarkers for Older Controls at Risk for Dementia [BIOCARD] data sets).

Conclusions and Relevance  This study characterized heterogeneity in LLD into 2 dimensions with distinct neuroanatomical, cognitive, clinical, and genetic profiles. This dimensional approach provides a potential mechanism for investigating the heterogeneity of LLD and the relevance of the latent dimensions to possible disease mechanisms, clinical outcomes, and responses to interventions.

Major depressive disorder is one of the most common mental health disorders and is a leading contributor to disability worldwide.^1,2 Late-life depression (LLD) refers to major depressive disorder that is present in individuals 60 to 65 years and older and can be early onset or late onset. LLD affects 1.8% to 7.2% of older adults in the general community.^3,4

There is considerable heterogeneity in clinical presentation and illness progression in LLD.^5,6 Pharmacological and psychological treatments tend to be less effective in LLD than in other adult age groups. Up to 50% of patients with LLD do not achieve remission with their first treatment.⁷ LLD is associated with cognitive impairment^5,6 and high rates of comorbidity, including cardiac and cerebrovascular disease⁸ and stroke,⁹ as well as increased risk of obesity, diabetes, frailty,¹⁰ and neurodegenerative diseases, such as Alzheimer disease and vascular dementia.^11-14

Magnetic resonance imaging has revealed gray matter (GM) reductions in bilateral anterior cingulate and medial frontal cortices, insula, putamen, and globus pallidus, extending into the parahippocampal gyrus, amygdala, and hippocampus. In contrast, larger GM volumes have been observed in the lingual gyrus,¹⁵ putamen, and caudate regions.¹⁶ Diffusion tensor imaging demonstrates widespread losses in white matter (WM) integrity, including in the anterior thalamic radiation, cingulum, corticospinal tract, superior and inferior longitudinal fasciculi, and uncinate fasciculus.¹⁷ Collectively, the findings support biological models of LLD being associated with cortical atrophy and WM abnormalities in specific brain networks, although the extent and magnitude vary.

Methodological advancements in data-driven biological subtypes^18-23 are challenging the traditional definition of neurological diseases, such as Alzheimer disease^18,19,21,22 and depressive disorder.^24-26 One of the advantages of semisupervised clustering methods^22,23,27,28 is that they perform subtyping via 1-to-k mapping from the domain of a reference group (ie, healthy control individuals) to the patient group, thereby avoiding clustering patients according to disease-irrelevant confounds. Distinct neuropathological mechanisms may underlie heterogeneity in the presentation and progression of the clinical phenotype.²⁹ Furthermore, the extent to which genetic heterogeneity influences or interacts with phenotypic expression has barely been explored,³⁰ and individual-level variability, including environment, genetic or other factors, may lead to different levels of disease liability.³¹

We sought to delineate heterogeneity in patients with LLD in a large multicenter sample (N = 996) using a semisupervised clustering method (heterogeneity through discriminative analysis [HYDRA]).²⁷ We hypothesized that multiple distinct dimensions can describe the underlying heterogeneity and that these dimensions might be prominent in the general population and longitudinal trajectories.

The Imaging-Based Coordinate System for Aging and Neurodegenerative Diseases (iSTAGING) is an international consortium consisting of various imaging protocols, scanners, data modalities, and pathologies,³² comprising harmonized magnetic resonance imaging data in more than 35 000 participants from more than 13 studies and encompassing a wide range of ages (22 to 90 years). The present study includes patients with LLD from 4 cohorts, including UK Biobank (UKBB),^33,36 Psychotherapy Response Study at the University of California San Francisco (UCSF), Baltimore Longitudinal Study of Aging (BLSA),^34,35 and Biomarkers of Cognitive Decline Among Normal Individuals (BIOCARD). The institutional review board at each site approved the study. All participants provided written informed consent to the studies contributing to this pooled mega-analysis.

