A Multi-Dimensional Comparison of Alzheimer’s Disease Neurodegenerative Biomarkers

Alzheimer’s Disease Neuroimaging Initiative (ADNI)

We conducted cross-sectional and longitudinal analyses of participants enrolled in the ADNI data-base (http://adni.loni.usc.edu). ADNI is a longitudinal, multicenter study launched in 2003 to assess serial changes in CSF biomarkers, blood biomarkers, neuroimaging markers, and neuropsychological assessments in three groups of elderly individuals: cognitively normal (CN), mild cognitive impairment (MCI) and AD. All AD individuals met the National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria for probable AD, with Mini-Mental State Examination (MMSE) scores between 20 and 26 and Clinical Dementia Rating (CDR) global scores of either 0.5 or 1. Criteria for amnestic MCI include MMSE scores between 24 and 30, and CDR scores of at least 0.5. CN individuals had MMSE scores of 24 or higher and a CDR score of 0. All individuals were recruited from more than 50 sites across the USA and Canada. Detailed diagnostic criteria are available in http://www.adni-info.org. All the data we used were from ADNI 1, 2 and GO.

Data used in preparation of this article were obtained from the ADNI database. The study was approved by institutional review boards of all participating institutions, and written informed consent was obtained from all participants or their guardians according to the Declaration of Helsinki (consent for research).

Participants

We extracted all information from the latest merged document “ADNIMERGE.csv” updated on May 24, 2019. In our study, individuals were included if they underwent the assessments of CSF Aβ (labeled A) and CSF P-tau (labeled T). A total of 1,238 participants were recruited from the ADNI database. In further cognitive and neuroimaging analyses, 13 participants without cognitive tests and 201 without imaging data were excluded (Supplementary Figure 1).

Biomarkers of neurodegeneration or neuronal injury

Studies have examined the following potential N markers: hippocampal volume atrophy (HV), FDG-PET, CSF total tau (T-tau), plasma neurofilament light (NFL), CSF NFL, CSF α-synuclein, CSF neurogranin (Ng), CSF progranulin (PGRN), CSF soluble triggering receptor expressed on myeloid cells 2 (sTREM2), CSF Visinin-like protein 1 (VILIP-1), CSF YKL-40 (or chitinase-3-like protein 1), and synaptosome-associated protein 25 (SNAP-25) at baseline (see Supplementary Table 1) [17].

CSF measurements

In the present study, CSF Aβ₄₂, p-tau, T-tau, and NFL were measured at the ADNI biomarker Core Laboratory (University of Pennsylvania) on the xMAP-Luminex multiplex platform (Luminex Corp, Austin, TX) using Innogenetics immunoassay kit-based reagents. CSF NFL (Unit: ng/L) was measured with a novel, sensitive sandwich ELISA method (NF-light ELISA kit, UmanDiagnostics AB, Sweden) in the University of Gothenburg, as described previously [18]. The lower limit of quantification for CSF NFL assay was 50 ng/L. Level of CSF α-synuclein was measured using LuminexMicroPlex [19]. CSF PGRN and sTREM2 (Unit: pg/mL) were measured with a MSD platform based ELISA assay, which was previously described and validated [20–22]. CSF Ng (Unit: pg/mL) was measured by electrochemiluminescence using the Ng-specific monoclonal antibody Ng7 as the coating antibody [23]. Both CSF VILIP-1and SNAP-25 were tested by a sandwich ELISA (together with the Erenna® immunoassay platform) [24]. CSF YKL-40 (Unit: ng/mL) was determined by the MicroVue YKL-40 ELISA assay at Washington University [25]. All CSF samples were performed in duplicate. Detailed information can be obtained at http://www.adni-info.org.

Plasma measurements

Blood samples were collected, centrifuged, ali-quoted, and stored at –80°C. Plasma NFL was analyzed by the single molecule array (Simoa) technique in Clinical Neurochemistry Laboratory (University of Gothenburg, Sweden) using the same methodology as previously described [26]. The plasma NFL assay used a combination of monoclonal antibodies and purified bovine plasma NFL as calibrator (details available in http://adni.loni.usc.edu). All tested samples were above the detection limit, analytical sensitivity was < 1.0 pg/mL. All samples were measured in duplicate.

