Validity and Performance of Blood Biomarkers for Alzheimer Disease to Predict Dementia Risk in a Large Clinic-Based Cohort

The MEMENTO cohort enrolled outpatients consulting in 26 French memory clinics.^25,26 The main inclusion criteria in the MEMENTO cohort were a Clinical Dementia Rating (CDR) scale score ≤0.5, a very mild to mild cognitive impairment (<1 SD below the age, sex, and education-level thresholds in one or more cognitive test(s)), or isolated SCC (for people older than 60 years). Cognitive complaints were assessed using visual analog scales. Exclusion criteria included history of head trauma with persistent neurologic deficits, stroke in the last 3 months or with persistent neurologic deficits, brain tumor, epilepsy, schizophrenia, known mutation in familial AD genes, and illiteracy.

All participants provided informed consent. The study protocol was approved by the ethics committee “CPP Sud-Ouest et OutreMer III” and registered at ClinicalTrials.gov (NCT01926249).

During a 5-year follow-up period, patients underwent a clinical and neurologic evaluation at least yearly. The Mini-Mental State Examination and the CDR were also systematically performed. At baseline and during annual follow-up visits, Free and Cued Selective Reminding Test, Delayed Matching-to-Sample 48, Letter and 2 minutes semantic fluencies (2 minutes task for letter P and animals), image naming, praxis assessment, and Trail Making Test (TMT) A and B were administered. Training sessions were organized to optimize standardization across centers for quoting of CDR and neuropsychological tests. All incident cases of dementia were reviewed by an expert committee, blinded to genetic and biological biomarkers using the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV criteria). The etiologic diagnosis of dementia was made according to the international criteria (National Institute on Aging - Alzheimer's Association criteria for AD,² DLB consortium for dementia with Lewy bodies,²⁷ and Rascovsky criteria for frontotemporal lobar degeneration²⁸).

Participants underwent brain scanning on 1.5 or 3 T MRI machines. The morphological protocol consisted of 1-mm isotropic 3D T1-weighted and 2D T2-weighted FLAIR images, acquired after a standardization of the imaging parameters by a dedicated team of neuroimaging specialists (Centre pour l’Acquisition et le Traitement des Images, Paris, France).²⁹ All scans were centralized, quality-controlled, and postprocessed to obtain standardized measurements for each participant. Intracranial volumes were extracted with Statistical Parametric Mapping; hippocampal volumes were estimated with the SACHA software and cortical thickness with FreeSurfer 5.3 in each region of interest of the Desikan atlas.²⁶ The MRI AD signature was calculated based on regionally specific cortical thinning in a list of Desikan's cortical areas (entorhinal, inferior temporal, middle temporal, inferior parietal, fusiform, and precuneus) corresponding to Dickerson's signature of AD.³⁰

Some participants were proposed to participate in an amyloid-PET ancillary study at baseline (Insight-PreAD) or during follow-up (AMYGING, NCT02164643, 2 years on average after inclusion in MEMENTO). Radioligands were ¹⁸F-florbetapir (¹⁸F-AV45) in some centers or ¹⁸F-flutemetamol (Vizamyl ) in the others. Amyloid positivity was defined for each radioligand according to a previously described procedure (thresholds: florbetapir > 0.88, flutemetamol > 1.063).³¹

Apolipoprotein E (ApoE) genotype was determined by KBiosciences (UK).²⁵

Lumbar puncture was optional and was performed in a subsample of participants, as previously described.²⁵ We restricted the analysis on CSF markers to participants who had their lumbar puncture within 1 year of blood biomarkers measurements. Amyloid-β 42 peptide (Aβ₄₂), Aβ₄₀, total tau, and phosphorylated-tau (p181-tau) levels were measured using the standardized commercially available INNOTEST (Fujirebio, Belgium). CSF NfL were measured using single molecular array ultra-sensitive immunoassay method (Simoa) with commercial kits on a Quanterix HD-X analyzer (Quanterix, MA).

