This analysis utilizes data from three cohorts participating in the NTK project, which were selected to provide data spanning the entire AD continuum. The ALFA+ study (NCT02485730) aimed to characterize preclinical AD in CU individuals, most with a family history of AD (n=398) [8]. The Wisconsin cohort (n=651) comprised several longitudinal studies that utilized the same preanalytic protocol and included CU individuals, participants with MCI, or AD-dementia, enriched for parental history of AD [10]. The Abby/Blaze cohort (n=164) comprised participants in the ABBY (NCT01343966) and the BLAZE (NCT01397578) studies for patients with mild to moderate AD-dementia [11, 12]. Full eligibility criteria for each of the respective cohorts are described in the Supplementary Methods. All cohorts in the present analysis excluded participants who had comorbidities that would affect cognition. Some medications that affected cognition, such as sleep aids, were permitted in the Wisconsin cohort.

For the purposes of this analysis, the correction reference group for each cohort was defined as participants who were CU, APOE-ε4 allele non-carriers, and aged <65 years. As the Abby/Blaze cohort only included participants with AD-dementia, a correction reference group could not be defined.

CSF biomarkers included chitinase-3-like protein-1 (YKL40), soluble triggering receptor expressed on myeloid cells 2 (sTREM2), glial fibrillary acidic protein (GFAP), interleukin (IL)-6, neurofilament light (NfL), neurogranin, S100, alpha-synuclein (α-Syn), amyloid-β1–40 (Aβ40), Aβ42, pTau, and tTau. CSF biomarker samples obtained at baseline/enrollment were included. All biomarkers were measured using the NTK panel of immunoassays, which currently includes the commercially available Elecsys β-amyloid (1–42) CSF, Elecsys total Tau CSF, and Elecsys phospho-Tau (181P) CSF immunoassays, and robust prototype assays for the nine remaining biomarkers. Biomarkers Aβ42, Aβ40, pTau, tTau, s100, and IL-6 were measured using the cobas e 601 analyzer, and the remaining biomarkers were measured using the cobas e 411 analyzer (both Roche Diagnostics International Ltd).

The preanalytical procedures employed by each cohort are detailed in the Supplementary Materials. Sample collection within the Wisconsin cohort was initiated ahead of standardized preanalytical protocol dissemination [9]; therefore, the correction factors are calculated in the respective correction reference groups (participants who were CU, APOE-ε4 allele non-carriers, and aged <65 years) of the Wisconsin and ALFA+ cohorts assuming the ALFA+ cohort being the “standard cohort.” The correction factor was calculated using the formula:

$$\textrm\ \textrm=\textrm\left(\textrm+\textrm\right)/\textrm\left(\textrm\ \textrm\right)$$

Application of the correction factor was deemed successful in accounting for preanalytical variations by assessment of biomarker distribution overlap before and after correction, i.e., if following correction the biomarker distributions had good overlap, the correction was a success. The correction was applied to CSF biomarkers: α-Syn, Aβ40, and Aβ42 (Table S1), which are known to be significantly affected by preanalytical protocols, specifically related to the ability of these biomarkers to stick to the tubes used during testing [4, 13]. Conversely, the remaining biomarkers, such as pTau and tTau, appear to be unaffected by the tubes employed [4]. Natural variations between the cohorts were unaffected. Variations between cohorts may also result from inherent cohort differences or from cultural bias, e.g., in cognitive assessments.

Amyloid-β pathology was determined by CSF Aβ42/Aβ40 ratio for this analysis; the results are provided in the Supplementary Materials. To derive the cut-off values for the CSF Aβ42/Aβ40 ratio, Gaussian mixture modeling was independently applied to the ALFA+ and Wisconsin cohorts. The optimal number of Gaussians was set as two, after testing models with two, three, and four Gaussians (Supplementary Materials). Derived cut-off values were defined as x±2*s with differently defined parameters for x and s: (i) x=μ, s=σ (mu, sigma; Gaussian parameters of the amyloid-β negative [A−] population); (ii) x=mean, s=SD of samples assigned to A− population; and (iii) x=median, s=rSD of samples assigned to A− population (Tables S2–S5). The resulting cut-off values for the Aβ42/Aβ40 ratio were defined as 0.071 for the ALFA+ cohort [8] and 0.060 (0.075 after correction) for the Wisconsin cohort. Only patients with AD-dementia were included in the ABBY and BLAZE studies; therefore, cut-off values were not defined as this cohort was not divided by amyloid-β status. For comparison, the cut-off values determined for the ALFA+ cohort were applied to the Wisconsin cohort after correction.

All participants completed the MMSE [14] and Clinical Dementia Rating scale Sum of Boxes (CDR-SB) [15] cognitive assessments during the respective studies. The time between CSF biomarker collection at baseline/enrollment and cognitive assessment varied for each participant and in some cases was up to 1 year. For the longitudinal studies, the cognitive assessment closest to the first lumbar puncture was used in this analysis, including those cognitive assessments performed before the lumbar puncture. The ALFA+ cohort only included participants with CDR-SB=0, per the study exclusion criteria [16]. Calculations of a modified Preclinical Alzheimer Cognitive Composite (PACC) were based on methods proposed by Donohue et al. [17], Papp et al. [18], and Jonaitis et al. [19]. Variables included in the composite in the ALFA+ cohort were Semantic Fluency (animal naming), Free and Cued Selective Reminding Test with Total Immediate Recall, and Wechsler Adult Intelligence Scale-Revised Coding subtest. In the Wisconsin cohort, Semantic Fluency (animal naming), Rey Auditory Verbal Learning Test Trials 1–5 Sum, and Wechsler Adult Intelligence Scale-Revised Coding subtest were included. PACC was not used for the Abby/Blaze cohort.

To compare biomarker concentrations across cohorts, the median concentration and interquartile range of all NTK CSF biomarkers before and after correction were calculated for all cohorts, therefore enabling the inclusion of outlying samples. The robust-to-outliers standard deviation (rSD) was estimated based on percentile values (rSD=[value of 84.13% percentile − value of 15.87% percentile]/2). The distributions of the CSF biomarker concentrations within the same disease stage across the cohorts were statistically compared before and after correction. To compare baseline/enrollment values for each CSF biomarker for both CU individuals and patients with AD-dementia separately, correlation values were computed using Spearman’s rho.

Fold change was calculated using the canonical fold change calculation in CSF biomarker concentrations in CU A− individuals from either the ALFA+ or Wisconsin cohorts with (i) CU amyloid-β-positive (A+) individuals (ALFA+), (ii) CU A+ individuals (Wisconsin), (iii) patients with MCI A+ (Wisconsin), (iv) patients with AD-dementia (Wisconsin), and (v) patients with AD-dementia (Abby/Blaze). Receiver operating characteristic (ROC) analyses, presented with an area under the curve (AUC) and 95% confidence intervals, comparing CSF biomarker concentrations in CU A− individuals with (i) CU A+ individuals (ALFA+), (ii) CU A+ individuals (Wisconsin), (iii) patients with MCI A+ (Wisconsin), and (iv) and patients with AD-dementia (Wisconsin) were performed.

Spearman’s rho correlation between the concentration of all the biomarkers and cognitive performance, reflected in MMSE and/or PACC scores, in the different disease stages was computed. To assess cognitive scores from each cohort on a similar scale relative to that cohort’s control participants, standardization of each individual raw score into z-scores was performed using the means and rSDs obtained from the A− samples of each control group as a reference. All three obtained z-scores were averaged. The obtained PACC values were re-standardized using the mean and rSD from the A− samples of the control group. Missing data in any of the raw scores led to a missing PACC value.