Functional network integrity presages cognitive decline in preclinical Alzheimer disease

Abstract

Objective:

To examine the utility of resting-state functional connectivity MRI (rs-fcMRI) measurements of network integrity as a predictor of future cognitive decline in preclinical Alzheimer disease (AD).

Methods:

A total of 237 clinically normal older adults (aged 63–90 years, Clinical Dementia Rating 0) underwent baseline β-amyloid (Aβ) imaging with Pittsburgh compound B PET and structural and rs-fcMRI. We identified 7 networks for analysis, including 4 cognitive networks (default, salience, dorsal attention, and frontoparietal control) and 3 noncognitive networks (primary visual, extrastriate visual, motor). Using linear and curvilinear mixed models, we used baseline connectivity in these networks to predict longitudinal changes in preclinical Alzheimer cognitive composite (PACC) performance, both alone and interacting with Aβ burden. Median neuropsychological follow-up was 3 years.

Results:

Baseline connectivity in the default, salience, and control networks predicted longitudinal PACC decline, unlike connectivity in the dorsal attention and all noncognitive networks. Default, salience, and control network connectivity was also synergistic with Aβ burden in predicting decline, with combined higher Aβ and lower connectivity predicting the steepest curvilinear decline in PACC performance.

Conclusions:

In clinically normal older adults, lower functional connectivity predicted more rapid decline in PACC scores over time, particularly when coupled with increased Aβ burden. Among examined networks, default, salience, and control networks were the strongest predictors of rate of change in PACC scores, with the inflection point of greatest decline beyond the fourth year of follow-up. These results suggest that rs-fcMRI may be a useful predictor of early, AD-related cognitive decline in clinical research settings.

Though assessment of β-amyloid (Aβ) burden is critical to the conceptualization of preclinical Alzheimer disease (AD), the presence of elevated Aβ alone is associated with variable levels of cognitive decline. It is now widely accepted that supplementing Aβ status with a marker of neurodegeneration or tau burden is likely a critical step in achieving this end.^1,2 In the current study, we focused on the ability of resting-state functional connectivity (rs-fcMRI) to provide evidence of early intrinsic network disconnectivity that may predict future cognitive decline in older adults, either alone or in concert with Aβ.

Cognitive decline has been associated with brain network alterations in rs-fcMRI of the default network in patients with AD dementia.^3–5 Decreases in default network connectivity are also seen with increasing clinical impairment from mild cognitive impairment to clinically evident AD dementia,^3,6,7 and the default network is among the sites of early Aβ deposition in AD.^4,8 Recent studies suggest that other cognitive rs-fcMRI networks, especially the salience, dorsal attention, and frontoparietal control networks, are also targeted in AD.⁹ We examined the hypothesis that disruptions in network integrity measurable by rs-fcMRI predict future cognitive decline in a longitudinal cohort of clinically normal elderly participants. We further hypothesized that the predictive effect of reduced network connectivity would be synergistic with neocortical Aβ burden, with the steepest cognitive decline evident in individuals with higher Aβ burden and reduced rs-fcMRI.

METHODS

Participants.

A total of 237 clinically normal participants from the Harvard Aging Brain Study (HABS) underwent baseline neocortical Aβ imaging with Pittsburgh compound B (PiB) PET imaging and structural and resting-state functional MRI (PiB-MRI interval = 158 ± 97 days; relatively slow amyloid accumulation reported in the literature¹⁰ supports this time interval as acceptable). To be enrolled in the HABS study, participants needed to score 0 on the Clinical Dementia Rating Scale global score, greater than 25 on the Mini-Mental State Examination, and less than 11 on the Geriatric Depression Scale, and to perform within validated education-adjusted norms on Logical Memory II delayed recall. Twenty-six participants (4%) were left-handed. To be included in this study, participants were additionally required to possess at least 2 annual neuropsychological assessments: 9 completed 2 visits, 58 completed 3 visits, 37 completed 4 visits, 75 completed 5 visits, and 58 completed 6 visits. The median neuropsychological follow-up period was 3 years. Baseline demographics can be found in table 1.

Table 1.

Baseline demographic and cognitive measures according to β-amyloid (Aβ) status

Standard protocol approvals, registrations, and patient consents.

Study protocols were approved by the Partners Healthcare Institutional Review Board and all participants provided written informed consent.

MRI.