Blood samples were collected in each memory clinic participating in MEMENTO. Standard biological measurements (including estimated glomerular filtration rate [eGFR]) were performed at local biochemistry departments. Study-specific blood samples were collected at baseline by venipuncture into gel-separator tubes for serum and EDTA tubes for plasma. They were left at room temperature for 30 minutes to coagulate, before centrifugation at 1,500g for 15 minutes at 4°C. Once separated, small volumes of serum and plasma were aliquoted to avoid thawing cycles (250 µL in 2 mL Sarstedt cryotubes) and stored at −80°C. Consequently, the analyses were performed after only 1 freeze/thaw cycle. Blood biomarkers were measured using Simoa technology with commercial kits on a Quanterix HD-X analyzer: plasma Aβ42, Aβ40, total tau using Neurology 3-Plex A Advantage Kit (item No. 101995), serum p181-tau using p181-tau Advantage V2 Kit (item No. 103714), and serum NfL using NF-light Advantage Kit (item No. 103186). Sensitivity cutoffs (functional lower limit of quantification) were (pg/mL) 0.69 for NfL, 0.33 for p181-tau, 0.25 for total tau, 2.7 for Aβ40, and 0.56 for Aβ42. All the datasheets and validation reports containing the technical characteristics of the immunoassays (calibration curves, coefficients of variation, reproducibility precision, etc) are available on the manufacturer's website (quanterix.com/simoa-assay-kits/). Measures were performed in Bordeaux University Hospital research platform in the centralized biobank (Bordeaux Biothèques Santé, Centre de Ressources Biologiques), blinded of any other data.

Descriptive data are presented as percentages for qualitative variables and as the median, 1st and 3rd quartiles for quantitative variables. Correlations between CSF and blood biomarkers concentrations were estimated with Spearman rank correlation coefficients. Associations between amyloid-PET status and blood biomarkers concentrations were tested with nonparametric Wilcoxon rank tests, and the performance of blood biomarkers to discriminate amyloid positivity on PET was determined with receiver operating characteristic (ROC) curve analyses.

We ran survival analysis to predict AD dementia risk and modeled time to dementia for incident cases, time to death, or time to end of follow-up, whichever occurred first. Individuals who developed either clinically defined AD or mixed AD dementia² were considered “AD dementia converters.” Participants who did not develop dementia during follow-up or who developed dementia of another etiology were classified as “non-AD dementia converters” and censored at time of dementia diagnosis. Kaplan-Meier survival curves were performed to model tertiles of blood and CSF biomarkers in relation to dementia risk. Tertiles were used because no “pathologic” cutoffs of blood biomarkers are available to date.

For multivariable analysis, we performed Cox proportional hazard models (with continuous measures of blood biomarkers). Blood biomarker concentrations were log-transformed. We defined a “clinical” reference prediction model based on data collected in usual clinical practice: age, sex, educational level, memory (total recalls of the Free and Cued Selective Reminding Test), and executive (TMT-B) performance. We also defined a “research” reference model including less accessible measures on top of the “clinical” prediction model: ApoE genotype and MRI cortical thickness in AD-signature regions. The predictive value of blood biomarkers were modelized alone or in combination with the “clinical” or “research” reference models. To reduce optimism in estimations, we computed 5-fold cross-validation probabilities. Subsequent indicators were based on these cross-validated probabilities. For each model, we calculated a c-index for discrimination capacities (from 0 to 1, the higher the better), a Brier score for prediction errors (the lower the better), and c-index difference between a model of interest and a reference one (0 indicates no difference, positive differences are in favor of the model of interest). Brier scores and c-index difference 95% confidence intervals (95% CI) were estimated using bootstrapping. As sensitivity analyses, we ran the same Cox proportional hazard models on subgroups of patients based on their baseline CDR (0 or 0.5) or based on the nature of MCI (amnesic or nonamnesic). As another sensitivity analysis, we also adjusted the Cox models on comorbidities known to affect blood biomarkers concentrations^32,33: eGFR and history of vascular event (stroke or myocardial infarction).

