Optimal Combinations of Biomarkers to Determine AT(N) in the Alzheimer’s Disease

Background: National Institute on Aging—Alzheimer’s Association (NIA-AA) proposed the AT(N) system based on β-amyloid deposition, pathologic tau, and neurodegeneration, which considered the denition of Alzheimer’s disease (AD) as a biological construct. However, the associations between different AT(N) combinations and clinical stage and progression have been poorly explored systematically. The aim of this study is to compare different AT(N) combinations using recognized biomarkers within the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. Methods: A total of 341 participants from ADNI cohort were classied into AT(N) groups, including 200 cognitively unimpaired (CU) participants and 141 cognitively impaired (CI) participants (101 mild cognitive impairment [MCI] and 40 Alzheimer’s disease [AD]). CSF Aβ42 and amyloid-PET ([18F]utemetamol) were used as biomarkers for A; CSF phosphorylated tau (p-tau) and tau-PET ([18F]ortaucipir) were used as biomarkers for T; CSF total tau (t-tau), FDG-PET, hippocampal volume, temporal cortical thickness and plasma neurolament light (NfL) were used as biomarkers for (N). Binarization of biomarkers was acquired from Youden index and public cutoffs. The relationship between different AT(N) biomarkers combinations and cognitive changes (longitudinal Mini-Mental State Examination scores and Clinical Dementia Rating Sum of Boxes) was examined using linear mixed modeling and coecient of variation. Results: Among CU participants, A−T−(N)− variants were most common. More T+ cases were shown using p-tau than tau PET, and more N+ cases were shown using uid biomarkers than neuroimaging. Among CI participants, A+T+(N)+ was more common. Tau PET combined with cortical thickness best predicted longitudinal cognitive decline in CI and MRI measurements in CU participants. Conclusion: These ndings suggest that optimal combinations of biomarkers to determine AT(N) are differed by clinical stage. Different biomarkers within a specic component 12 . ITC and Braak V/VI indicated early and late stage of tangle pathology respectively. [18F]ortaucipir data were corrected for partial volume effects using the Geometric Transfer Matrix (GTM) approach and divided by the inferior cerebellar GM reference region 13 . The pre-dened meta-ROIs in FDG PET of AD were composed of the angular gyrus, posterior cingulate, and inferior temporal cortical normalized to pons and vermis 14


Background
Alzheimer's disease (AD) is the most common cause of dementia, and one of the main causes of complications and death in the aging population. A series of complex pathobiology are involved in the pathogenesis of AD, including the deposition of extracellular amyloid plaque, tau-related intracellular neuro brillary tangles (NFTs), neuronal loss and atrophy 1 . Recently, National Institute on Aging-Alzheimer's Association (NIA-AA) proposed a research framework based on the pathological characteristics mentioned above 2 . The framework establishes a classi cation system consisted of biomarkers of Aβ (A), tau (T), and neurodegeneration (N), and lists a classic AD biomarker grouping including CSF, MRI and PET. However, it's not perfect concordant among biomarkers within a speci c component (A, T, or N) [2][3] , and it's usually di cult to perform all examinations on patients, which may limit its clinical application. Therefore, how to choose AT(N) biomarkers for patients with different clinical stages is an urgent problem to be solved. Though a lot of researches compared different biomarkers in a certain component [4][5][6] , only one study assessed different combinations of AT(N) using BioFINDER participants 7 . Here, we use a more comprehensive biomarkers group and suppose that AT(N) category prevalence and cognitive prediction would vary by combinations of different biomarkers and clinical stage.