We applied a harmonized LLD definition criterion to consolidate participants with LLD and excluded those with comorbid medical and neurological diseases that were potential confounders as follows: all participants from all 4 sites were restricted to be 60 years or older; for UKBB, we excluded individuals who were diagnosed with schizophrenia, bipolar disorder, psychotic symptoms, anxiety, obsessive-compulsive disorder, posttraumatic stress disorder, Huntington disease, Alzheimer disease, epilepsy and stroke, diabetes, or hypertension; for BLSA, we excluded individuals diagnosed with hypertension, anxiety, bipolar disorder, or schizophrenia; for BIOCARD, we excluded individuals diagnosed with diabetes or hypertension; and for UCSF, we excluded individuals with substance misuse, psychotic features, cognitive-enhancing substance use, neurological diseases, or posttraumatic stress disorders (Table). We defined 2 additional populations: a general population (12 518 participants from UKBB) and a longitudinal population (1431 participants from ADNI, BLSA, and BIOCARD). A total of 996 participants (501 patients with LLD and 495 healthy control individuals) were included. Image protocols and acquisition parameters for all sites are presented in the eMethods in Supplement 1. The search terms for this study were late-life depression, heterogeneity, semisupervised clustering, and dimensional representation. Data were collected from July 2017 to July 2020 and analyzed from July 2020 to December 2021.

Quality-controlled images were corrected for magnetic field intensity inhomogeneity³⁷ (eMethods in Supplement 1). A multiatlas parcellation method (multiatlas region segmentation utilizing ensembles of registration algorithms and parameters, and locally optimal atlas selection [MUSE])³⁸ was used to extract region-of-interest values of the segmented GM tissue maps (eTable 2 in Supplement 1). Voxelwise regional volumetric maps for each tissue volume³⁹ were generated by spatially aligning skull-stripped images to a template residing in the Montreal Neurological Institute (MNI) space using a registration method.⁴⁰ Fractional anisotropy maps were used to examine microstructural integrity disruptions in the WM (eMethods in Supplement 1). The mean fractional anisotropy values were extracted within the 48 WM tracts of the Johns Hopkins University international consortium for brain mapping diffusion tensor imaging-81 WM label atlas.⁴¹ Intersite image harmonization of the GM MUSE regions of interest is detailed in the eMethods in Supplement 1.

We consolidated an imaging-genetic data set from UKBB that passed the quality-control protocol, resulting in 20 438 participants and 8 430 655 single-nucleotide variants (eMethods in Supplement 1). We then selected 774 UKBB participants who overlapped with the LLD population for genetic analyses.

We applied HYDRA²⁷ to the harmonized MUSE regions of interest (eMethods in Supplement 1). Briefly, HYDRA aims to cluster disease effects instead of directly clustering participants by comparing the patterns between healthy control individuals and patients with LDD, thus resulting in a 1-to-k mapping from the control domain to the patient domain.

We chose the optimal number of dimensions or clusters (k), ranging from 2 to 8 clusters, using the adjusted Rand index.⁴² We performed additional analyses to evaluate the robustness of the optimal k clusters scheme. First, split-sample analyses⁴³ were carried out to assess whether the dimensions in each half exhibited similar neuroanatomical patterns, given that the 2 halves had similar cohort characteristics in terms of age, sex, and site. Second, we conducted leave-site-out validation⁴⁴ to examine whether the dimensions were consistent across sites, training on UKBB only and training on all sites. Lastly, a permutation test was performed to test the statistical significance with the optimal k cluster scheme (eMethods in Supplement 1). Statistical significance was set at .05 for a 2-tailed test.

We subsequently investigated the characteristics of the multiple dimensions regarding GM volume, WM integrity, cognitive functioning and depression-related variables, and genetic architecture. Moreover, we investigated the expression of the k dimensions in the general population and longitudinal data.

Voxelwise regional volumetric GM maps from all sites were used to assess differences in GM tissue volumes. The 3dttest++ program⁴⁵ in Analysis of Functional Neuroimages software version 17.2.10 (Medical College of Wisconsin and National Institutes of Health)⁴⁶ was used to detect the distinct neuroanatomical patterns of the corresponding dimensions vs the control group, considering age, sex, site, and intracranial volume as covariates. For those voxels that survived the adjustment (Benjamini-Hochberg procedure), voxelwise effect-size maps (ie, Cohen f²) were estimated for each paired comparison.

