Longitudinal atrophy in early Braak regions in preclinical Alzheimer's disease

Abstract A major focus of Alzheimer's disease (AD) research has been finding sensitive outcome measures to disease progression in preclinical AD, as intervention studies begin to target this population. We hypothesize that tailored measures of longitudinal change of the medial temporal lobe (MTL) subregions (the sites of earliest cortical tangle pathology) are more sensitive to disease progression in preclinical AD compared to standard cognitive and plasma NfL measures. Longitudinal T1‐weighted MRI of 337 participants were included, divided into amyloid‐β negative (Aβ−) controls, cerebral spinal fluid p‐tau positive (T+) and negative (T−) preclinical AD (Aβ+ controls), and early prodromal AD. Anterior/posterior hippocampus, entorhinal cortex, Brodmann areas (BA) 35 and 36, and parahippocampal cortex were segmented in baseline MRI using a novel pipeline. Unbiased change rates of subregions were estimated using MRI scans within a 2‐year‐follow‐up period. Experimental results showed that longitudinal atrophy rates of all MTL subregions were significantly higher for T+ preclinical AD and early prodromal AD than controls, but not for T− preclinical AD. Posterior hippocampus and BA35 demonstrated the largest group differences among hippocampus and MTL cortex respectively. None of the cross‐sectional MTL measures, longitudinal cognitive measures (PACC, ADAS‐Cog) and cross‐sectional or longitudinal plasma NfL reached significance in preclinical AD. In conclusion, longitudinal atrophy measurements reflect active neurodegeneration and thus are more directly linked to active disease progression than cross‐sectional measurements. Moreover, accelerated atrophy in preclinical AD seems to occur only in the presence of concomitant tau pathology. The proposed longitudinal measurements may serve as efficient outcome measures in clinical trials.

thus are more directly linked to active disease progression than cross-sectional measurements. Moreover, accelerated atrophy in preclinical AD seems to occur only in the presence of concomitant tau pathology. The proposed longitudinal measurements may serve as efficient outcome measures in clinical trials. Structural MRI is a promising candidate because, compared to PET, it is more accessible, less expensive and has higher resolution and lower repeat measurement error. Longitudinal change in the medial temporal lobe (MTL), the earliest region affected by neurofibrillary tangle (NFT) pathology, quantified from structural MRI has repeatedly demonstrated sensitivity for early diagnosis and monitoring of patients with mild cognitive impairment (MCI; Chincarini et al., 2016;Iglesias et al., 2016;Jack et al., 2000;Kulason et al., 2019;Ledig, Schuh, Guerrero, Heckemann, & Rueckert, 2018;Leung et al., 2010;Morra et al., 2009;Schuff et al., 2009;Tward et al., 2017;Wolz et al., 2010).
However, there is limited evidence that the more subtle longitudinal change in preclinical AD can be reliably detected with longitudinal structural MRI.
Four recent prospective papers have evaluated the sensitivity of longitudinal MRI to preclinical AD. Donohue et al. (2017) reported a significant difference in hippocampal volume change between amyloid-β positive (Aβ+) and negative (Aβ−) cognitively normal individuals. However, this effect reached significance only after 4 years and was weaker than that of the preclinical Alzheimer's cognitive composite (PACC) score, a standard cognitive measure used in preclinical AD. Conversely, Pegueroles et al. (2017) reported significant atrophy rate difference in the MTL using only scans within 2-year follow-up in a relatively small sample of preclinical AD patients compared to controls. Similarly in a small sample of preclinical AD with evidence of tau pathology, Holland, McEvoy, Desikan, and Dale (2012) found significant differences in longitudinal atrophy rates compared to Aβ− controls over a 3-year period. However, both these two studies did not report how their measurements compared with cognitive scores and replication in larger cohorts needs to be done. A significant interaction between cerebrospinal fluid (CSF) Aβ, phospho-tau biomarkers and cross-sectional cortical thickness in preclinical AD was observed by Fortea et al. (2014), but they did not report a significant thickness difference between Aβ+ and Aβ− groups. Additionally, several retrospective studies have reported longitudinal and crosssectional structural changes in cognitively normal individuals who progressed to cognitive impairment (Miller et al., 2013;Roe et al., 2018;Younes et al., 2019). In particular, Miller et al. (2013) found significantly increased atrophy rates in the hippocampus and entorhinal cortex of cognitively normal individuals who progressed to MCI. This suggests that MTL longitudinal markers are sensitive to asymptomatic disease, but since clinical trials do not have access to information on who will or will not progress to cognitive impairment, it remains critical to evaluate the sensitivity of longitudinal MRI to preclinical AD, as defined by Aβ positivity, in a prospective setting.
Several barriers that may hinder the sensitivity of structural measures in preclinical AD: (a) Most studies have been cross-sectional, which may be suboptimal because they are influenced by nondisease effects, such as inter-subject variability due to developmental and other lifespan factors. Longitudinal measurement reflects active neurodegeneration and thus should be more sensitive to evidence of underlying AD pathology. (b) Most studies have not included Brodmann area 35, which approximates the transentorhinal cortex, the earliest area of NFT pathology (Braak & Braak, 1995). (c) Measurements of the MTL cortical subregions have associated confounds (cortex oversegmentation due to the dura and error in subregion boundaries due to anatomical variability; Figure 1) limiting measurement accuracy. In this study, we hypothesize that a tailored MRI computational pipeline focused on MTL subregions will yield longitudinal biomarkers, measured within a practical timeframe for theoretical clinical trial (2-year followup), that are more sensitive to disease progression in preclinical AD than standard cognitive measures. We also wanted to compare to an emerging blood based biomarker, plasma neurofilament light chain (NfL), which has displayed considerable promise as a potentially easily accessible biomarker of neurodegeneration (Mattsson et al., 2017).

