|Home | About | Journals | Submit | Contact Us | Français|
To empirically assess the concept that Alzheimer’s disease (AD) biomarkers significantly depart from normality in a temporally ordered manner.
Multi-site, referral centers
We studied 401 elderly cognitively normal (CN), Mild Cognitive Impairment (MCI) and AD dementia subjects from the Alzheimer’s Disease Neuroimaging Initiative. We compared the proportions of three AD biomarkers – CSF Aβ42, CSF total tau (t-tau), and hippocampal volume adjusted by intra-cranial volume (HVa) - that were abnormal as cognitive impairment worsened. Cut-points demarcating normal vs. abnormal for each biomarker were established by maximizing diagnostic accuracy in independent autopsy samples.
Within each clinical group in the entire sample (n=401) CSF Aβ42 was abnormal more often than t-tau or HVa. Among the 298 subjects with both baseline and 12 month data, the proportion of subjects with abnormal Aβ42 did not change from baseline to 12 months in any group. The proportion of subjects with abnormal t-tau increased from baseline to 12 months in CN (p=0.05) but not in MCI or dementia. In 209 subjects with abnormal CSF AB42 at baseline, the percent abnormal HVa, but not t-tau, increased from baseline to 12 months in MCI.
Reduction in CSF Aβ42 denotes a pathophysiological process that significantly departs from normality (i.e., becomes dynamic) early, while t-tau and HVa are biomarkers of downstream pathophysiological processes. T-tau becomes dynamic before HVa, but HVa is more dynamic in the clinically symptomatic MCI and dementia phases of the disease than t-tau.
Biomarkers of Alzheimer’s disease (AD) occupy an essential place in recently formulated diagnostic criteria for AD (1–5) where their role is to identify the pathophysiological processes underlying cognitive impairment or to help predict time to dementia (6–15). AD biomarkers are also increasingly used in clinical trials both as inclusion criteria and as outcome measures.
At present five AD biomarkers are sufficiently validated to be employed in therapeutic trials, large observational research studies and, on occasion, for clinical diagnostic purposes (16–19). The major biomarkers of brain Aβ deposition are low CSF Aβ42 and positive PET amyloid imaging (20–30). Biomarkers of neuronal injury or neurodegeneration are elevated CSF total tau (t-tau) and phosphorylated tau (31–33), decreased FDG uptake on PET in temporo-parietal cortex (34–36), and atrophy on structural MRI (sMRI) in medial, basal and lateral temporal lobes, and medial parietal isocortex (37–43).
Some of the authors recently proposed a hypothetical model of Alzheimer’s pathophysiology (44, 45) describing the temporal evolution of these five biomarkers based on the assumption that they do not change suddenly or simultaneously but rather over many years in an ordered, more sequential, manner and like-wise approach a pathological level in an ordered manner. The model does not assume a strict sequence whereby one biomarker changes then stops, the next changes then stops, etc. Rather the model assumes that the maximum rate of change moves from one class of biomarker to the next, and, as the disease progresses all biomarkers become progressively more abnormal simultaneously, albeit at rates that change over time in an ordered manner. It was proposed as a hypothetical biomarker cascade with validation awaiting additional data.
Empirical testing of this hypothetical biomarker cascade model (44, 45) can be approached in various ways (46); however all require that different biomarkers be directly compared with each other in the same subjects. This can be conceptualized as plotting all biomarker values on a common graph with the vertical axis representing biomarker severity and the horizontal axis representing disease stage and/or time. Our present objective was to evaluate some aspects of the hypothetical biomarker cascade model by characterizing the prevalence of biomarker abnormalities at different disease stages defined by clinical cohort and by MMSE. Comparing the proportions of subjects with abnormal biomarker values allowed us to express significant biomarker deviations from normal in the same units for each biomarker. Cut points denoting abnormality for each biomarker were derived from independent autopsy cohorts which limited our analysis to three of the five major AD biomarkers – CSF Aβ, CSF tau, and sMRI. Our objective was to test the hypothesis that CSF Aβ, CSF tau, and sMRI significantly depart from normality in a temporally ordered manner as disease progresses.
All Alzheimer’s Disease Neuroimaging Initiative (ADNI) subjects who had usable baseline CSF and sMRI data were considered for our analysis. We also analyzed serial (baseline and 12 month) data if both CSF and sMRI were obtained at the 12 month visit. Written informed consent was obtained for participation in these studies, as approved by the Institutional Review Board (IRB) at each of the participating centers http://www.ADNI-info.org.