MRI were collected on a 3T Trio Tim scanner (Siemens Medical Systems, Erlangen, Germany) using 12-channel phased-array head coils. fMRI data were acquired using a gradient-echo echoplanar imaging sequence sensitive to blood oxygen level–dependent contrast. Whole-brain coverage was acquired using repetition time (TR) 3 seconds, flip angle 85°, echo time (TE) 30 ms, matrix 72 × 72, field of view 216 × 216 mm, and 47 × 3 mm axial slices, which resulted in isotropic voxels of 3 mm. Two 124 volumes were acquired in 2 6:12 runs for a total of 12:24. A crosshair was projected on a screen visible to participants during the scan, and participants were asked to lie flat, keep eyes open, and remain awake.

We processed resting-state data using SPM8 (fil.ion.ucl.ac.uk/spm/). We excluded the first 4 volumes of each run for T1 equilibration. We slice-time corrected each run (using fifth order or polynomial splines, with the first slice as the reference slice), realigned to the first volume of each run with INRIAlign, and normalized each run nonlinearly to the Montreal Neurological Institute (MNI) 152 EPI template (using statistical parametric mapping–defined normalization parameters). This was conducted to avoid warping artifact that can occur when normalizing to the T1 image. We smoothed with a 6-mm Gaussian kernel, as this boosts the signal-to-noise ratio and improves intersubject registration. Additional processing included (1) movement correction using the coregistration parameters (plus first derivatives) in order to reduce movement artifacts and (2) temporal bandpass filtering to remove frequencies outside of the 0.01–0.08 Hz band. Further details can be found at Schultz et al.¹¹ In addition, outlier volumes were scrubbed from the connectivity analysis if change in global signal was more than 2.5 SDs from the mean change, if change in position was greater than 0.75 mm from the previous volume, or if change in rotation was greater than 1.5° from the previous volume. Individual runs were dropped if more than 20 volumes in a run were flagged as outliers (n = 2) or if the average temporal signal to noise was less than 115 (n = 1).

Template maps for the default mode, salience, dorsal attention, frontoparietal control, primary visual, extrastriate visual, and motor networks were derived from an out-of-sample population of 675 participants in the Brain Genomics Superscript Project¹² (neuroinformatics.harvard.edu/gsp/). Participants were an average of 21 years old (range 18 and 35 years), and were 58% female, 12% left-handed, with approximately 14 years of education.^11,13 These demographics did not substantially depart from a recent publication of the larger study cohort.¹³ Derivation, reproducibility, and use of these templates to generate connectivity measures have previously been described in detail.^11,14 Maps and relevant code for template-based rotation analysis (TBR) are publicly available at mrtools.mgh.harvard.edu. TBR connectivity measures for each participant and each network were calculated by computing the average correlation strength of all voxels within a particular thresholded network mask (threshold defined as greater than 40% of the maximum value in the corresponding template map), yielding a whole network average connectivity measure. The default network comprised an aggregate of posterior midline, parietal, medial frontal, lateral temporal, and parahippocampal regions,¹¹ similar to previous reports of this network.^15,16 Although some studies have fractionated the default mode into subnetworks^17,18 (particularly anterior and posterior or ventral and dorsal subsystems), our unified default network accords well with prior work using seed-based analyses¹⁹ and group independent component analysis,¹⁸ and is highly similar to the parcellation arrived at via cluster analysis.¹⁵ Notably, we have demonstrated that connectivity measures derived using this unified default network template are significantly related to cognitive performance in the HAB study.¹⁴ The salience network is predominantly composed of dorsal anterior cingulate, prefrontal, frontal, and insula regions, and is similar to that reported by Seeley et al.¹⁶ The dorsal attention network, including the parietal and fusiform regions, closely mirrors that of Yeo et al.¹⁵ MNI seeds for all templates can be found in our methods publication.¹¹

Structural T1-weighted images were acquired as magnetization-prepared rapid gradient echo with the following acquisition parameters: TR/TE/inversion time 2,300/2.95/900 ms, flip angle 9°, 1.1 × 1.1 × 1.2 mm resolution, 2 × (GRAPPA) acceleration. The structural MRI data were processed with Freesurfer v5.1. For the current study, total gray matter (GM) volume was used as a covariate, and was adjusted according to each individual's estimated total intracranial volume (ICV) (GM and white matter measured) using the following algorithm:

where b indicates the regression coefficient when GM is regressed against ICV.¹

PET imaging.