| METHODS
This section provides information on participants, data processing pipeline and statistical analysis. Details on the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, MRI acquisition, quality control, statistical analyses and sample size estimation are available in Supplementary S1.

| Participants
Participants from the ADNI-GO and ADNI-2 studies who had longitudinal T1-weighted (T1w) MRI and Florbetapir PET scans at baseline available were included. To simulate a realistic clinical trial, only longitudinal scans within 2-year follow-up of baseline were analyzed for each participant. Participants who had no scans beyond 1 year after baseline were excluded, since longitudinal atrophy over only 1 year is unlikely to be detected in preclinical AD. A summary standardized uptake value ratio (SUVR) derived from Florbetapir PET 1 was used to determine the Aβ status of each participant (threshold of 1.11 ;Landau et al., 2012). After quality control (Supplementary S1.3), 337 participants (summarized in Table 1) were selected and grouped into Aβ− cognitively normal controls, preclinical AD (Aβ+ controls), early prodromal AD (Aβ+ early MCI, or EMCI). To investigate whether the presence of tau pathology is associated with rate of neurodegeneration in preclinical AD, we further divided this group into tau positive (T+) and negative (T−) subgroups based on CSF p-tau measurement (threshold of 23 pg/mL; Shaw et al., 2009, 68 out of 76 preclinical AD patients have CSF p-tau measurements available).

| Cross-sectional and longitudinal quantitative measures of MTL subregions
The longitudinal MRI scans were processed using a tailored pipeline (summarized in Figure S1) that accounts for common confounds of conventional approaches (described in Figure 1 F I G U R E 1 Common confounds in automatic segmentation of medial temporal lobe (MTL) subregions using T1-weighted MRI. Confound 1: the dura mater (indicated by purple lines) has similar intensity with gray matter (GM) in T1-weighted MRI (a) but can be easily separated in T2-weighted MRI (b), is commonly mislabeled as GM (c). Confound 2: Large anatomical variability exists in the MTL defined by the pattern of the collateral sulcus (CS), which influences the borders and extent of the subregions of the MTL cortex. Our segmentation pipeline is able to reliably separate dura from GM (1d) and account for anatomical variability.  (Xie et al., 2017;. For longitudinal analysis, symmetric diffeomorphic registration (Avants, Epstein, Grossman, & Gee, 2008) was performed between the baseline MRI scan and each of the followup MRI scans. The volume of each MTL subregion in each follow-up scan was estimated by applying the spatial transformation computed by the registration to the ASHS-T1 segmentation of the baseline scan in a manner that is unbiased (Das et al., 2012). For each subregion in each subject, the annualized volume atrophy rate was computed by linear regression of all available longitudinal volume measurements vs. scan date differences from baseline. This was then converted to a relative volume atrophy rate (in %) by dividing by the baseline subregion volume. Bilateral measurements of each subregion were averaged to increase reliability. The conclusions do not change when using measurements of left or right hemisphere separately.

