Longitudinal pathways of cerebrospinal fluid and positron emission tomography biomarkers of amyloid-β positivity

Mismatch between CSF and PET amyloid-β biomarkers occurs in up to ≈20% of preclinical/prodromal Alzheimer’s disease individuals. Factors underlying mismatching results remain unclear. In this study we hypothesized that CSF/PET discordance provides unique biological/clinical information. To test this hypothesis, we investigated non-demented and demented participants with CSF amyloid-β42 and [18F]Florbetapir PET assessments at baseline (n = 867) and at 2-year follow-up (n = 289). Longitudinal trajectories of amyloid-β positivity were tracked simultaneously for CSF and PET biomarkers. In the longitudinal cohort (n = 289), we found that participants with normal CSF/PET amyloid-β biomarkers progressed more frequently toward CSF/PET discordance than to full CSF/PET positivity (χ2(1) = 5.40; p < 0.05). Progression to CSF+/PET+ status was ten times more frequent in cases with discordant biomarkers, as compared to csf−/pet− cases (χ2(1) = 18.86; p < 0.001). Compared to the CSF+/pet− group, the csf−/PET+ group had lower APOE-ε4ε4 prevalence (χ2(6) = 197; p < 0.001; n = 867) and slower rate of brain amyloid-β accumulation (F(3,600) = 12.76; p < 0.001; n = 608). These results demonstrate that biomarker discordance is a typical stage in the natural history of amyloid-β accumulation, with CSF or PET becoming abnormal first and not concurrently. Therefore, biomarker discordance allows for identification of individuals with elevated risk of progression toward fully abnormal amyloid-β biomarkers, with subsequent risk of neurodegeneration and cognitive decline. Our results also suggest that there are two alternative pathways (“CSF-first” vs. “PET-first”) toward established amyloid-β pathology, characterized by different genetic profiles and rates of amyloid-β accumulation. In conclusion, CSF and PET amyloid-β biomarkers provide distinct information, with potential implications for their use as biomarkers in clinical trials.


Introduction
The abnormal deposition of amyloid-β into plaques is considered one of the earliest neuropathological events in Alzheimer's disease. Current research diagnostic criteria for Alzheimer's disease promote the use of either CSF (low levels of CSF amyloid-β 42 ) or PET imaging (high brain retention of amyloid-β PET tracers) as equivalent measures of amyloid-β pathology [1][2][3][4]. Amyloid-β PET and CSF amyloid-β 42 usually provide highly concordant information, which has justified their interchangeable use in the diagnosis of Alzheimer's disease dementia [5,6]. However, a significant proportion of cases showing discordant CSF and PET results has been consistently reported across multiple studies  (Supplementary Table 1), reaching up to ≈20% in prevalence [15] in preclinical/prodromal disease phases (e.g., [13-15, 18, 19, 28]). To date, the interpretation of the CSF/ PET discordant findings is still unclear and the longitudinal evolution of the concordant and discordant CSF/PET groups remains unknown. While it has been suggested that discordance in amyloid-β biomarkers might be related to measurement issues, due to different detection thresholds of CSF vs. PET techniques [13], recent studies have indicated that biomarker discordance might be biologically relevant, with each amyloid-β biomarker providing distinct information [15,19,35]. There are several reasons to hypothesize that CSF and PET biomarkers measure different aspects of amyloid-β pathology. CSF amyloid-β 42 is a marker of soluble amyloid-β, only indirectly related to fibrillar amyloid-β in the brain [36]. Furthermore, CSF amyloid-β 42 can be affected by variations in amyloid precursor protein and amyloid-β production, and by non-fibrillar aggregation [12]. On the other hand, amyloid-β PET is a direct measure of fibrillar amyloid-β deposition in the brain, as reported in antemortempostmortem correlative studies [37][38][39][40].
Hence, we hypothesize that discordance in amyloid-β biomarkers provides unique information and that the existence of discordant groups may result from differences in amyloid-β processing and kinetics in the CSF vs. in the brain. In particular, while CSF amyloid-β 42 represents an instantaneous measure of ongoing amyloid-β accumulation, reflecting the availability of soluble amyloid-β in the CSF, amyloid-β PET represents an integral measure of previously accumulated amyloid-β, providing information on the resulting fibrillar amyloid-β plaque accumulation. Thus, we expect the instantaneous CSF amyloid-β 42 concentration to be predictive of the subsequent rate of change in amyloid-β PET standardized uptake value ratio (SUVr), but not vice-versa. To test this hypothesis, we aimed to: (a) test whether baseline CSF amyloid-β 42 levels significantly predict future rate of change in amyloid-β PET SUVr; (b) test whether amyloid-β PET SUVr at baseline does not significantly predict rate of change of CSF amyloid-β 42 ; (c) investigate the longitudinal evolution of CSF amyloid-β 42 and amyloid-β PET biomarkers in a longitudinal cohort of participants stratified into four groups of concordant and discordant biomarkers (csf−/pet−, CSF+/pet−, csf−/PET+, and CSF+/PET+).

