Unbiased estimates of cerebrospinal fluid β-amyloid 1–42 cutoffs in a large memory clinic population

We selected 2462 subjects from the Amsterdam Dementia Cohort (ADC) [12] with subjective cognitive decline (SCD; n = 448), mild cognitive impairment (MCI; n = 490), AD dementia (n = 1031), and dementia other than AD (n = 493) who had CSF measurements available at the time of their first visit at the memory clinic between August 1997 and July 2015. All patients underwent standardized dementia screening at baseline, including physical and neurological examinations, electroencephalograms, magnetic resonance imaging, and laboratory tests. Cognitive screening included the Mini Mental State Examination and, in over 90% of the subjects, a comprehensive neuropsychological test battery. Diagnoses were made by consensus among a multidisciplinary team that did not have knowledge of the CSF results, and they were based on the following clinical criteria. Subjects were diagnosed with SCD when cognitive complaints were present but criteria for MCI, dementia, or any other neurological or psychiatric disorders were not met and all other examinations were normal [13]. Subjects were diagnosed with MCI according to the established MCI criteria [14]. Subjects with AD-type dementia were diagnosed according to the criteria of the National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association [4, 15]. Subjects with non-AD-type dementia included subjects with behavioral variants of frontotemporal dementia (n = 204) [16, 17], dementia with Lewy bodies (n = 113) [18], vascular dementia (n = 66) [19], corticobasal degeneration or syndrome (n = 31 or n = 10, respectively) [20], progressive supranuclear palsy (n = 44) [21], alcohol-related dementia (n = 3), Huntington’s disease (n = 3), Parkinson’s disease (n = 1), normal pressure hydrocephalus (n = 3), CADASIL (cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy; n = 1), Creutzfeldt-Jakob disease (n = 3), tauopathy (n = 2), and dementia without a known cause (n = 9). All subjects gave written informed consent for the use of their clinical and biomarker data for research purposes, and the ethical review board of the VU University Medical Center approved the study.

Subjects were followed according to clinical needs. The standard follow-up procedure included a 6-month follow-up examination for subjects with dementia and a 12-month follow-up visit for subjects without dementia [12]. Neuropsychological tests were repeated every 12 months. Diagnoses at follow-up were made on the basis of clinical criteria listed above by consensus among a multidisciplinary team.

CSF was collected from 67% of the subjects in the ADC [12]. Reasons for not collecting CSF were refusal, contraindications, technical failure, or CSF having been collected elsewhere. CSF was obtained by lumbar puncture using a 25-gauge needle with a syringe into 10-ml polypropylene tubes (Sarstedt, Nümbrecht, Germany). Within 2 h, CSF samples were centrifuged at 1800 × g for 10 minutes at 4 °C. The CSF supernatant was transferred to new polypropylene tubes and stored at −20 °C until further analysis (within 2 months). Baseline Aβ_1–42 was measured using a commercially available enzyme-linked immunosorbent assay (Innotest β-amyloid(1-₄₂; Innogenetics, Ghent, Belgium) on a routine basis as described before [11]. The intra-assay coefficient of variation (mean ± SD) for Aβ_1–42 was 2.0 ± 0.5%, calculated by averaging the coefficient of variation of duplicates from five runs randomly selected over 2 years. The inter-assay coefficient of variation (mean ± SD) was 10.9 ± 1.8%, as analyzed in a high and low pool from 13 consecutive pool preparations used in total in 189 runs. The team performing the CSF analysis was unaware of the clinical diagnoses.

Baseline characteristics were compared between diagnostic groups with analysis of variance and Kruskal-Wallis or chi-square tests, where appropriate, using IBM SPSS Statistics version 20.0 software (IBM, Armonk, NY, USA). A difference with a p value less than 0.05 was considered significant. Gaussian mixture modeling was used to define a cutoff for abnormal CSF Aβ₄₂ using the R statistical software program version 3.2.1 mixtools package. First, the number of distributions that best described the data was determined with the R boot.comp function. Next, we defined a data-driven cutoff as the point where the lines of two fitted normal distributions crossed each other. The main analyses included all subjects. We repeated mixture modeling in subgroups based on diagnosis, age (dichotomized based on the median age of 66.5 years), and APOE ε4 allele carriership. Bootstrap sampling was used to determine 95% CIs of the cutoff. Cutoffs were considered to be statistically different between subgroups when their 95% CIs did not overlap.

New cutoff values were compared with our previously clinically defined cutoff of 550 pg/ml, which was based on subjects seen between 2001 and 2007 (n = 1070) [11]. For this comparison, we repeated mixture modeling in a subset of data including subjects from this time period. Using this subset, we further tested differences between the clinical and new cutoff values in discrimination between nondemented subjects with or without AD-type dementia at follow-up. In addition, using Cox proportional hazard models, we compared the association of the old cutoff and new cutoff with time to AD-type dementia progression, including age and sex as covariates. For these analyses, nondemented patients were included when they had at least 6 months of follow-up available. We compared both models with chi-square tests of the log-likelihood ratio. A difference with a p value less than 0.05 was considered significant. Statistical analyses for multivariate Cox regression were performed using R version 3.2.3 software.