Participants
All participants in this study were from the Alzheimer's Disease Neuroimaging Initiative (ADNI), which is a longitudinal multicenter study designed to develop clinical, imaging, genetic, and biospecimen biomarkers for tracking the progression of AD. Cognitively unimpaired (CU) participants must be free of memory complaints and cognitively normal. And cognitively impaired (CI) participants must have a subjective memory concern, including mild cognitively impaired (MCI) participants, whose general cognition and functional performance su ciently preserved, and AD dementia participants according to NINCDS/ADRDA criteria for probable AD 8 . Demographic and clinical information, neuroimaging, and biomarkers data were downloaded from the ADNI data repository (adni.loni.usc.edu).
CSF and plasma biomarkers analysis CSF β-Amyloid (1-42), phospho-tau (181P), and total tau were analyzed by the electrochemiluminescence immunoassays (ECLIA) Elecsys following a Roche Study Protocol 4 . Plasma neuro lament light (NfL) was obtained by the Single Molecule Array (Simoa) technique. This assay used a combination of monoclonal antibodies and puri ed bovine NfL as a calibrator.
Neuroimaging acquisition and processing 3T MRI scans were processed before download as previously described [9][10] . FreeSurfer (ADNI phase 1, grand opportunity and phase 2 data was run with FreeSurfer version 5.1, while phase 3 with version 6.0) was used for further analysis.
Regions of interest (ROIs) were extracted, including the bilateral hippocampal volumes (adjusted for intracranial volume [ICV] by calculating the residual term from a linear regression of hippocampal volume versus ICV among ApoE negatively CU participants) and an AD signature cortical thickness (mean thickness in the entorhinal, inferior temporal, middle temporal, and fusiform cortices) 11 4 . And for amyloid PET, we selected a cutoff of 1.11, which is the upper 95% con dence interval above the mean of a young normal control group 15 . Furthermore, mean ±2 SD from Aβ-negative CU controls (+2 SD for amyloid PET, tau PET, CSF tau, and plasma NfL; -2 SD for CSF Aβ42, hippocampal volume, temporal cortical thickness, and FDG PET), along with 90% sensitivity for AD were used as a sensitivity analysis.

Statistical analyses
Demographics and continuous biomarkers between different groups were compared using Kruskal-Wallis test, and binary biomarkers using Fisher exact test. Associations between biomarkers were analyzed using Spearman rank correlation (ρ), Cohen's kappa coe cient (κ), and percentage agreement (concordance). Prevalence estimates for AT(N) categories were calculated in CU, CI, MCI and AD participants with 95% con dence intervals generated using bootstrap resampling (n=1,000). The relationship between AT(N) variants and cognitive trajectories (longitudinal MMSE and CDRSB) was examined by linear mixed-effects (LME) model (including age, sex and education as covariates, and time as a categorical variable) with subject-speci c intercepts and slopes. The goodness of LME models with different AT(N) variants was assessed by marginal R 2 . All analyses were performed in IBM SPSS Statistics 20, with signi cance set at p<0.05, 2-tailed.

Study participants
Demographics are presented in Table 1, more detailed information are shown in Supplement Table 1. Between CU and CI participants, there was no signi cant difference in age, while there were more females, longer time for education, and less prevalence of APOE e4 in the CU group. No signi cant differences were found between MCI and AD (subgroups of CI) in age, gender, education, or APOE e4. MMSE, Aβ42, hippocampal volume, temporal cortical thickness and FDG PET decreased sequentially, while CDRSB, amyloid and tau PET, CSF tau and NfL increased sequentially among CU, MCI, CI and AD groups. However, there was no signi cance in NfL between MCI and AD. As plasma NfL level was reported to be positively associated with age (ρ=0.471, p<0.01) [16][17] , we divided participants into younger and older groups using a median split (age=72.25y) and found there was a signi cant difference in NfL levels between the resulting groups (p<0.001). Therefore the prevalence of (N)+ using NfL was likely to vary by age in the present cohort, so we calculated the cut-point based on age strati cation.

Prevalence measures in CU participants
Prevalence for AT(N) categories in CU and CI participants are summarized in Figure 2, Figure 3 and Supplement Table 3-4. When only considering A and T in CU, A-T-were the most common categories (range 43.5% [A1T1; 95% con dence interval, 36.6%-50.5%] to 62.0% [A2T2; 95% con dence interval, 55.0%-68.8%]). Comparing A biomarkers, slightly more were negative when using CSF Aβ42 than amyloid PET. Positivity in T was highest when using CSF p-tau both in the case of A+ or A-, while the prevalence of T+ was much less when using tau PET ( Figure 2A). These results indicate that using CSF p-tau may greatly increase the positive rate of T component compared to tau PET in CU participants.