WM microstructural abnormality was assessed using the mean fractional anisotropy values of the 48 regional tracts from UKBB. Group comparisons were performed with multiple linear regression models using R version 3.4.0 (R Foundation) (eMethods in Supplement 1). Age and sex were fixed effects, and group was the variable of interest. P values were corrected, and Cohen f² was computed with the same procedure as above.

Group comparisons for demographic, cognitive, and clinical variables (eTable 5 in Supplement 1) were examined separately between the dimensions. Mann-Whitney U (Wilcoxon) test was used for continuous variables (eg, age) and χ² test of independence for categorical variables (eg, sex). Global effect size (ie, Cohen d) was also reported for continuous variables.

We performed genome-wide associations with the derived binary dimension traits (ie, dimension 1 or dimension 2) vs control using PLINK version 2.0 (Massachusetts General Hospital). The Functional Mapping and Annotation (FUMA) online platform version 1.3.6b (Center for Neurogenomics and Cognitive Research)³ was then used to annotate the genomic risk loci and independent significant single-nucleotide variants (eMethods in Supplement 1).

The trained model was applied to the external validation samples in the general population (Table). Dimension membership and expression scores of the k dimensions were derived (eMethods in Supplement 1).

We examined neuroanatomical patterns using voxelwise regional volumetric GM maps. In addition, demographic and cognitive functioning of the k dimensions was investigated in the general population. We calculated the genome-wide single-nucleotide variant–based heritability coefficient (h²) using Genome-wide Complex Trait Analysis software version 1.93.2β (School of Life Sciences)⁴ (eMethods in Supplement 1).

The cross-sectionally trained model was applied to the longitudinal population (Table). Dimension membership was derived to evaluate its longitudinal changes in MUSE GM regions of interest using the spatial patterns of atrophy for recognition of Alzheimer disease index⁴⁷ and the spatial patterns of atrophy for recognition of brain age index.⁴⁸ Specifically, the rate of change over time in these variables for each participant was derived with a linear mixed-effects model and compared across dimensions using a linear regression model (eMethods in Supplement 1).

The highest adjusted Rand index (0.58) was achieved by a HYDRA model for k = 2 clusters (eFigure 1 in Supplement 1). The cluster assignment distribution for k = 2 to 8 across sites is presented in eTable 3 in Supplement 1. For the optimal k = 2 clustering scheme, 227 participants with LLD were assigned to dimension 1 and 274 to dimension 2. The optimal k = 2 clustering scheme was replicated in split-sample and leave-site-out analyses (eFigure 1 in Supplement 1). In the leave-site-out analyses, the percentage overlap for participants assigned to the same dimension was 446 of 501 (89.02%) (369 of 402 [92%] for UKBB, 22 of 29 [76%] for BLSA, 4 of 5 [80%] for BIOCARD, and 51 of 65 [78%] for UCSF). The neuroanatomical patterns of the 2 dimensions were similar (eFigure 3 in Supplement 1) to the original dimension patterns (Figure 1). In split-sample analyses, the GM patterns for the 2 splits were similar (eFigure 2 in Supplement 1) compared with the original dimension patterns (Figure 1). The adjusted Rand index at k = 2 was higher than the null distribution in the permutation test (Cohen d, 0.31; 95% CI, 0.13-0.49; P < .001). Lastly, we presented the results without excluding comorbidities in UKBB, which yielded similar imaging patterns for the 2 dimensions (eFigure 4 in Supplement 1). Therefore, we present the results of k = 2 for all subsequent analyses.

Patients in dimension 1 demonstrated greater GM tissue volume in bilateral thalamus, putamen, and caudate relative to healthy control individuals. Those in dimension 2 demonstrated reduced GM tissue volume in widespread cortical regions, including bilateral anterior and posterior cingulate gyri, superior, middle, and inferior frontal gyri, gyrus recti, insular cortices, superior, middle, and inferior temporal gyri compared with control individuals (Figure 1). The split-sample and leave-site-out analyses in support are detailed in eFigures 2 and 3 in Supplement 1, respectively.