A standardized protocol was implemented in ADNI to quantify biomarker concentrations in each of the CSF baseline aliquots using a multiplex xMAP Luminex platform (Luminex Corp, Austin, TX) with Innogenetics (INNO-BIA AlzBio3, Ghent, Belgium) immunoassay kit-based reagents which was validated in Vanderstichele et al (47) and Shaw et al (32). Details can be found at (http://www.adni-info.org/index.php).
All subjects were scanned at 1.5T with a 3D magnetization prepared rapid acquisition gradient echo (MPRAGE) imaging sequence (48). All images were corrected for image distortion due to gradient non-linearity using “GradWarp” (49), for B1 non-uniformity as necessary (48), and for residual inhomogeneity using “N3” (50) with a software pipeline running at the Mayo Clinic Rochester. Hippocampal and intracranial volumes for both the autopsy sample (see below) and the main ADNI analysis sample were measured at Mayo Clinic using FreeSurfer software (version 4.5.0) (51). Each subject’s raw hippocampal volume (HV) was adjusted by his/her total intracranial volume (HVa) (52).
In order to create a common analytic framework to compare different biomarkers we elected to define each biomarker measure as either normal or abnormal. This requires that a cut point be established in the continuous distribution of values for each biomarker. Arguably the least biased and most valid approach to establishing biomarker cut points is to use an independent cohort in which ground truth is established by autopsy. Cut points for CSF Aβ and t-tau were established by Shaw et al (32, 53) using an ADNI-independent autopsy cohort of subjects followed in the University of Pennsylvania Alzheimer’s Disease Core Center who were diagnosed by NIA-Reagan criteria (20). The cut points were chosen in order to maximize accuracy in separating high from low probability autopsy proven AD.
To our knowledge, however, cut points have not been established for hippocampal atrophy using an independent autopsy data set that employed imaging methods identical to those used in ADNI. To obtain a HVa cut point we used an independent sample of 53 subjects at the Mayo Clinic who had ante mortem MRI within 3.5 years of death and an autopsy diagnosis of high or low probability of AD using NIA-Reagan criteria. These subjects had been enrolled in the Mayo Alzheimer’s Disease Patient Registry or the Alzheimer’s Disease Research Center. Our cut point was based on first calculating an adjusted hippocampal volume using the formula HVa = HV − (2.546201 + 0.002139314 * intra-cranial volume). Here, HVa is the residual value obtained after regressing hippocampal volume versus intra-cranial volume. Next we performed a receiver operating characteristic (ROC) curve analysis and chose a cut point in order to maximize accuracy in separating high from low probability autopsy proven AD. That is, we chose the cut point to maximize Sensitivity − (1-Specificity).
Each subject had three binary outcomes for their baseline visit defined as normal (y = 0) or abnormal (y = 1) for each of the three biomarkers. Because these can be considered repeated measures data having a binary outcome, we used generalized estimating equations (GEE) with the logit link and an exchangeable working correlation matrix to estimate and compare the proportion of subjects having an abnormal biomarker. Our predictors were clinical group (CN vs. MCI vs. AD), biomarker and their interaction. We used likelihood ratio tests to perform a global test of biomarker differences after including clinical group in the model and to test for an interaction between clinical group and biomarker. We used Wald tests to perform pair-wise comparisons of abnormality of a biomarker separately within each clinical group. This analysis was then repeated replacing clinical group with MMSE, modeled as a restricted cubic spline with knots at the 10th, 50th, and 90th percentiles of the MMSE distribution. We also performed a subset analysis among those with abnormal baseline CSF Aβ and examined t-tau and HVa within group.
Among those subjects who also had a 12-month visit, we used baseline and 12-month data and fit a GEE model that included group, visit, and biomarker along with all interactions. From this model we estimate separate rates of biomarker abnormality by group, time point, and biomarker. We also performed a subset analysis among those with abnormal baseline CSF Aβ and examined within group change in t-tau and HVa from baseline to 12 months.
The cut point for defining abnormal hippocampal volume was based on a sample of 43 Mayo subjects with ante mortem sMRI and autopsy diagnoses of high-probability AD (56% women, median age at death 85 years) and 10 low probability (60% women, median age at death 85 years) (Supplementary Table S1). The optimal HVa ratio cut point value was 0.48 (Supplementary Fig. S1). With this cut point, NIA-Reagan low AD probability subjects in the MRI training sample were separated from high AD probability with a classification accuracy of 79%, sensitivity of 74%, and specificity of 100%. Overall discrimination was high with an area under the ROC curve of 0.90. Previously established CSF cut point values of Aβ42 192 pg/ml and t-tau 92 pg/ml were used to define abnormal studies (32).