Acquisition of ¹¹C PiB-PET data in HABS has previously been described in detail.²⁰ In brief, PiB-PET images were acquired with an 8.5–15 mCi bolus injection and a subsequent 1-hour dynamic acquisition over 69 volumes (12 × 15 seconds, 57 × 60 seconds). PET data underwent reconstruction and attenuation correction, were evaluated for head motion, and were coregistered to the corresponding T1 image for each participant using a 6 degrees of freedom rigid body registration. Further details on our PET methodology can be found in our previous publication.²¹ A summary distribution volume ratio (DVR) of frontal, lateral, and retrosplenial tracer uptake (FLR),²² as defined by Freesurfer v5.1, was calculated for each participant by averaging the median PiB uptake value across voxels in the precuneus, rostral anterior cingulate, medial orbito-frontal, superior frontal, rostral middle frontal, inferior parietal, inferior temporal, and middle temporal regions of interest from both hemispheres divided by the median PiB DVR from cerebellar GM. The regions comprising the FLR are known to show elevated PiB binding in AD dementia patients.²³ The cutoff of PiB FLR DVR >1.2 for Aβ positivity was derived via a Gaussian mixture model approach detailed in a prior report.¹

Cognitive measures.

The Preclinical Alzheimer Cognitive Composite (PACC) was created using the following neuropsychological tests²⁴: the Logical Memory IIa delayed recall (LM-DR) index from the Wechsler Memory Scale,²⁵ the Mini-Mental State Examination (MMSE) total score,²⁶ the Digit Symbol Substitution Test score from the Wechsler Adult Intelligence Scale–Revised,²⁷ and a combined summation of the Free and Cued Selective Reminding Test (FCSRT) free recall and total recall (free + cued) indices.²⁸ Measures were z score transformed based on the mean (SD) from baseline data and averaged, with a higher score on the PACC indicating better performance. For the current study, each of the measures were also examined as dependent variables to determine which components of the PACC might be most strongly related to functional connectivity (FC) networks.

Analyses.

We used the nlme statistical package in R (version 3.3.0) software to implement linear and nonlinear mixed models. Models were constructed to ascertain the influence of FC networks and Aβ on longitudinal PACC performance or on the PACC components. Both random effects of intercept and slope for each participant were modeled using maximum likelihood estimation, with linear and quadratic terms for time. For each FC network, we examined the same set of models:

1. Cognition ∼ time + covariates × time

2. Cognition ∼ FC_network × time + covariates × time

3. Cognition ∼ FC_network × Aβ × time + covariates × time

4. Cognition ∼ FC_network × Aβ × time + FC_network × Aβ × time² + covariates × time

   • Cognition = PACC (within the PACC: MMSE, LM-DR, Digit Symbol, or FCSRT)

   • Covariates = age, years of education, GM_adjusted

The fit of each increasingly complex model was tested against the fit of the prior model using analysis of variance (ANOVA). A more complex model was rejected in favor of a simpler model if the ANOVA revealed p > 0.05. Differences in the fit indices were reported as the difference in Akaike information criteria (ΔAIC) and the difference in Bayesian information criteria (ΔBIC) between the 2 models. Only significant covariate-by-time interactions were included in final models. Whole network connectivity measurements were regressed against movement and number of outlier volumes and z scored prior to inclusion into the models above. Multiple independent comparisons were accounted for according to a Sidak correction of α = 0.007. For ease of interpretation of the continuous rs-fcMRI and Aβ predictors, our figures represent model estimates of cognitive slopes according to differing levels of connectivity or Aβ burden.

RESULTS

FC networks on PACC change.

Baseline PACC scores were weakly positively associated with baseline default network, r(235) = 0.16, p = 0.01, salience, r(235) = 0.15, p = 0.02, and control network connectivity, r(235) = 0.2, p = 0.001. According to model 2, lower baseline default, salience, and control network functional connectivity at enrollment in the study was associated with PACC decline over time, even after accounting for covariates (age, education, GM volume). By contrast, dorsal attention network connectivity was not a significant predictor of longitudinal PACC performance. Connectivity in the sensory networks, primary visual, motor, and extrastriate visual, also did not predict change in PACC scores over time (see table e-1 at Neurology.org for details). These results suggest that the effect of connectivity on the PACC is limited to a subset of cognitive networks. As default, salience, and control networks showed the clearest relationship with PACC score decline, further analyses were conducted using only these networks. A diagrammatic depiction of the model slopes for each network can be found in figure 1A; forest plot of the slope estimates can be found in figure 2 (with full model estimates and effect sizes in table e-1).