| Cognitive and plasma NfL data processing
To compare our longitudinal measurements with cognitive measurements commonly used in preclinical and prodromal AD and an alternative blood-based neurodegeneration biomarker, we included the PACC (computed as in Donohue et al., 2017) using standardized z score composite of the ADAS-Cog subscale delayed word recall, delayed recall score on logical memory test, MMSE, and the log-transformed trailmaking test B time to completion), Alzheimer's Disease Assessment Scale-Cognitive (ADAS-Cog, ADAS11-Cog was used) and plasma NfL.
All measurements are publicly available from the ADNI database. Annualized longitudinal rate of change was computed using linear regression in the same manner as in Section 2.2. Since we found that normalizing by baseline measures did not improve discriminability, we used the absolute change rate of these measurements.

| Statistical analysis
The longitudinal and cross-sectional measurements (neuroimaging, cognitive and plasma NfL) of each patient group were compared to Aβ− controls separately using general linear models with each measurement as the dependent variable, group membership as the factor of interest, and age as covariate. Intracranial volume was included as an additional covariate for cross-sectional volume measurements. Sex was not included as a covariate in this study, but including it did not significantly change the reported findings. Holm-Bonferroni correction for multiple comparisons was performed (Holm, 1979). In addition, we estimated the sample size (details in Supplementary S1.4) required to detect both 50%/year and 25%/year reduction in the atrophy rate of each patient group relative to that of Aβ− controls (power 1 − β = 0.8, one-sided significance level α = 0.05). The 95% confidence interval of each sample size estimate was computed using the bootstrap method (Efron, 1979). Since we have a strong hypothesis on the direction of the effect, the above analyses were one-sided.

| RESULTS
The results of group comparisons using longitudinal and crosssectional measurements are summarized in Figure 2 and described in detail below.
In order to visualize the raw data, spaghetti plots of representative 3.2 | Cross-sectional volume/thickness and plasma NfL differences between patients and Aβ− controls The group comparison results using cross-sectional measurements of MTL subregions and plasma NfL measurements are summarized in Table 3 and Figure 2b. Compared to longitudinal MTL subregional measurements, cross-sectional discrimination is much weaker, with no Holm-Bonferroni-corrected significant differences observed in preclinical AD (including the T+ preclinical AD subgroup). Only BA35 displayed a p-value <.05 before correction in the whole preclinical AD sample compared to controls (F = 3.5, p = .035), which appears to be driven by the T− preclinical AD individuals (F = 3.9, p = .025) rather than the T+ ones (although both groups had more cortical thinning than controls in absolute terms). In the early prodromal stage, signifi-

| Sample size estimations
The sample size required to detect both 50%/year and 25%/year reduction in atrophy rate is reported in

| Difference in MTL subregion atrophy in preclinical AD
From the longitudinal analysis, we observed that the atrophy rate of BA35 was faster than that of the other subregions in preclinical AD patients ( Figure 3). We performed a post hoc exploratory analysis to investigate whether this effect was statistically significant.