The main analysis of 401 ADNI subjects included 116 CN, 196 MCI, and 89 AD subjects with baseline data. In total, 298 subjects had baseline and 12 month data. The baseline age distributions varied somewhat among the three groups of subjects (ANOVA p = 0.026) and there were differences in the proportion of women that ranged from 33% of MCI to 49% of CN (Table 1). MMSE score and proportion of APOE e4 carriers varied by clinical group in the expected manner.
For each of the three biomarkers evaluated, the median baseline group values became more abnormal (p<0.001, linear trend test) (Table 1, Fig. 1) and the percentage of subjects with abnormal biomarker studies (Table 1, Fig. 2A) increased in an ordered manner in the CN, MCI, AD groups. The same pattern held at 12 months. The percentage of subjects with abnormal biomarker studies increased monotonically for each biomarker with decreasing MMSE (Fig. 2B). A Supplement to Figure 2 shows plots of individual patient trajectories for each biomarker within clinical diagnosis.
Among all 401 subjects at baseline, CSF Aβ42 was abnormal more often than t-tau or HVa in each clinical group (p<0.001 across all pair-wise tests, Table 2). The percentage of abnormal t-tau was greater than the percentage of HVa among CN (21% vs. 8%, p=0.003) but did not differ among MCI or AD dementia subjects (Table 2).
We performed a sub-analysis of baseline CSF t-tau and HVa among only those subjects (47 CN, 145, MCI and 82 AD) who had abnormal CSF Aβ42 values at baseline (Supplementary Table S2). The percentage of these CSF Aβ42 positive subjects with abnormal t-tau and HVa increased monotonically by clinical group in the following order, CN, MCI, AD dementia (Fig. 2C) and increased monotonically with decreasing MMSE (Fig. 2D). In this sub-analysis, t-tau was abnormal more often than HVa in CN. There was a trend for more abnormal t-tau than HVa in MCI but no difference in AD dementia (Table 2).
Among the 298 subjects with both baseline and 12 month data, the proportion of subjects with abnormal CSF Aβ42 did not change from baseline to 12 months in any diagnosis group (CN: p=0.15, MCI: p=0.15, AD: p=0.33) (Fig. 3A). The proportion of subjects with abnormal t-tau increased from baseline to 12 months in CN subjects (p=0.05) but not in MCI or AD dementia subjects (p=0.40, p>0.99) although the absolute levels of t-tau in the CN group were far lower than in the MCI group. There was no difference in the proportion of subjects with abnormal HVa in CN subjects (p>0.99) or AD dementia subjects (p=0.16) from baseline to 12 months, but a significant increase in the proportion of subjects with abnormal HVa for MCI.
The preceding paragraph describes the proportion of subjects by group with abnormal biomarker values at baseline and 12 months. The change in biomarker values in native units is seen in Fig 1. CSF Aβ42 did not change from baseline to 12 months in any group (p=0.52, p=0.13, p=0.51 for CN, MCI, and AD). T-tau increased in CN (p=0.002) but not in MCI or AD (p=0.12, p=0.36). HVa decreased in all groups (p<0.001).
We also performed a sub-analysis among only those subjects (n=209) who had abnormal CSF Aβ42 values at baseline and who also had both baseline and 12 month data (Fig. 3B). Neither the proportion of subjects with abnormal Aβ42 nor t-tau changed from baseline to 12 months in any diagnosis group. The proportion of subjects with abnormal HVa increased from baseline to 12 months for MCI (p<0.001) but did not differ for CN or AD dementia subjects (p=0.32, p=0.16).
Our overall objective was to test for evidence of temporal ordering of CSF AB42, t-tau, and HVa (54, 55). We were limited to evaluating these three AD biomarkers for which independent autopsy cohorts were available to select normal/abnormal cut points in an unbiased manner. A biomarker value for an individual subject at a given point in time (and by extension, the percent abnormal across a group at a given disease stage) is a function of two phenomena: the elapsed time from the initial deviation of the biomarker away from normality to the present, and the average rate of change of the biomarker over this period of time. An analogy to motion would be distance traveled from abnormality, average rate of change, and elapsed time. Our data consisted of measures of each biomarker value at a fixed point or points in time in subjects who entered the study at different stages of the disease. We observed whether a subject had reached a certain distance from normality, but cannot individually identify the contributions from average rate of change and elapsed time. We can, however, draw valid inferences from our data about the combined effect of time elapsed from onset and average rate of change, and refer to this as relative dynamic ordering of biomarkers.