Patients in dimension 1 exhibited similar fractional anisotropy values compared with control individuals. However, those in dimension 2 showed widespread WM disruptions, with 31 of the 48 WM tracts demonstrating significantly lower fractional anisotropy values than in control individuals but small effect sizes (0.01 ≤ Cohen f² ≤ 0.05, Figure 1B). Specifically, the middle cerebellar peduncle tract obtained the highest effect size (Cohen f² = 0.05). Other affected WM tracts mainly involved the frontal lobe and subcortical limbic regions (eTable 4 in Supplement 1).

Patients in dimension 1 showed statistically higher scores in fluid intelligence scores (Cohen d, 0.25; 95% CI, 0.09-0.41), fewer errors in pairs matching test (Cohen d, −0.28; 95% CI, −0.41 to −0.15), and fewer depressive symptoms in patient health questionnaire responses (Cohen d, −0.45; 95% CI, −0.95 to 0.05) relative to dimension 2. The 2 dimensions did not significantly differ in age, sex, site, or other clinical variables (details in eTable 5 in Supplement 1).

Dimension 1, but not dimension 2, was significantly associated with 1 de novo independent variant (rs13120336; chromosome: 4, 186387714; minor allele, G)⁵ (odds ratio, 2.35; SE, 0.15; P = 3.14 ×10⁸) (Figure 2). Quantile-quantile plots are presented in eFigure 5 in Supplement 1.

Applying the trained model to UKBB samples resulted in 2269 participants in dimension 1, 3786 in dimension 2, 2963 in whom both dimensions were expressed, and 3500 in whom neither dimension was expressed (eTable 6 in Supplement 1; Figure 3).

The neuroanatomical patterns of the 2 dimensions were stable (Figure 3). Participants in dimension 1 showed higher scores in fluid intelligence scores (Cohen d, 0.28; 95% CI, 0.22-0.34; P < .001), but lower errors in pairs matching (Cohen d, −0.13; 95% CI, −0.18 to −0.08; P < .001) compared with those in dimension 2 (eTable 6 in Supplement 1). The expression scores of the 2 dimensions were significantly heritable in the general population. Specifically, h² was 0.27 (SE, 0.04; P < .001) for dimension 1 and 0.18 (SE, 0.04; P < .001) for dimension 2.

Applying the trained model to ADNI, BLSA, and BIOCARD, which also had longitudinal follow-up data, yielded 301 participants in dimension 1, 390 in dimension 2, 330 in whom both dimensions were expressed, and 410 in whom neither dimension was expressed in baseline images (eTable 7 in Supplement 1). The neuroanatomical patterns of the 2 dimensions were stable (eFigure 6 in Supplement 1). The GM rate of change in participants in dimension 2 decreased more rapidly than it did in those in dimension 1 or neither dimension (−0.1 < Cohen f² < 0.1), specifically in the left precentral gyrus, temporal pole, and right anterior insula (eFigure 6 in Supplement 1). Moreover, the 2 dimensions remained independent and stable along longitudinal trajectories (Figure 4). Patients in dimension 2 showed progression of both spatial patterns of atrophy for recognition of Alzheimer disease (Cohen f² = 0.03) and spatial patterns of atrophy for recognition of brain atrophy (Cohen f² = 0.03) compared with those in dimension 1 (Figure 4B), but not at baseline.

Two reproducible and distinct dimensions characterized neuroanatomical heterogeneity in patients with LLD. Patients in dimension 1 showed relatively preserved brain anatomy with larger subcortical regional volumes and were more likely to have with 1 de novo genetic variant, while those in dimension 2 displayed widespread brain atrophy and WM integrity disruptions with impaired cognitive functioning and increased depressive severity. Moreover, the 2 dimensions were manifested in the general population and were significantly heritable. Notably, patients in dimension 2 demonstrated a higher degree of progression to AD and brain aging signatures relative to dimension 1.