Neuroimaging

Acquisition protocols and preprocessing steps for structural MRI and FDG-PET are available at http://adni.loni.ucla.edu/. Structural MRI was performed using a Vision 3.0T or 1.5T scanner (Siemens, Erlangen, Germany). Regional brain volume estimates were processed using Free-surfer software package version 4.3 and 5.1 image processing framework for the 1.5T and 3.0T MRI images, respectively. Middle temporal lobe (MidTemp) volume, entorhinal cortex thickness (Entorhinal), whole brain, ventricular volume and fusiform volume were selected for further analysis to compare the measures of brain atrophy.

FDG-PET data for each subject were pre-pro-cessed by a series of steps as described in detail elsewhere [7, 27]. In this study, the mean standardized uptake value ratio (SUVR) of previously validated AD-typical hypometabolism regions (angular, temporal, and posterior cingulate) was estimated as FDG SUVR of each participant for further analysis [27].

Cognitive scores

MMSE, Alzheimer Disease Assessment Scale 11 score (ADAS11), Alzheimer Disease Assessment Scale 13 score (ADAS13), Rey Auditory Verbal Lear-ning Test (RAVLT) Immediate, and Functional Activities Questionnaire (FAQ) were used to assess overall cognitive ability and evaluate outcome measures.

AT(N) measurements

As for A and T categories, we adopted the established cutoffs based on the ADNI database to define the diagnostic test results: positive or negative [28]. CSF amyloid positive (A+) and negative (A–) were determined by a cutoff value of 192 pg/ml for CSF Aβ₄₂ [28]. CSF p-tau positive (T+) and negative (T–) were defined as a score above and below a cutoff value of 23 pg/ml. Binaryzation of N markers (+/–, abnormal/normal) was obtained from a Youden index-derived cutoff (ROC analyses included AD dementia as cases and CN participants as controls).

Statistical analysis

To find the best N marker(s), we conducted a three-step analysis in our study.

In the first step, we included common neurodegenerative biomarkers generated from blood test, CSF, MRI, and PET. We compared the changing trend of each N marker in the preclinical, prodromal, and dementia stages of AD: A–CN, A + CN, A + MCI, and A + AD. Then, we filtered out those non-significant marker(s) and calculated the diagnostic accuracies of selected N markers using area under the receiver operating characteristic curve (AUROC) with binary logistic regression models. Receiver operating characteristic curve (ROC) and logistic regression (LR) analyses were done using IBM SPSS Statistics 26.

Secondly, in the cross-sectional analyses, the effects of each candidate N biomarker on cognitive (MMSE, ADAS11, ADAS13, RAVLT, and FAQ) were investigated using a linear regression model. Longitudinally, the correlations of those candidate N biomarkers with cognitive performance over time were further compared by linear mixed-effects models. In the cross-sectional and longitudinal analyses, all the included biomarkers and outcome variables (cognitive scores) were all Z log-transformed to normalize the distributions, a facilitating the comparison of biomarkers. In these results, β coefficients refer to standardized effects (β= 1 implies that an increase of Z log biomarker was associated with a 1-SD increase in the dependent variable). All regression analyses were adjusted for age, gender, APOE ɛ4, years of education and diagnosis at baseline for cognitive performance.

Finally, unadjusted Kaplan-Meier (KM) analysis with the log-rank test to determine cognitive decline was performed. Clinical progression was defined as followings: 1) CN converted to MCI or AD, or their CDR scores rose to 0.5 or more, 2) MCI subjects converted to AD at follow-up or their MMSE scores decrease more than 3 points. More precisely, we conducted the subgroup analyses as follows: 1) using N markers only (N + versus N-); 2) using the combination of “A” marker and N markers, i.e., A–N–versus A–N+ versus A + N- versus A + N+. Then, we ran multivariate Cox proportional hazard models adjusted for age, gender, APOE ɛ4, and years of education at baseline.