CSF amyloid-β 42
Following overnight fasting, CSF samples were acquired in the morning through lumbar puncture. CSF amyloid-β 42 measurement was performed based on the multiple xMAP Luminex platform (Luminex Corp) with INNO-BIA Alz-Bio3 immunoassay kit (Innogenetics) [41]. Longitudinal CSF samples belonging to the same participant were measured on the same plate. All data were derived from the UPENNBIOMK_MASTER.csv dataset.

Amyloid-β PET
[18F]Florbetapir PET acquisition and analysis are thoroughly described elsewhere [13]. Following recommendations for longitudinal studies [42] we selected a composite reference region, including the whole cerebellum, brainstem/pons, and subcortical white matter, for SUVr estimation. A summary measure of global cortical uptake was calculated as the weighted-average uptake across regions of interest including frontal, anterior/posterior cingulate, lateral parietal, and lateral temporal lobes. All data were derived from the UCBERKELEYAV45_08_09_18.csv dataset.

Other data
Additional baseline and longitudinal measures of neurodegeneration, tau pathology, cognition, and neuropsychiatric symptomatology were available; genetic information was also retrieved. Number of participants with additional data available, at baseline and at follow-ups, is reported in the next paragraphs.

FDG-PET
FDG-PET acquisition and analysis are thoroughly described elsewhere [43]. A composite index for Alzheimer's diseaselike hypometabolism was available in n = 857 participants (n = 417 at 2-year follow-up). Data were retrieved from the ADNIMERGE.csv dataset.

Structural MRI
T1-weighted MRI image acquisition and analysis are thoroughly described elsewhere [44]. Hippocampal, ventricles and whole-brain volume data were available in n = 799 participants (n = 614 at 1-year, n = 427 at 1.5-year, n = 335 at 2-year follow-up). Volumes of target regions were corrected for total intracranial volume. Data were retrieved from ADNIMERGE.csv dataset.

Other MRI measures
Presence, localization and size of infarcts were evaluated by trained physicians on T1-and T2-weighted and FLAIR MRI images in n = 818 participants [45]. Volume of white matter hyperintensities, indexing vascular brain injury, was quantified based on a Bayesian approach to segment highresolution T1 and FLAIR MRI images in n = 816 participants [46]. Data were retrieved from the MRI_IN-FARCTS_11_16_15.csv and UCD_ADNI2_WMH_10_ 26_15.csv datasets.
Other CSF measures CSF phosphorylated-tau 181 (n = 867 at baseline; n = 307 at 2-year follow-up) and total-tau (n = 866 at baseline; n = 306 at 2-year follow-up) measurements were derived through AlzBio3 immunoassay. Data were retrieved from the UPENNBIOMK_MASTER.csv dataset. CSF amyloidβ 40 measurements were obtained in n = 822 participants at baseline through 2D-UPLC-tandem mass spectrometry. Data were retrieved from the UPENNMSMSABETA.csv and UPENNMSMSABETA2.csv datasets.