Longitudinal cognition
Overall ndings for longitudinal cognition using continuous predictors are summarized in Figure 4, Figure 5 and Supplement  Figure 4G, H). The best AT(N) variants capturing changes in cognition in CDRSB and MMSE were A2T3[N]2 (R 2 =7.84%) and A2T1[N]1 (R 2 =12.29%) respectively, but not all included biomarkers contributed signi cantly ( Figure 4B, E). For the marginal R 2 in CU participants relatively low, we considered whether random effects (i.e., individual heterogeneity) accounted for more variance. Then we calculated conditional R 2 using MRI imaging biomarkers ([N]2 for CDRSB and [N]1 for MMSE). Adding individual heterogeneity and slope for time as the random effect, conditional R 2 increased to 19.32% and 33.55% in CDRSB and MMSE respectively. These results indicated that longitudinal cognition in CU participants was mainly associated with individual characteristics; and MRI imaging measurements were the best biomarkers to predict cognitive changes.
In CI participants, individual characteristics were not signi cantly associated with cognitive decline. Almost all single AT(N) biomarkers could predict longitudinal cognition except CSF p-tau (p=0.061) and t-tau (p=0.051) in CDRSB, and marginal R 2 using MRI imaging and tau PET was relatively higher than others. The AT(N) variants combining CSF Aβ42, tau PET, and temporal cortical thickness were the best predictors in both CDRSB and MMSE, and almost all included variables contributed signi cantly ( Figure 5B

Sensitivity analyses
We repeated the AT(N) prevalence analyses using alternative cut-points (Supplement Table 8). Using cutoffs from 90% sensitivity for AD, except for more amyloid positivity using CSF Aβ42 in CU participants, other results were in concordance with main cutoffs. However, cut-points de ned by mean ± 2 SD from Aβ-negative CU controls were more conservative.
There was the least tau positivity using CSF rather than PET, and temporal cortical thickness in all participants was negative.