All tests were two-tailed, and statistical significance was set at p < 0.001. All statistical analyses were performed using the R statistical software (version 3.5.1) and IBM SPSS Statistics 26.

Basic characteristics of the population

A total of 1,238 individuals (including 372 CN, 632 MCI, and 234 AD) were enrolled in our study. The basic demographic, clinical, and psychometric characteristics of our study population were summarized in Table 1. The total participants had a median age of 73.5 years (interquartile range IQR, 68.3, 78.1 years), a median of 16.0 years of education (IQR 14, 18 years), and a female proportion of 44.5% (Table 1). Of these participants, 782 (63.17%) were assigned to A positive (A+) group, and 644 (52.01%) were assigned to T positive (T+) group, 905 participants were categorized into AD continuum (161 A–CN, 116 + CN, A + MCI, and A + AD) when we further added the amyloid marker.

Screening the candidate N biomarkers

We primarily selected several reported markers of neurodegeneration or neuronal injury: HV, FDG-PET, T-tau, plasma NFL, CSF NFL, α-synuclein, Ng, PGRN, STREM2, YKL-40, VILIP-1, and SNAP-2 (Supplementary Figure 2). We explored whether these biomarkers could be the candidate N biomarkers. N biomarkers were closely tied with aging during the preclinical, prodromal, and dementia stages of AD. We compared levels of baseline N markers from A–CN to A + CN, to A + MCI, and to A + AD (see Fig. 1a). Supplementary Figure 2 and Supplementary Table 2 showed the levels of these 12 makers in these four subgroups. To better compare their trends, combined models were showed in Fig. 1a. In this study, we found hippocampal volume (mean: A–CN 7447.31, A + CN 7317.07, A + MCI 6622.94, and A + AD 5845.55, mm³) and FDG-PET (mean: A–CN 1.33, A + CN 1.29, A + MCI 1.22, and A + AD 1.06, SUVR) declined significantly as AD progressed (p < 0.0001). Moreover, CSF T-tau, Ng, CSF NFL, and plasma NFL were also increased significantly (p < 0.0001, see Supplementary Table 2). STREM2, PGRN, α-synuclein, YKL-40, VILIP-1, and SNAP-25 did not show significant change from the preclinical to dementia stages of AD (Fig. 1a). Finally, we included six candidate N biomarkers for further analysis: N1 HV, N2 FDG-PET, N3 T-tau, N4 plasma NFL, N5 CSF NFL, and N6 Ng.

OPEN IN VIEWER

Fig. 1

Screening the candidate N biomarkers. A) The trajectories of primarily candidate N biomarkers from the preclinical, prodromal, and dementia stages of AD. Based on the baseline levels of each biomarker (mean±SD) in four subgroups (A–CN, A + CN, A + MCI, and A + AD), we delineated an approximate trend Graph. Control: Aβ- controls (A–CN); AD continuum: Aβ+ controls (A + CN), patients with Aβ+ MCI (A = MCI), and patients with Aβ+ AD dementia (A + AD). A–indicates Aβ negative; A + indicates Aβ positive, definite A: CSF Aβ₄₂ < 192 ng/L. B) Receiver operating characteristic curve (ROC) curves for N biomarkers for the Alzheimer’s disease (AD) cases versus cognitively normal (NC) subjects. HV, hippocampal volume; FDG-PET, ¹⁸F-fluorodeoxyglucose-positron emission tomography; T-tau, CSF total tau; plasma NFL, plasma neurofilament light chain; CSF NFL; α-synuclein, Ng, neurogranin; PGRN, progranulin; SD, standard deviation.