Genetic data
APOE genotyping was performed using DNA extracted by Cogenics from a 3 ml aliquot of whole blood (n = 867). We considered both the number of APOE-ε4 and APOE-ε2 alleles. Polygenic hazard score (n = 783) was computed based on a combination of APOE and 31 other genetic variants, as detailed elsewhere [47]. Data were retrieved from the APOE.csv and DESIKANLAB.csv datasets.

Multivariable regression analyses
We fitted linear, quadratic, and cubic regression models, testing whether baseline CSF amyloid-β 42 predicted the rate of change in [18F]Florbetapir PET SUVr (over 2-year and 4-year follow-up; n = 608 and n = 348 participants, respectively), and whether baseline [18F]Florbetapir PET SUVr predicted the rate of change in CSF amyloid-β 42 (over 2-year follow-up; n = 305 participants). CSF amyloid-β 42 and [18F]Florbetapir PET were included in the regression models as continuous variables. Sex, age, number of APOE-ε4 alleles and clinical group were entered as nuisance covariates. The assumptions of normal distribution and homoscedasticity of residuals, as well as the absence of multicollinearity between predictors, were tested and met in all regression models.

Classification of participants
Participants were classified as either csf−/pet−, csf−/PET+, CSF+/pet− or CSF+/PET+, based on cutoffs for abnormality of CSF amyloid-β 42 and [18F]Florbetapir PET biomarkers. CSF amyloid-β 42 measurements were classified as abnormal (CSF+) when <192 pg/ml. This cutoff was initially established to allow optimal discrimination between healthy controls and Alzheimer's disease cases confirmed at autopsy [48] and later validated using a diagnosis-free data-driven approach [49]. [18F] Florbetapir PET images were classified as abnormal (PET +) when global cortical uptake was >0.79 SUVr [42]. This cutoff (optimized for longitudinal studies) was derived using linear regression from a well-validated cutoff for cross-sectional assessment (1.11 SUVr) [42], initially derived in healthy controls [50], then validated for discrimination between autopsy-confirmed healthy controls and Alzheimer's disease patients [51], and again using mixture model analysis [35].

Longitudinal analyses
To investigate the simultaneous longitudinal changes in CSF amyloid-β 42 and [18F]Florbetapir PET biomarkers, stratified by csf−/pet−, csf−/PET+, CSF+/pet− or CSF+/PET+ subgroups, we selected all participants (n = 289) with CSF and PET amyloid-β biomarkers available both at baseline and at the 2-year follow-up examinations, and obtained within 90 days. Differences in prevalence of longitudinal outcomes across subgroups were tested by means of χ 2 tests.

Group comparisons
Group comparisons for continuous variables were performed via general linear modeling. Post-hoc tests were performed using Bonferroni correction for multiple comparisons. The assumptions of normality of residuals and homoscedasticity were tested; when not met, the estimates were bootstrapped with 1000 replicates, generating bias-corrected and accelerated 95% confidence intervals. For categorical variables, χ 2 tests were applied. Cox regression analysis was used for estimation of hazard risks. When the assumption of proportionality of hazards was not met, we run a stratified model. Sex, age, number of APOE-ε4 alleles, and clinical group were included as nuisance covariates.
All statistical analyses were performed in PASW Statistics 18 (SPSS Inc.), setting the significance level at p < 0.05 (two tailed). Graphical renderings were performed in RStudio (v.