Analyses were performed with SAS 9.4, and programs were derived from Nancy Cook and colleagues' SAS macros (ncook.bwh.harvard.edu/sas-macros.html).

Anonymized data will be shared by request from any qualified investigator for the sole purpose of replicating procedures and results presented in the article and as long as data transfer is in agreement with EU legislation on the general data protection regulation.

From 2011 to 2014, 2,323 nondemented outpatients from 26 French memory clinics were consecutively enrolled in the MEMENTO study.²⁵ Among them, 2,277 (98.0%) had at least 1 AD or neurodegeneration blood biomarker measured. Among them, 357 (15.7%) had a lumbar puncture within 1 year of blood sampling. Amyloid-PET was available for 649 individuals (150 [23.1%] were amyloid-positive). The demographic, clinical, neuropsychological, biological, and MRI characteristics at baseline of the main analytical sample and of the CSF subsample are presented in Table 1. Data on sex differences in CSF and blood biomarkers measures are provided in eTable 1 (links.lww.com/WNL/C445).

The follow-up rates for the whole sample were 91.4% at 1 year, 85.3% at 2 years, 78.6% at 3 years, 73.0% at 4 years, and 64.3% at 5 years. Over the 5-year follow-up period, 257 incident cases of clinically defined AD (or mixed) dementia were diagnosed, amounting to an incidence rate of 2.84 per 100 person-years (PY) (95% CI = 2.50; 3.20). In the CSF subsample, the incidence of AD dementia was 4.40 per 100 PY [95% CI = 3.39; 5.62] (n = 64). Characteristics of “AD converters” and “nonconverters” are presented in Table 1 for both samples. In the AD “nonconverters” group, 63 patients developed dementia of another etiology in the whole cohort, of whom 13 patients were in the CSF subsample.

Blood Aβ42/40 ratio and concentrations of p181-tau and NfL were associated with amyloid-PET status (Figure 1A), with the highest effect size for p181-tau: Mean concentration (SD) was 0.88 (0.55) pg/mL in patient with negative amyloid-PET vs 1.44 (0.85) in patients with positive amyloid-PET (+63%, Cohen's d = 0.82). Mean Aβ42/40 ratio was 6.12 (2.13) vs 5.13 (1.28) (−16%, d = 0.49), and mean NfL concentration was 19 (8.2) vs 24 (15) pg/mL (+26%, d = 0.48). Total tau concentrations did not differ according to amyloid-PET status.

ROC analyses of blood biomarkers to identify amyloid-PET positivity are shown in Figure 1B. Blood p181-tau was the best discriminator with an area under the curve (AUC) of 0.74 (95% CI = 0.69; 0.79).

CSF and plasma biomarkers were mainly intercorrelated. Except for total tau, blood biomarkers were correlated with their equivalence in the CSF (Table 2).

Tertiles of blood and CSF AD and neurodegeneration biomarkers were modeled in relation to AD dementia risk over 5-year follow-up using Kaplan-Meier survival analyses (Figure 2 and eTable 2 [links.lww.com/WNL/C445] for incidence rates). Blood and CSF p181-tau, NfL, and Aβ42/40 ratio were associated with incident AD dementia risk following a dose-response pattern (p < 0.0001, p < 0.0001, and p = 0.0033, respectively, on log-rank tests). For p181-tau and NfL, blood or CSF tertile concentrations led to similar incidence rates. Blood Aβ42/40 ratio was less efficient than CSF Aβ42/40 ratio. CSF t-tau was associated with incident AD dementia (p < 0.0001) but not blood t-tau (p = 0.43).