Accuracy of N biomarkers in predicting AD

ROC analyses of AD patients versus CN group provided cutoffs concentrations which showed the greatest diagnostic accuracy. Detailed information on the diagnostic sensitivity and specificity was summarized in Fig. 1b (Supplementary Table 3). The greatest value of the area under the ROC curve (AUC) was obtained for N2 FDG-PET (0.922). FDG-PET had the greatest sensitivity value of 86.34% and greatest specificity value of 85.44% (cutoff value, 1.199 SUVR). However, the diagnostic specificity for N1 HV was 89.09%, which was greater than the other five biomarkers (cutoff value, 6594 mm³). For T-tau, the AUC value and sensitivity value were 0.826 and 81.14, respectively (cutoff value, 74.4 pg/ml). For plasma NFL, CSF NFL, and Ng, the AUC values were 0.760, 0.768, and 0.704, respectively (see Fig. 1b). Prevalence of AT(N_x) categories based on these above N cutoffs were showed in the Supplementary Figure 1. To further compare diagnostic utilities of N markers, we compared their diagnostic accuracy in the amyloid positive subgroup. We compared those N markers in A + subgroups (A + T + and A + T-). In A + T+, the diagnostic accuracy of FDG-PET in differentiating AD from CN and MCI was much better than other markers (Supplementary Figure 3). Besides, the diagnostic accuracy of HV in differentiating AD from CN and MCI was comparable to FDG-PET. In A + T- subgroup, there were no significant differences in diagnostic accuracy between the six biomarkers (Supplementary Figure 4).

Associations between N markers and cognitive scores

In multivariable models adjusting for age, gender, years of education, APOE ɛ4, and diagnosis at baseline, the associations between cognitive scores (MMSE, ADAS11, ADAS13, RAVLT, and FAQ) and N markers were shown in Supplementary Table 4 and Fig. 2a. The levels of HV, FDG-PET, and T-tau were all correlated with all the above cognitive variables (MMSE, ADAS11, ADAS13, RAVLT, and FAQ). Notably, FDG-PET showed strongest associations with 4 cognitive variables (MMSE, ADAS11, ADAS13, and FAQ), followed by HV and T-tau (abs-olute value of β: FDG-PET > HV > T-tau > CSF NFL > plasma NFL > CSF Ng; see Fig. 2a). A few cogni-tive variables (ADAS11 and ADAS13) were associated with plasma NFL (p < 0.01) and CSF NFL (p < 0.01). No cognitive variables were associated with CSF Ng.

OPEN IN VIEWER

Fig. 2

Associations between candidate N markers and clinical variables (cognitive scores and imaging markers). A) Association between N biomarkers and cognitive scores cross-sectionally. B) Association between N biomarkers and cognitive scores longitudinally. Beta values were all transformed to absolute values of β. All analyses were adjusted for age, gender, education, APOE ɛ4 status, and baseline diagnosis. All data were z log transformed. N1, MRI Hippocampal volume; N2, ¹⁸F-fluorodeoxyglucose-positron emission tomography; N3, CSF total tau; N4, plasma neurofilament light chain; N5, CSF NFL; N6, CSF neurogranin.

In the longitudinal analyses, the associations of all cognitive variables with HV, FDG-PET, and T-tau measures remained significant (Supplementary Table 5 and Fig. 2b). FDG-PET was closely associated with cognitive variables (abstract value of beta value: MMSE, RAVLT, and FAQ: FDG-PET > T-tau>HV>CSF NFL > plasma NFL > CSF Ng). Plasma and CSF NFL had moderate associations with these cognitive variables (p < 0.008).

Above all, three biomarkers were significantly associated with cognitive decline and neuroimaging: FDG-PET, HV, and T-tau. We also conducted a comparison between N biomarkers and brain atrophy (MRI measurements: volumes of ventricles, whole brain, entorhinal, fusiform and MidTemp), which yield similar results that FDG-PET, HV, and T-tau were best N biomarkers (see Supplementary Material). Accordingly, these three biomarkers were further compared in the following studies.

Ability of N markers to predict future clinical progression

The inter-group comparison (as evaluated with Cohen’s Kappa values) of these three valuable biomarkers (FDG-PET, MRI HV, and CSF T-tau) across diagnostic groups is shown in Fig. 3. For these three N biomarkers, there was no agreement among the diagnostic groups (all Kappa value < 0.4).