Results
Longitudinal changes in amyloid-β CSF and PET measures in the whole cohort Table 1 shows baseline demographic and clinical characteristics of the participants, divided by clinical group (n = 867).  Longitudinal changes in amyloid-β CSF and PET measures for each of the CSF/PET subgroups Table 2  At baseline (n = 867), when compared to the remaining groups, the groups with discordant amyloid-β biomarkers presented with "intermediate" levels of the pathological amyloid-β biomarker, i.e., (i) the CSF+/pet− group had significantly lower CSF amyloid-β 42 levels than those with normal CSF amyloid-β 42 values (both csf−/pet− and csf−/PET+), but significantly higher CSF amyloid-β 42 levels as compared to the CSF+/PET+ group; (ii) the csf−/PET+ group had significantly higher [18F]Florbetapir PET retention compared to either the csf−/pet− or CSF+/pet− groups but significantly lower as compared to the CSF+/PET+ level (Table 2). This pattern was consistently observed at the regional level as well (Supplementary  Table 2).
When systematically comparing discordant and concordant groups considering (i) biomarkers for tau pathology and neurodegeneration, (ii) clinical and neuropsychological measures, (iii) genetic assessments and (iv) confounding factors (data available in up to n = 867 participants; see "Participants and Methods" for the precise values), we found that: (i) A systematic comparison of biomarkers for tau pathology and neurodegeneration among CSF/PET groups revealed no significant differences between the CSF+/pet− and csf−/PET+ groups. Only the CSF+/PET+ group presented with increased baseline tau pathology (as measured by CSF phosphorylated-tau 181 ) and neurodegeneration (as measured by CSF total-tau, [18F] Fluorodeoxyglucose PET, and MRI), and faster rate of neurodegeneration as compared to the csf−/pet− group (Supplementary Table 3).
(ii) Groups with discordant biomarkers were not at increased risk of clinical progression (csf−/PET+: hazard risk = 1.14 [95% CI: 0. 44 Table 4). When examining the severity of neuropsychiatric depressive symptoms, we found that the GDS score was significantly higher in the csf−/PET+ group ( Table 4).