The 2 dimensions demonstrated the extent of underlying GM heterogeneity in patients with LLD. GM atrophy evident in dimension 2 has been widely reported in previous case-control studies.^49-51 Regional atrophy in the frontal lobes has been observed,^52,53 which is associated with cognitive deficits as well as reports of psychotic symptoms.⁵⁴ Striatal atrophy has been associated with degeneration in the dopaminergic connections between caudate and cortical limbic areas involved in mood regulation,⁵⁵ although increased caudate and putamen volumes have been found in UKBB depression phenotypes.¹⁶ Dimension 2 showed atrophy in hippocampal regions, perhaps indicative of future neuroprogressive degeneration associated with Alzheimer disease.

The 2 identified neuroanatomical dimensions differed significantly in microstructural integrity. Dimension 1 showed no significant WM abnormalities, while dimension 2 demonstrated widespread WM abnormalities. WM lesions may play a key role in conferring vulnerability or perpetuating depressive syndromes in patients with LLD and contributing to the observed microstructural disturbance.⁵⁶ Widespread WM disruptions can persist in patients with LLD, even excluding WM lesions from the diffusion tensor imaging analysis.⁵⁷ WM tracts connecting frontosubcortical and frontolimbic regions are most frequently affected, including the uncinate fasciculus,^58,59 anterior thalamic radiation, superior longitudinal fasciculus,^56,58,60 and posterior cingulate cortex.⁶¹ Dimension 2 demonstrated clinical features of patients with LLD that are frequently associated with more severe cognitive deterioration.^62-64 Interestingly, previous studies using depressive symptom and cognitive scores,²⁵ or metabolic-inflammatory profile,²⁶ derived 1 subtype that was a healthy group and other subgroups that demonstrated higher depressive symptom scores or a more specific immune-inflammatory dysregulation profile.

The detected genetic variant (rs13120336) was uniquely associated with dimension 1. Two mapped genes (CCDC110 and LOC105377590) have been previously associated with cancer and diabetes.^65,66 We speculate that these genetic factors may play a key role in the heterogeneity of imaging phenotype and cognitive dysfunctions in the 2 dimensions. Many studies have shown that depression is associated with different genetic variants, some of which were not replicated.^67-70 Replication needs to be performed to confirm this detected variant. In general, our dimensional approach might provide another approach for genetic associations in depression.

The 2 dimensions showed significant genetic heritability of 18% to 27%, potentially suggesting genetic underpinnings of neuroanatomical phenotypes associated with depression in the general population. Multimorbidity, such as schizophrenia and anxiety disorders, in the UKBB population⁷¹ might partially account for the expression of the 2 dimensions. Major depressive disorder is a common and complex syndrome with an estimated genetic heritability of approximately 40%,⁷² and prevalence rates range from 7% to 13%.⁷⁰ Our findings confirm the high risks and prevalence of depression in the general population.

The 2-dimensional representation proposed in this study emphasizes the prognostic potential to distinguish LLD that cooccurs with or precedes neurodegenerative diseases. Patients in dimension 2 progressed toward Alzheimer disease or brain aging, whereas those in dimension 1 expressed a preserved brain anatomy. Epidemiological studies^73,74 have consistently found that shared risk factors exist in Alzheimer disease and LLD, supporting depression as a prodromal feature or a risk factor associated with Alzheimer disease. The 2 dimensions did not longitudinally differ in cognitive impairment, perhaps supporting the Alzheimer disease pathological cascade model.⁷⁵

To ensure the reproducibility of the findings, we performed split-sample analysis, leave-site-out analysis, and applying the model trained on LLD to independent UKBB and a combined ADNI, BLSA, and BIOCARD cohort with the same age range as the LLD population. From a technical perspective, applying the trained LLD model to a younger population would be possible, but this could lead to a trivial solution owing to the significant difference in age ranges rather than a disease effect of interest, as aging might play a crucial role in driving these dimensions. We believe that applying the model to external data requires careful consideration of potential confounds, such as demographic differences.