Discussion
In this study, we found that different combinations of AT(N) biomarkers have different effects on category prevalence and predictions of cognitive decline. First, it is not surprising that the composition of AT(N) categories is different between CU and CI. Categories representing AD continuum was the most common in CI while more subjects with non-AD pathologic change were observed in CU 2 . Moreover, different AT(N) variants give considerable differences in prevalence, such as less prevalence of T+ when using tau PET than CSF p-tau in all groups, and more prevalence of N+ using uid biomarkers in CU. Finally, different AT(N) combinations have different associations with cognitive changes, with differences between CU and CI (MRI imaging was more in uential in CU participants, and tau PET in CI participants). and amyloid PET is "L-shaped" rather than linear ( Figure 1A) [23][24] . This may be owing to a temporal offset between them 6,25-26 . And in T biomarkers, the correlation between CSF p-tau and tau PET is imperfect, because p-tau seems to plateau later in the disease 27 while the tau PET signal keeps increasing continuously 28  Since cognition is also a continuum and the de nition of CU is independent from biomarker ndings according to the NIA-AA research framework 2 . In our study, the overall prevalence of A+ in CU participants is similar, in consistent with a metaanalysis demonstrated 35 . But the increments of amyloid positivity between 2 groups were higher when using amyloid PET. This may be due to CSF analysis detecting cerebral Aβ accumulation earlier than PET 6,[25][26] . Same ndings were shown in tau positivity by comparing CSF and PET owing to temporal lag 28,36 . Among the neurodegeneration biomarkers, CSF t-tau and plasma NfL are more common in CU participants, while there are no evident differences in CI. These results in line with several studies that found CSF t-tau and blood NfL are increased before symptom onset 28,37 .
In order to verify the prevalence ndings across AT(N) categories, we repeated prevalence calculations in different cutpoints strategies and found the results were not completely consistent. This nding highlights the optimization of categorization strategies is important for future studies.
Here, we analyzed the prediction of different AT(N) variants on longitudinal cognition which evaluated by both CDRSB and MMSE. CDRSB may enable a more detailed analysis of subtle changes with different staging of dementia severity 38 .
First of all, optimal variants differ by clinical stage. Only MRI imaging measures were signi cantly associated with cognition changes in CU participants, whereas the best model for predicting cognition in CI included CSF Aβ42, tau PET and cortical thickness. When using a single AT(N) biomarker for prediction, there was no obvious difference between CSF and PET amyloid plaque. This nding may indicate CSF Aβ42 and amyloid PET can be used interchangeably in practice as several literatures reported 4,39 , which is consistent with the characteristic of "A" as state biomarkers 3 . However, CSF ptau is increased earlier in the disease stage than tau PET 5,7,39 . Therefore, between 2 subgroups of CI, the difference of tau PET was more signi cant than that of CSF p-tau. This may be the reason why tau PET far exceeded CSF p-tau on longitudinal cognitive prediction in CI. And early tangle pathology of tau PET was better for prediction on CDRSB than MMSE, which is consistent with the characteristics of the scales. Compared to other N biomarkers, we found MRI imaging measures were the best, especially cortical thickness. Since the hippocampal volume is highly related to ICV 11 , and differing methods of adjusted volume by ICV associated with gender, age and study populations may affect study power 40 . A study proposed to use thickness measurements rather than volumes to assess neurodegeneration in AD cohorts with a large age range 40 . Our results also suggested that cortical thickness may predict cognition more precisely. Same ndings were shown when considering interactions in CI, but the interactions dominated the main effects. This result demonstrates that although AT(N) variants can predict cognitive changes, their marginal effects rely on the time level. Overall, we got relatively robust results in this cohort (MRI imaging for CU and the combination of tau PET and cortical thickness using MRI for CI). Compared to a recent study recruiting participants from Swedish BioFINDER 7 , we con rmed the importance of tau PET in AD diagnosis and staging, and highlight that cortical thickness may be of great signi cance to cognition declines and staging severity.

Limitations
This study has several limitations. First, the sample size in our study was moderate, which may have some effects on study power. Further, the greater individual heterogeneity of CU participants may be a reason of low marginal R 2 . Then, differences were observed among different cut-points strategies, and between binary or continuous biomarkers as another study reported 7 . So, more approaches to selecting normal/abnormal cutoffs or alternatives of the binarization (semicontinuous scale, i.e. the centiloid scale) 21 are needed to be tested. Finally, we only determined typical AD biomarkers in this study. With the emergence of more and more biomarkers, they may also need to be included.

Conclusions
Collectively, the proposal of the A/T/N framework makes a more precise division of the Alzheimer's continuum from the pathology 2 , but different biomarkers for de ning AT(N) cannot be used identically. Each component of biomarkers for AT(N) system classi cation plays different roles in the stating and staging of AD, and the optimal combinations for cognitive prediction may differ by clinical stage. Furthermore, different strategies for discontinuous biomarkers will be an important area for future studies.

Contributors
Rong-Rong Lin: analysis and interpretation of the data, and drafting the manuscript; Yan-Yan Xue, Xiao-Yan Li, and Yi-He Chen: data acquisition, analysis and interpretation of the data; Qing-Qing Tao: funding, designed the study, critical revision of the manuscript; Zhi-Ying Wu: funding, conceptualized and designed the study, critical revision of the manuscript. All authors reviewed the manuscript. All authors have contributed to the manuscript revising and editing critically for important intellectual content and given nal approval of the version and agreed to be accountable for all aspects of the work presented here. All authors read and approved the nal manuscript.

Funding
This study was supported by grants from the Key Research and Development project of Zhejiang Province (2019C03039) and the National Natural Science Foundation of China (81970998 Data are presented as mean (SD) or n (%).