Figure 1. Model estimates of Preclinical Alzheimer Cognitive Composite (PACC) decline according to resting-state functional connectivity networks.

Slopes represent PACC trajectories according to connectivity in all functional networks. Blue line indicates baseline connectivity 1 SD above the group mean (i.e., high connectivity); purple line indicates mean baseline connectivity; red line indicates baseline connectivity 1 SD below the group mean (i.e., low connectivity).

Figure 2. Model estimates of Preclinical Alzheimer Cognitive Composite (PACC) slopes.

Above the horizontal line: estimates of the linear slopes of cognitive decline according to each resting-state functional connectivity network. Below the horizontal line: estimates of the quadratic slopes of cognitive decline according to the interaction between β-amyloid (Aβ) and default mode/salience/frontoparietal networks. Estimate represents the change in PACC (by SD/year) according to a unit increase in network disconnectivity (above horizontal line) or a unit increase in the Aβ × network disconnectivity interaction (below the horizontal line).

Interaction between default, salience, and control network connectivity and Aβ on PACC change.

As the PACC was developed to be a sensitive measure of Aβ-related cognitive decline,²⁴ we hypothesized that interactions between Aβ and default network connectivity would significantly predict PACC decline. In this context, we next modeled longitudinal PACC scores by including 3-way linear and quadratic interactions between default, salience, or control networks and Aβ burden with time. While both linear and quadratic time interaction terms were significant predictors of PACC change (table e-1), inclusion of a quadratic term improved model fit as compared with the linear-only model for all 3 networks examined (default network: model 3 vs 4: ΔAIC = 18.4, ΔBIC = −1.17, log-likelihood ratio = 13.2, p < 0.0001). An illustrative depiction of this 3-way interaction can be found in figure 3.

Figure 3. 3D representation of the rate of change of Preclinical Alzheimer Cognitive Composite (PACC) over time according to default network connectivity and level of β-amyloid (Aβ) burden.

Each bar represents rate of PACC change according to default network connectivity (x axis) and Aβ burden (z axis), with rates of change in PACC slope (on the y axis) represented by each bar in the figure. Darker blue bars indicate greater rates of PACC decline (evidenced in cases of highest Aβ burden and lowest default connectivity), while lighter yellow bars indicate practice effects. PiB DVR = Pittsburgh compound B distribution volume ratio.

Considering the quadratic model, we observed that lower baseline default network connectivity interacted with higher Aβ burden to predict poorer PACC scores over time (see figure 3). The improved model fit came as a result of a quadratic inflection at the third year of follow-up, with poorer default network connectivity and higher Aβ burden relating to rapidly declining PACC scores after this point (refer to figure 4A). By contrast, in individuals with low Aβ burden, no influence of default network connectivity was found on longitudinal PACC scores. Similar quadratic interactions were found between salience network and Aβ (figure 4B), and control network and Aβ (figure 4C), again with a quadratic inflection at the third year of follow-up. Tables e-2 and e-3 present the estimates for the linear-only and full quadratic models, respectively. Figures e-1 and e-2 depict the raw data of PACC score change over time.

Figure 4. Nonlinear model estimates of Preclinical Alzheimer Cognitive Composite (PACC) slopes according to default, salience, or control networks by β-amyloid (Aβ) interaction.

Slopes represent nonlinear trajectories according to different levels of Aβ burden and network connectivity in (A) default, (B) salience, and (C) control networks; low Aβ burden (M_{PiB DVR} = 1.08), high Aβ burden (M_{PiB DVR} = 1.40), low default connectivity (M_{Below median split across networks} = −0.77), high default connectivity (M_{Above median split across networks} = 0.78). Error bars are 95% confidence interval. PiB DVR = Pittsburgh compound B distribution volume ratio.

Default network and Aβ on change in individual measures comprising the PACC.

We next decomposed the PACC into its 4 constituent measures and assessed the longitudinal relationship of each to baseline default network connectivity. Three-way interactions among default network connectivity, Aβ, and time significantly predicted longitudinal FCSRT total, MMSE, and Digit Symbol Substitution scores. Furthermore, models including quadratic time terms fit significantly better than similar linear models. Longitudinal FCSRT free recall performance showed a similar pattern (model 4, default × Aβ × time² term, p = 0.06). The 3-way interaction also significantly predicted longitudinal Logical Memory delayed recall performance, though for this measure the fit of linear and quadratic models was statistically similar (model 3 vs 4: ΔAIC = −0.92, BIC = −20.50, log-likelihood ratio = 7.08, p = 0.13). For Logical Memory, those with low Aβ burden showed a clear practice effect, whereas a lack of evident practice effect was seen in individuals with combined higher Aβ burden and lower default network connectivity (see table e-4 and figure e-2).

