BACKGROUND: Low cerebrospinal fluid (CSF) amyloid-β 1-42 (Aβ 1-42) concentrations indicate amyloid plaque accumulation in the brain, a pathological hallmark of Alzheimer disease (AD). Innotest assay values of Aβ 1-42 have gradually increased over the past 2 decades, which might lead to misclassification of AD when a single cutpoint for abnormality is used. We propose an unbiased approach to statistically correct for drift.
METHODS: We determined year-specific cutpoints with Gaussian mixture modeling, based on the cross-section of bimodal distributions of Aβ 1-42 concentrations in 4397 memory clinic patients. This allowed us to realign year-specific cutpoints as an unbiased method to remove drift from the data. Sensitivity and specificity to detect AD dementia were compared between corrected and uncorrected values.
RESULTS: Aβ 1-42 values increased 22 pg/mL annually, and this could not be explained by changes in cohort composition. Our approach removed time dependencies [β (SE) = 0.07 (0.59);P= 0.91]. Statistically correcting for drift improved the sensitivity to detect AD dementia to 0.90 (95% CI, 0.89-0.92) from at least 0.66 (95% CI, 0.64-0.69) based on uncorrected data. Specificity became lower (0.69; 95% CI, 0.67-0.70) vs at most 0.80 (95% CI, 0.79-0.82) for uncorrected data.
CONCLUSIONS: This approach may also be useful to standardize Aβ 1-42 CSF concentrations across different centers and/or platforms, and to optimize use of CSF biomarker data collected over a long period.