OPEN IN VIEWER

Fig. 3

Inter-group comparison between three top neurodegeneration biomarkers in different diagnostic groups: MRI Hippocampal volume, FDG-PET, and CSF T-tau. Precentral of concordances and discordances between MRI, PET, and CSF were compared in five diagnostic groups (CN, SMC, EMCI, LMCI, and AD). A) MRI HV versus FDG-PET (concordant MRI-PET-, concordant MRI + PET + , discordant MRI + PET-, and discordant MRI-PET+). B) MRI HV versus CSF T-tau (concordant MRI-CSF-, concordant MRI + CSF+, discordant MRI + CSF-, and discordant MRI-CSF+). C) FDG-PET versus CSF T-tau (concordant PET-CSF-, concordant PET + CSF+, discordant PET + CSF-, and discordant PET-CSF+). Cohen’s Kappa statistics allowed numerical comparisons between pairs of profiles obtained using different N biomarker. Agreement was defined as coefficient values > 0.4 (fair agreement) ranging up to 1 (perfect agreement). CN, cognitively normal; SMC, subjective memory concern; MCI mild cognitive impairment; EMCI, early MCI; LMCI, late MCI; AD, Alzheimer’s disease; HV, hippocampal volume; FDG-PET, ¹⁸F-fluorodeoxyglucose-positron emission tomography; T-tau, cerebrospinal fluid total tau. MRI- indicate MRI HV negative; MRI + indicates MRI HV positive; CSF- indicate CSF T-tau negative; CSF + indicates CSF T-tau positive; PET- indicate FDG-PET negative; PET + indicates FDG-PET positive.

The results from the Kaplan-Meier analysis of N positive versus N negative and A–N–versus A–N+ versus A + N–versus A + N+ were shown in Fig. 4. Controlling for age, gender, APOE ɛ4 status, MMSE scores, and years of education at baseline, Cox proportional-hazards models were conducted to access the conversion risk. The corresponding hazard ratios were given in Supplementary Table 6. Using three N biomarkers (N1, HV; N2, FDG-PET; N3, CSF T-tau) to define N, we found that all three N + subgroups had a greater conversion rate than the corresponding N- subgroups (p < 0.0001, Fig. 4, Supplementary Table 6). The FDG-PET positive subgroup was more likely to progress than FDG-PET negative subgroup with an HR of 3.45 (95% CI = 2.50–4.77, Fig. 4a), which is greater than those of HV (HR = 2.59, 95% CI = 1.95–3.43, Fig. 4b) and T-tau (HR = 2.24, 95% CI = 1.66–3.01, Fig. 3c). When we added the A biomarker and divided participants into four subgroups (A–N–, A–N+, A + N–, and A + N+), the HRs derived for HV was bigger than FDG-PET and T-tau (p < 0.0001 HR = 3.15, 95% CI = 1.76–5.64). A + N2 + (using FDG-PET to define N2) subjects showed a 7.05-fold risk of cognitive decline compared with A–N2–individuals. We also assessed the conversion risk of cognitive impairment (Event: MMSE score decline more than 3 points) (see Supplementary Figures 6 and 7). Similarly, the results were in accordance with conversion risk of clinical progression. There was a trend that FDG-PET positive group had a faster rate of clinical progression, but when the A marker was added, HV was a better predictor for clinical progression.

OPEN IN VIEWER

Fig. 4

Kaplan-Meier curves showing cumulative probability of clinical disease progression. The comparisons of cumulative probability of clinical progression in (A) N1- versus N1+, (B) N2- versus N2+, (C) N3- versus N3+, and (D, E, F) add Amyloid status. The numbers of subjects at different time points were presented. N1, MRI Hippocampal volume; N2, FDG-PET; N3, CSF total tau; N-, neurodegeneration marker normal; N+, neurodegeneration marker abnormal; A–, amyloid normal; A+, amyloid abnormal.

In addition, we also compared the abilities of six N + markers (N1+, HV positive; N2+, FDG-PET positive; N3+, T-tau positive; N4+, plasma NFL positive; N5+, CSF NFL positive; N6+, Ng positive) to predict future cognitive impairment in Cox regression models (Supplementary Figure 7). No significant intergroup differences were detected among N positive subjects.