Table 3 shows accuracy of blood biomarkers alone or in combination to predict AD dementia. Baseline blood p181-tau was the best blood biomarker alone to predict AD dementia (c-index = 0.731 [95% CI = 0.694; 0.768]). The model prediction performance was improved to 0.757 (95% CI = 0.726; 0.789) for the combination of NfL + p181-tau + Aβ42/40 ratio (difference in c-index difference = 0.027 [95% CI = 0.010; 0.043]). The “clinical” reference model (age, sex, education level, memory and executive performance) predicted AD dementia risk with a c-index = 0.884 (95% CI = 0.862; 0.905). Model prediction performance rose to 0.899 (95% CI = 0.882; 0.917) when adding blood p181-tau and Aβ42/40 ratio (difference in c-index 0.016 [95% CI = 0.006; 0.026]). The “research” reference model (clinical data + ApoE genotype and MRI cortical thickness in AD-signature regions) predicted AD dementia with a c-index = 0.907 (95% CI = 0.888–0.926). Performance increased to 0.917 (95% CI = 0.901; 0.933) when adding blood p181-tau and Aβ42/40 ratio (difference in c-index 0.009 [95% CI = 0.001; 0.018]).

Most of the memory care centers do not have access to quantitative volumetric analyses but uses visual read of atrophy such as the Scheltens scale.³⁴ Thus, we added this grading of hippocampal atrophy to the “clinical” reference model (eTable 3, links.lww.com/WNL/C445). Interestingly, predictive performances were similar to the “research” reference model, suggesting that brain MRI assessment methods (quantitative measure of cortical thickness in AD-signature regions or hippocampal grading) are equivalent to predict AD dementia risk in this large clinic-based cohort.

Because it can be argued that clinical status at baseline (SCC or MCI) and the nature of MCI can influence the predictive value of biomarkers, we performed a first sensitivity analysis where we split the analytical sample in 2 subgroups according to baseline CDR (CDR = 0 or CDR = 0.5) and another where we divided the MCI group in 2 subgroups (amnestic or nonamnestic). The accuracy of blood p181-tau to predict AD dementia was higher in patients with CDR = 0 (c-index = 0.830 [95% CI = 0.694; 0.967]) than in patients with CDR = 0.5 (c-index = 0.697 [95% CI = 0.658; 0.737]). However, in both subgroups (as in the whole cohort), blood biomarkers added little to 5-year dementia risk prediction models including traditional predictors (eTable 4, links.lww.com/WNL/C445). Regarding the nature of MCI, blood biomarkers were equivalent to predict dementia risk in nonamnestic MCI and in amnestic MCI, but clinical and research models were slightly more performant in amnestic MCI (eTable 5, links.lww.com/WNL/C445).

To test whether comorbidities known to affect blood biomarkers concentration may change our findings, we performed the same statistical analyses after the exclusion of patients with chronic kidney diseases (eGFR<60 mL/min/1.73 m²) and we found the same results (eTable 6, links.lww.com/WNL/C445). We also added in the “clinical” and “research” models the eGFR and the history of cardiovascular event. We found no substantial change in the predictive performances of the models (eTable 7, links.lww.com/WNL/C445).

Finally, we ran the same Cox proportional hazard models in the CSF subsample. The “clinical” reference model predicted AD dementia risk with a c-index = 0.831 [95% CI = 0.769; 0.893]. Performance reached 0.856 [95% CI = 0.807; 0.905] after addition of all CSF biomarkers and 0.856 [95% CI = 0.801; 0.911] for the best blood biomarker combination (p181-tau + Aβ42/40 ratio + NfL). The “research” reference models had c-index of 0.858 [95% CI = 0.811–0.904], 0.870 [95% CI = 0.836; 0.904] with addition of CSF biomarkers, and 0.862 [95% CI = 0.818; 0.907] with addition of blood p181-tau + Aβ42/40 ratio + NfL). The addition of either CSF or blood biomarkers to the “research” reference model did not significantly improve its performance (c-index differences 0.011 [95% CI = −0.018; 0.040] and 0.005 [95% CI = −0.015; 0.025], respectively).