DISCUSSION

Our results add to a body of literature demonstrating relationships between Aβ burden and rs-fcMRI measures of network connectivity (particularly with respect to the default,^{4,9,20,29–31} salience, and control networks), and further suggest that connectivity measures across cognitive networks can act synergistically with Aβ burden to predict cognitive decline in preclinical AD. Previous studies indicate that Aβ preferentially aggregates in areas of high intrinsic connectivity,³⁰ especially in the default network.^8,32 In individuals with high Aβ burden, lower cognitive network connectivity may represent evidence of Aβ-related synaptic dysfunction,³³ suggesting these individuals are further along the pathway to AD-related cognitive decline.³⁴ Alternatively, measures of intrinsic connectivity may serve as an Aβ-independent marker of reserve capacity,^5,35 such that individuals with lower levels of connectivity (and, potentially, lower reserve) may be more susceptible to Aβ-related cognitive decline. More broadly, the results motivate further studies focusing on the relationship between connectivity and reserve, including metrics of brain³⁶ and cognitive reserve (i.e., education, premorbid intelligence, occupational attainment, cognitive activities, or some combination of these).³⁷

In addition to the heavily studied default network, we found that altered connectivity in the salience and control networks was predictive of Aβ-related cognitive decline. These relationships were subtly weaker than with the default network, though the overall patterns with respect to linear and quadratic slopes of PACC decline were similar. This result is perhaps unsurprising given that these networks show relationships similar to the default network with respect to performance on executive and episodic memory tests,¹⁴ and are also thought to be susceptible to disruption by early Aβ deposition.^9,38 Further evidence for cognitive network specificity in the prediction of cognitive decline was provided by the null result when examining the influence of visual or motor networks as predictors of longitudinal PACC scores. Notably, previous studies have not found a relationship between connectivity in visual and motor networks and Aβ burden.³¹

It remains uncertain whether disconnectivity of intrinsic networks appears prior to Aβ deposition, as suggested by recent studies,³ or whether the connectivity changes are a consequence of abnormal protein deposition in nodes within targeted networks.^33,39 This question is further complicated by studies suggesting that hyperactivity and hyperconnectivity within affected networks may precede a later stage of disease in which progressive declines in connectivity mirror progressive cognitive decline.²⁹ If present, hyperconnectivity and hyperactivity in targeted networks³⁰ may reflect early Aβ-induced alterations in neural processing, and this increased activity may itself promote further production of toxic Aβ species. Longitudinal rs-fcMRI studies will address these questions, and help elucidate cause-and-effect relationships between Aβ deposition and changes in connectivity.

Generalizability beyond the demographic characteristics of the cohort and the potential for attrition bias are possible limitations of the present study. The HABS cohort has relatively high levels of education and performance on verbal IQ testing, and, accordingly, the levels of cognitive reserve in this cohort may affect the observed levels of connectivity and pattern of cognitive decline. In addition, some participants did not complete all follow-up assessments, and considering that individuals with cognitive impairment are more likely to drop out of longitudinal studies,⁴⁰ it is possible that systematic attrition of this type may lead to an underestimation of baseline rs-fcMRI and Aβ as predictors of cognitive decline.

Our results indicate that early evidence of intrinsic network dysfunction, especially in the setting of co-occurring AD pathology, may presage future cognitive decline. The demonstrable synergy between connectivity and Aβ burden in predicting decline suggests that connectivity measurements may complement imaging biomarkers of AD pathology by helping to distinguish individuals who are more (low connectivity) or less (high connectivity) susceptible to progressive, pathology-induced cognitive decline. Though further work is needed to operationalize connectivity measurements and to establish reliable cutoffs, the results here also suggest that rs-fcMRI could potentially be integrated with other available biomarkers to improve the identification of individuals at the highest risk of impending cognitive decline.