This study has limitations. First, longitudinal data in LLD are needed to confirm the added value of the proposed multidimensional representation. Additionally, replication of the genome-wide association findings is required when additional data are available.

In this study, LLD was characterized into 2 dimensions associated with neuroanatomy, cognitive functioning, and genetic profiles. The 2-dimensional representation offers a potential system for future research on the underlying etiology mechanisms and heterogeneity of genetic architectures as well as personalized clinical care.

Accepted for Publication: December 19, 2021.

Published Online: March 9, 2022. doi:10.1001/jamapsychiatry.2022.0020

Corresponding Author: Junhao Wen, PhD, Center for Biomedical Image Computing and Analytics, Perelman School of Medicine, University of Pennsylvania, 3700 Hamilton Walk, Philadelphia, PA 19104 (junhao.wen89@gmail.com); Christos Davatzikos, PhD, Center for Biomedical Image Computing and Analytics, Perelman School of Medicine, University of Pennsylvania, 3700 Hamilton Walk, Philadelphia, PA 19104 (christos.davatzikos@pennmedicine.upenn.edu).

Author Contributions: Dr Wen and Dr Davatzikos had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Study concept and design: Davatzikos, Wen, Fu

Acquisition, analysis, or interpretation of data: Davatzikos, Wen, Fu, Tosun, Resnick

Drafting of the manuscript: Wen, Fu

Critical revision of the manuscript for important intellectual content: Wen, Fu, Tosun, Veturi, Yang, Mamourian, Srinivasan, Skampardoni, Singh, Nawani, Bao, Erus, Shou, Abdulkadir, Habes, Doshi, Varol, Mackin, Sotiras, Fan, Sheline, Saykin, Shen, Ritchie, Wolk, Albert, Resnick, Davatzikos

Statistical and genetic analysis: Wen

Obtained funding: Davatzikos

Administrative, technical, or material support: Davatzikos

Supervision: Davatzikos, Fu

Conflict of Interest Disclosures: Dr Mackin reported grants from the National Institutes of Health during the conduct of the study and outside the submitted work and support from Johnson and Johnson outside the submitted work. Dr Sotiras reported support from TheraPanacea Equity outside the submitted work. Dr Saykin reported grants from the National Institutes of Health paid to Indiana University during the conduct of the study. Dr Saykin receives support from multiple National Institutes of Health grants, has received support from Avid Radiopharmaceuticals, a subsidiary of Eli Lilly (in kind contribution of PET tracer precursor); Bayer Oncology (scientific advisory board); Siemens Medical Solutions (dementia advisory board); Springer-Nature Publishing (editorial office support as Editor-in-Chief, Brain Imaging and Behavior). Dr Wolk served as site principle investigator for studies by Biogen, Merck, and Eli Lilly/Avid; has received consulting fees from General Electric Healthcare and Neuronix; is on the DSMB for a trial sponsored by Functional Neuromodulation; and has consulted for Qynapse. Dr Albert reported grants from the National Institute on Aging during the conduct of the study and other support from Eli Lilly Consultant outside the submitted work. No other disclosures were reported.

Funding/Support: This work was supported in part by National Institutes of Health grants 1RF1-AG054409-01 and U01-AG068057. Alzheimer’s Disease Neuroimaging Initiative (ADNI) is supported by National Institutes of Health grants U01-AG024904 and RC2-AG036535. The Biomarkers for Older Controls at Risk for Dementia (BIOCARD) study is supported by National Institutes of Health grant U19-AG033655. The Baltimore Longitudinal Study of Aging (BLSA) is supported by the Intramural Research Program of the National Institute on Aging and research and development contract HHSN-260-2004-00012C. This research has been conducted using the UK Biobank Resource under application number 35148.

Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Group Information: The members of ADNI can be found in Supplement 2, members of BIOCARD in Supplement 3, members of BLSA in Supplement 4, and members of the iSTAGING consortium in Supplement 5.