| DISCUSSION
In this study, using a highly tailored MRI processing pipeline that accounts for common confounds (Figure 1)  summarized in a meta-analysis of longitudinal atrophy in normal older adults (Fraser, Shaw, & Cherbuin, 2015), the 0.72%/year in (Holland, McEvoy, Desikan, & Dale, 2012), and the 0.59%/year reported by our prior study using longitudinal T2-weighted MRI (Das et al., 2012).
While the hippocampus has long been the focus of biomarker research, ERC and BA35 are the first cortical sites of NFT pathology (Braak & Braak, 1995) and have been less frequently included in longitudinal analyses. In general, our longitudinal atrophy rates in ERC  Tward et al., 2017: 2.35 and 6.42%/year in controls and MCI), probably due to, again, differences in segmentation protocol, definition of preclinical AD, the use of EMCI rather than a more typical MCI cohort, and the length of follow-up time.
Being able to measure atrophy rates of granular MTL subregions allows for investigation into the spatial pattern of atrophy, which is a unique aspect of our processing pipeline. When comparing preclinical AD to Aβ− controls, we found significant differences in multiple MTL subregions, which is consistent with the results reported by prior studies (Miller et al., 2013;Pegueroles et al., 2017). Across the subregions, BA35 exhibited the greatest volume loss in absolute terms, followed by ERC and hippocampus. BA36 and PHC had the slowest atrophy rate among the MTL subregions in absolute terms. This result fits well with the pattern of spreading of NFT (Braak & Braak, 1995), that is, beginning at the transentorhinal region (approximates the BA35 in our segmentation protocol), spreading to ERC and hippocampus and then to the BA36. Thus, this pattern supports the notion that our longitudinal measurements are sensitive to tau-mediated neurodegeneration, along with this effect found only in the T+ preclinical AD subgroup.
Of note, while BA35 demonstrated the largest atrophy rate of any region in the preclinical AD group (significant compared to ERC, BA36, PHC and anterior hippocampus while only in absolute terms compared to posterior hippocampus, Figure 3) and had a marginally larger absolute difference with Aβ− controls compared to that of the hippocampal measurements, the latter had stronger statistical significance (smaller p-value). This is likely due to the larger variability in measurement of BA35 atrophy compared to that of the hippocampus, likely because BA35 is much harder to measure due to its smaller size and variability of its location (depending on the collateral sulcus pattern shown in Figure 1). This suggests that by improving the accuracy of BA35 measurements through advances in imaging and image analysis technology, we could derive more sensitive biomarkers of preclinical AD in the future. Nonetheless, BA35 was the only region in the cross-sectional analysis to approach significance in differentiating the preclinical group from the controls and its longitudinal change demonstrated the strongest statistical significance in separating early prodromal AD patients from Aβ− controls.
In this study, we did not observe significant group differences in longitudinal change of cognitive measures, including the PACC, which has previously demonstrated sensitivity to the cognitive decline of preclinical AD (Donohue et al., 2017). Indeed, the current finding may appear inconsistent that reported by Donohue et al. (2017), in which they found a significant group effect using the PACC for discrimina-

| Accelerated atrophy in preclinical AD is associated with concomitant tau pathology
In this study, we only observed significant increases in atrophy rates in T+ preclinical AD patients, not T− ones, suggesting that MTL neurodegeneration only occurs in subjects with evidence of concomitant tau pathology. In fact, the magnitude of most of the MTL longitudinal measures in T+ preclinical AD patients is comparable to that of the early prodromal AD (except for BA35 which is larger in the early prodromal AD group). A similar finding was reported in prior crosssectional (Fortea et al., 2014) or longitudinal (Desikan et al., 2011;Holland, McEvoy, Desikan, & Dale, 2012;Pegueroles et al., 2017) studies, in which structural abnormality were only observed in subjects with evidence of both amyloid and tau pathologies. Our finding further supports the notion that, compared to amyloid pathology, tau pathology is more directly linked to neurodegeneration. Additionally, it also demonstrates that our longitudinal measurements are sensitive to tau-mediated neurodegeneration.

| Sample size estimation in MCI and preclinical AD
In this study, we computed sample size relative to controls (as in Cash et al., 2015;Das et al., 2012;Leung et al., 2010) rather than to zero atrophy rate (as in Chincarini et al., 2016;Wolz et al., 2010), because the differential effect between the treatment and placebo groups is of greater interest in clinical trials.
Our results shown in Table 4  Biosciences.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available in Alzheimer's Disease Neuroimage Initiative (ADNI) at adni.loni.usc.edu.