GLOSSARY

Aβ

β-amyloid

AD

Alzheimer disease

AIC

Akaike information criteria

ANOVA

analysis of variance

BIC

Bayesian information criteria

DVR

distribution volume ratio

FC

functional connectivity

FCSRT

Free and Cued Selective Reminding Test

FLR

frontal, lateral, and retrosplenial tracer uptake

GM

gray matter

HABS

Harvard Aging Brain Study

ICV

intracranial volume

LM-DR

Logical Memory IIa delayed recall

MMSE

Mini-Mental State Examination

MNI

Montreal Neurological Institute

PACC

Preclinical Alzheimer Cognitive Composite

PiB

Pittsburgh compound B

rs-fcMRI

resting-state functional connectivity

TBR

template-based rotation analysis

TE

echo time

TR

repetition time

AUTHOR CONTRIBUTIONS

Rachel Buckley, Aaron Schultz, Jasmeer Chhatwal: study concept and design. All authors: acquisition of data. Rachel Buckley, Aaron Schultz, Reisa Sperling, Jasmeer Chhatwal: analysis and interpretation of data. Reisa Sperling, Keith Johnson: study supervision. All authors: critical revision of manuscript for intellectual content.

STUDY FUNDING

Dr. Buckley is funded by the NHMRC Dementia Research Fellowship (APP1105576). This work was supported with funding from the NIH, including P01 AG036694 (R.A.S., K.A.J.), P50 AG005134 (R.A.S., K.A.J., T.H.), K23 EB019023 (J.S.), K23 AG049087 (J.P.C.), K24 AG035007 (R.A.S.), and K01 040197 (T.H.). This research was carried out in part at the Athinoula A. Martinos Center for Biomedical Imaging at the Massachusetts General Hospital, using resources provided by the Center for Functional Neuroimaging Technologies, P41EB015896, a P41 Biotechnology Resource Grant supported by the National Institute of Biomedical Imaging and Bioengineering, NIH. This work also involved the use of instrumentation supported by the NIH Shared Instrumentation Grant Program and/or High-End Instrumentation Grant Program; specifically, grants S10RR021110, S10RR023401, and S10RR023043.

DISCLOSURE

R. Buckley is funded by the NHMRC Dementia Research Fellowship (APP1105576). A. Schultz has been a paid consultant for Janssen Pharmaceuticals and Biogen. T. Hedden is funded by NIH (K01 040197). K. Papp has been a paid consultant for Biogen. B. Hanseeuw is funded by the Belgian American Education Foundation (BAEF), Belgian National Fund for Scientific Research (FNRS), and Saint-Luc Foundation. G. Marshall has received research salary support from Eisai Inc. and Eli Lilly and Company. J. Sepulcre is funded by NIH (K23 EB019023). E. Smith reports no disclosures relevant to the manuscript. D. Rentz served as a consultant for Eli Lilly, Biogen Idec, and Lundbeck Pharmaceuticals, and serves as a member of the Scientific Advisory Board for Neurotrack. K. Johnson has served as a paid consultant for Bayer, GE Healthcare, Janssen Alzheimer's Immunotherapy, Siemens Medical Solutions, Genzyme, Novartis, Biogen, Roche, ISIS Pharma, AZTherapy, GEHC, Lundberg, and Abbvie. He is a site coinvestigator for Lilly/Avid, Pfizer, Janssen Immunotherapy, and Navidea. He has spoken at symposia sponsored by Janssen Alzheimer's Immunotherapy and Pfizer. K. Johnson receives funding from NIH grants R01EB014894, R21 AG038994, R01 AG026484, R01 AG034556, P50 AG00513421, U19 AG10483, P01 AG036694, R13 AG042201174210, R01 AG027435, and R01 AG037497 and the Alzheimer's Association grant ZEN-10-174210. R. Sperling has served as a paid consultant for Abbvie, Biogen, Bracket, Genentech, Lundbeck, Roche, and Sanofi. She has served as a coinvestigator for Avid, Eli Lilly, and Janssen Alzheimer Immunotherapy clinical trials. She has spoken at symposia sponsored by Eli Lilly, Biogen, and Janssen. R. Sperling receives research support from Janssen Pharmaceuticals and Eli Lilly and Co. These relationships are not related to the content in the manuscript. She also receives research support from the following grants: P01 AG036694, U01 AG032438, U01 AG024904, R01 AG037497, R01 AG034556, K24 AG035007, P50 AG005134, U19 AG010483, R01 AG027435, Fidelity Biosciences, Harvard NeuroDiscovery Center, and the Alzheimer's Association. J. Chhatwal is funded by NIH (K23 AG049087). Go to Neurology.org for full disclosures.