Discussion
Discordance in CSF and PET measures of amyloidosis is fairly common, occurring in ≈20% of individuals at preclinical and prodromal stages of Alzheimer's disease [15]. Yet, guidelines on how to correctly interpret discordant CSF/PET results are lacking. In this study, we validated the hypothesis that discordance in CSF/PET amyloid-β biomarkers provides unique information and that the existence of discordant groups may result from differences in amyloid-β processing and kinetics in the CSF vs. in the brain. Through systematic assessment of discordant amyloid-β biomarker results at both crosssectional and longitudinal levels, we showed that biomarker discordance represents a natural phase in the progression of amyloid-β pathology, with either CSF or PET amyloid-β biomarkers becoming abnormal first and not concurrently. We subsequently identified two mutually exclusive trajectories ("CSF-first" vs. "PETfirst") toward established amyloid-β pathology, characterized by different genetic profiles and rates of amyloid-β accumulation ("fast" vs. "slow" accumulators, respectively). In addition, we showed that biomarker discordance was not due to systematic errors in CSF and PET measurements. Instead, biomarker discordance carried prognostic relevance as a marker of progression toward established amyloid-β pathology, thus allowing identification of individuals at subsequent risk of clinical progression, cognitive decline, increased neurodegeneration, and tau pathology.
First, for the longitudinal cohort as a whole, decreased CSF amyloid-β 42 levels at baseline were predictive of subsequent increases in amyloid-β plaques in the brain as measured by longitudinal [18F]Florbetapir PET (Fig. 1). This observation supports the concept that decreased CSF amyloid-β 42 might be a proxy for the rate at which soluble amyloid-β isoforms are being trapped into nonsoluble amyloid-β aggregates in the brain as measured by PET. Consistent with this observation, and in participants with normal CSF/PET amyloid-β biomarkers, CSF amyloid-β 42 , and [18F]Florbetapir PET levels did not become pathological at the same time. Preferentially, CSF amyloid-β 42 became abnormal first without detectable brain [18F]Florbetapir PET uptake (7% of csf−/pet− cases progressed to CSF+/pet−). This evidence, coupled with the finding that about 30% of CSF+/pet− cases progressed to CSF+/PET+, qualifies isolated CSF positivity as a preferential pathway from fully normal to fully pathological amyloid-β biomarkers, consistent with previous reports [19,24]. However, our cross-sectional and longitudinal analyses in this large cohort provided evidence for the existence of an additional alternative pathway, where [18F]Florbetapir PET became abnormal first. Firstly, and consistent with previous reports [6-8, 13-15, 17, 18, 20-32, 34] (Supplementary Table 1), we observed isolated [18F]Florbetapir PET positivity in a non-negligible percentage of cases. This cross-sectional evidence alone questions whether isolated CSF amyloid-β 42 positivity represents the only possible pathway toward fully abnormal amyloid-β biomarkers. Indeed, our longitudinal results showed that-in a proportion of participants with normal amyloid-β biomarkers at baseline-[18F]Florbetapir PET became abnormal first, while levels of CSF amyloid-β 42 were still normal (3.5% of csf−/pet− cases progressed to csf−/PET+). In addition, about 20% of the participants with isolated pathological [18F]Florbetapir PET progressed to CSF+/PET+. These longitudinal data suggest that cases with isolated [18F] Florbetapir PET positivity are part of an alternative (PET-first) pathway toward established amyloid-β pathology (Fig. 3).
Of note, the frequency of progression from csf−/pet− to either CSF+/pet− or PET+/csf− was not significantly different; this finding supports the concept that both CSFfirst and PET-first pathways are relevant. Furthermore, our study found a slower rate of [18F]Florbetapir PET accumulation in the csf−/PET+ compared to the CSF+/pet− group. Thus, the existence of cases with discordant biomarkers could be partly explained by differences in amyloid-β accumulation dynamics, with csf−/PET+ and CSF+/pet− cases representing "slow" and "fast" amyloid-β accumulators, respectively. This explanation is also supported by the lower prevalence of APOE-ε4ε4 genotype among csf−/PET+ individuals found in this study ( Table 2), considering that a slower rate of amyloid-β plaque accumulation is found in non-APOE-ε4 carriers compared to APOE-ε4 carriers [54]. In contrast, CSF+/pet− cases, despite presenting no evidence of significant amyloid-β PET tracer retention, accumulated plaques in the brain at a relatively fast rate, a finding consistent with the higher prevalence of APOE-ε4ε4 genotype found in the CSF+/pet− group compared to that in the csf−/PET+ group (Table 2).
On a technical note, it has been argued that occurrence of discordant findings might be inherently dependent on the choice of cutoffs to establish biomarkers' abnormality [36]. Still, independently of the type of cutoffs adopted, previous studies consistently reported a proportion of discordant amyloid-β biomarker results, even when using cutoffs specifically designed to optimize agreement between biomarkers [14, 16-18, 20, 21, 23, 25-27, 29-31] (Supplementary Table 1). Similarly, it has been proposed that discordant cases might be due to false negative or false positive results, CSF analytical factors or failed PET scans [19]. Our assessment of a comprehensive series of confounding factors in this study argues against this explanation. Altogether, our results argue against the claim that discordant findings in amyloid-β biomarkers systematically represent false positive/negative results, suggesting instead that amyloid-β biomarkers discordance, per se, carries relevant information with respect to brain amyloid-β burden and accumulation rate. Most importantly, amyloid-β biomarkers discordance represents a strong predictor of future progression toward full amyloid-β biomarkers abnormalities (Fig. 2), thus allowing prompt identification of individuals that will be at higher subsequent risk for clinical progression and cognitive decline, steeper rates of neurodegeneration, and tau pathology [14,15,19,28].
This study has some limitations. We acknowledge that the sample size of discordant biomarker groups was relatively limited (n = 62 for CSF+/pet− and n = 44 for csf−/ PET+ cases). Although we used all longitudinal CSF/PET data available from ADNI at the time of our study, joint CSF/PET examinations were available at two longitudinal time points only, thus future studies incorporating additional longitudinal CSF/PET time points will be valuable to confirm our findings.
In conclusion, biomarker discordance allows for identification of individuals with elevated risk of progression toward fully abnormal amyloid-β biomarkers, with subsequent risk of neurodegeneration and cognitive decline. Our results also suggest that there are two alternative pathways ("CSF-first" vs. "PET-first") toward established amyloid-β pathology, characterized by different genetic profiles and rates of amyloid-β accumulation. In conclusion, CSF and PET amyloid-β biomarkers provide complementary rather than redundant information, with potential implications for their use as biomarkers in clinical trials. or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_ Acknowledgement_List.pdf.

Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of interest.
Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons. org/licenses/by/4.0/.