Finally, we studied the minimum number of subjects per clinical population necessary to reliably estimate the cutoff in a data-driven way. To this end, we simulated CSF Aβ₄₂ values from a bimodal distribution with mean and SD values as estimated from our dataset. We recalculated the cutoffs and 95% CIs for sample sizes with varying numbers starting from n = 300 to 3000 with steps of 100. The minimum sample size required to obtain a reliable cutoff was determined as the sample size for which 95% CI lines were larger than the mean cutoff ±10%, which is currently used as a rule-of-thumb indication of acceptable variability in CSF Aβ₄₂ levels.

Table 1 shows the baseline characteristics according to diagnostic group. Briefly, patients with AD and patients with MCI were older and included a higher percentage of APOE ε4 allele carriers than the other groups. CSF Aβ₄₂ levels were highest for SCD, followed by non-AD-type dementia, MCI, and AD-type dementia. Of the nondemented subjects, over an average follow-up period of 3.2 (SD 2.04) years, 21 (9%) of the subjects with SCD progressed to MCI and 13 (5%) to AD-type dementia, and 146 (39%) of the subjects with MCI progressed to AD-type dementia.

In the total sample and all subgroups, a bimodal distribution best fitted the data (Table 2). In the total sample, this yielded a cutoff of 680 pg/ml (95% CI 660–705 pg/ml) (Fig. 1a). With this cutoff, 55% of our population fell into the abnormal amyloid distribution. Similar cutoffs were found when we repeated mixture modeling within the dementia group (694 pg/ml, 95% CI 670–721 pg/ml), the pooled sample of subjects with SCD and MCI (664 pg/ml [95% CI 621-712 pg/ml), subjects with SCD (621 pg/ml (95% CI 526-901 pg/ml) and subjects with MCI (696 pg/ml [95% CI 654-758 pg/ml) (Fig. 1b–e). A lower cutoff was found for subjects younger than 66.5 years (645 pg/ml [95% CI 617-678 pg/ml) than subjects older than 66.5 years (723 pg/ml [95% CI 691-762 pg/ml) (Fig. 1f and g). The cutoff for CSF Aβ₄₂ was higher in APOE ε4 carriers than in noncarriers (resp. 716 pg/ml [95% CI 684-756 pg/ml; 650 pg/ml [95% CI 611-689 pg/ml), but this did not reach statistical significance (Fig. 1h and i).

The cutoff of 680 pg/ml based on mixture modeling was substantially higher than our previously clinically defined cutoff of 550 pg/ml (95% CI 531–570 pg/ml) [11]. Repeated mixture modeling in the subset of subjects with CSF analysis in the 2001–2007 period (in which subjects were selected for the clinically based cutoff calculation) also resulted in a higher cutoff (615 pg/ml, 95% CI 573–673 pg/ml) (Fig. 2) than the clinically defined cutoff.

In this subset, using a cutoff value of 615 pg/ml, 439 subjects were classified as having abnormal amyloid, which was a 13% increase compared with the 390 subjects classified by the clinically defined cutoff. The sensitivity of the cutoff of 615 pg/ml for AD-type dementia was 0.89 with a specificity of 0.62. The clinically defined cutoff of 550 pg/ml resulted in a sensitivity of 0.86 with a specificity of 0.65. Of the nondemented subjects who later progressed to AD-type dementia, 87% had CSF Aβ₄₂ levels less than 615 pg/ml, compared with 76% with CSF Aβ₄₂ levels less than 550 pg/ml. For nondemented subjects who did not progress to AD-type dementia, these proportions were 30% versus 21%, respectively (Table 3). Survival analyses showed that both cutoffs were predictive of the time to development of AD-type dementia (550 pg/ml cutoff HR = 5.14, 95% CI 2.96–8.93; pg/ml; versus 615 pg/ml cutoff HR 7.44, 95% CI 3.74–14.79) (Table 4). The HR for the development of AD-type dementia was significantly greater for the new cutoff of 615 pg/ml than for the cutoff of 550 pg/ml (p < 0.001).

We further explored why the data-driven cutpoint was somewhat lower in the subset of subjects with CSF analysis in the 2001–2007 period than in the total sample. Additional file 1: Figure S1 shows that the peaks of CSF Aβ₄₂ level distributions seem to shift over subsequent years. This could not be explained by a difference in patient population, because the distribution of diagnoses remained comparable over time (χ² (45) = 61.32, p > 0.05) (Additional file 1: Table S1). We further explored whether this shift was due to an assay drift, and we repeated all subgroup analyses stratified for the time period when the lumbar puncture was obtained (2001–2007 versus 2008–2015) (see Additional file 1: Table S2 for baseline characteristics). Briefly, the cutpoint in the total group was higher in the 2008–2015 subsample (697 pg/ml [675–723 pg/ml]) than in the 2001–2007 subsample (615 pg/ml [573–673 pg/ml]) (Additional file 1: Table S3 and Figure S2 and S3). Subgroup analyses showed that the cutpoint for the dementia group and for APOE ε4 allele carriers was also higher in the 2008–2015 group than in the 2001–2007 group. Cutpoints for other subgroups did not differ between time periods.

Additional file 1: Figure S4 shows the mean and SD values for different sample sizes, varying from 300 to 3000 subjects with steps of 100. The average CSF Aβ₄₂ cutoff was 679 pg/ml and remained similar with increasing sample size, whereas the 95% CI became narrower. Accepting a maximum deviation from the cutoff of ±10%, a minimum sample size of 800 is required for an acceptable 95% CI of 637–744 pg/ml. Our determined cutoff of 680 pg/ml and 95% CI fell within this range.