The data presented here supports several key principles in our recently proposed hypothetical biomarker cascade model (44, 45). These include: 1) all biomarkers become progressively more abnormal as subjects worsen clinically. 2) Reduction in CSF Aβ42 denotes an upstream pathophysiological process that significantly departs from normality (i.e., is dynamic) early in the pathophysiological cascade while subjects are clinically asymptomatic, but does not change greatly during the clinically symptomatic MCI and dementia phases of the disease. 3) T-tau and HVa are biomarkers of downstream neurodegenerative pathophysiological process that are dynamic later as subjects approach and move through the clinically symptomatic phases of the disease (56–64). 4) Cognitive decline is more closely related to biomarkers of neuronal injury than brain Aβ load (65–72). 5) T-tau is more dynamic earlier than HVa, but the proportion of abnormal biomarker studies are similar in symptomatic subjects such that HVa “catches up” to t-tau as symptoms worsen. This is supported visually by the steeper slope of HVa vs ttau in Fig 2.
The hypothetical model represents an idealized trajectory of an individual subject who progresses to pathologically pure AD dementia. Our sample, however, almost certainly consists of a mixture of subjects, many of whom are on the AD pathway, but many, particularly in the CN and MCI groups, who are not (73–78). The fact that both elevated t-tau and hippocampal atrophy can occur in other conditions that lead to dementia (79) such as cerebro-vascular disease has led to the belief that of the three biomarkers we examined, CSF Aβ42 should have the greatest specificity for AD (80). Consequently, we performed a sub-analysis of subjects with abnormal CSF Aβ at baseline in order to isolate those subjects who we were somewhat more confident were in the AD pathophysiological pathway (56, 81–86). Our results concerning evidence for biomarker ordering led to similar conclusions in the “all-subjects” and the abnormal CSF Aβ 42 samples.
Using the percent-abnormal metric might not seem to be an obvious first choice for comparing biomarkers. Other options however prove to be untenable. For example, using biomarker values in native measurement units precludes direct comparisons of biomarkers because they are not on a common scale. Z scores or percentiles are also not tenable since by construction half the subjects in the sample must be above and half below average for each biomarker. This constraint would make it impossible to test our major question – i.e., is one biomarker abnormal more often than another at different stages of the disease. The obvious advantage of comparing biomarkers on a percent-abnormal basis is that the scoring method is standardized for all biomarkers over the entire cognitive continuum. A limitation is that the results are highly sensitive to the cut point values and hence the validity of the analysis depends on selecting valid cut points for each biomarker. We used cut point values for each biomarker that were based on an independent autopsy verified sample, and used the same pathological criteria for all biomarkers – i.e., NIA Reagan low vs. high AD probability. While using cut points based on diagnostic sensitivity, rather than accuracy, in the autopsy reference standard might seem a better approach, that is not the case. Imagine a biomarker with identical distributions in autopsy proven high vs low probability AD. Fixing a cut point at a sensitivity of 90% in high probability autopsy proven AD would lead to the conclusion that 90% of autopsy low probability AD cases had abnormal biomarker values, and the biomarker in question comes “early” in the pathophysiological cascade. This would clearly be erroneous. Thus our choice of selecting cut points for this study in a manner that is constrained by both sensitivity and specificity, as is done for all clinically applied diagnostic tests, seems prudent.
Our hypothetical model (44, 45) was intended to represent an idealized trajectory that an individual subject follows from a time prior to appearance of any AD pathophysiology in the brain through end stage AD dementia when all biomarkers have become maximally abnormal. The optimal way to test this model would be to follow the trajectory of multiple biomarkers over several decades in a large group of subjects who enter the study prior to the first appearance of any AD pathophysiology, are followed through the symptomatic stages of the disease to autopsy. Given that the total course of the disease may span 30 years or more it will take many years to accumulate the necessary data to construct a temporally accurate disease model. While the data is being accumulated, perhaps the only realistic approach to empirical analysis is to attempt to build models of disease in a piece-wise fashion from individual subjects who are at various stages in the disease as we have done here.
Support from NIH AG11378 and ADNI. All authors had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Manuscript preparation by Samantha Wille.
*Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/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.ucla.edu/wp-content/uploads/how_to_apply/ADNI_Authorship_List.pdf