Biological variation estimates of Alzheimer's disease plasma biomarkers in healthy individuals

Abstract INTRODUCTION Blood biomarkers have proven useful in Alzheimer's disease (AD) research. However, little is known about their biological variation (BV), which improves the interpretation of individual‐level data. METHODS We measured plasma amyloid beta (Aβ42, Aβ40), phosphorylated tau (p‐tau181, p‐tau217, p‐tau231), glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL) in plasma samples collected weekly over 10 weeks from 20 participants aged 40 to 60 years from the European Biological Variation Study. We estimated within‐ (CVI) and between‐subject (CVG) BV, analytical variation, and reference change values (RCV). RESULTS Biomarkers presented considerable variability in CVI and CVG. Aβ42/Aβ40 had the lowest CVI (≈ 3%) and p‐tau181 the highest (≈ 16%), while others ranged from 6% to 10%. Most RCVs ranged from 20% to 30% (decrease) and 25% to 40% (increase). DISCUSSION BV estimates for AD plasma biomarkers can potentially refine their clinical and research interpretation. RCVs might be useful for detecting significant changes between serial measurements when monitoring early disease progression or interventions. Highlights Plasma amyloid beta (Aβ42/Aβ40) presents the lowest between‐ and within‐subject biological variation, but also changes the least in Alzheimer's disease (AD) patients versus controls. Plasma phosphorylated tau variants significantly vary in their within‐subject biological variation, but their substantial fold‐changes in AD likely limits the impact of their variability. Plasma neurofilament light chain and glial fibrillary acidic protein demonstrate high between‐subject variation, the impact of which will depend on clinical context. Reference change values can potentially be useful in monitoring early disease progression and the safety/efficacy of interventions on an individual level. Serial sampling revealed that unexpectedly high values in heathy individuals can be observed, which urges caution when interpreting AD plasma biomarkers based on a single test result.


Highlights
• Plasma amyloid beta (Aβ42/Aβ40) presents the lowest between-and within-subject biological variation, but also changes the least in Alzheimer's disease (AD) patients versus controls.
• Plasma phosphorylated tau variants significantly vary in their within-subject biological variation, but their substantial fold-changes in AD likely limits the impact of their variability.
• Plasma neurofilament light chain and glial fibrillary acidic protein demonstrate high between-subject variation, the impact of which will depend on clinical context.
• Reference change values can potentially be useful in monitoring early disease progression and the safety/efficacy of interventions on an individual level.
• Serial sampling revealed that unexpectedly high values in heathy individuals can be observed, which urges caution when interpreting AD plasma biomarkers based on a single test result.

BACKGROUND
Novel technologies to measure brain pathophysiological processes in the blood have revolutionized the Alzheimer's disease (AD) research landscape. 1,2Established and highly accurate methods for tracking such processes face barriers to their large-scale implementation, such as the high costs, radiation exposure, and limited availability of positron emission tomography (PET) scans, as well as the relative invasiveness of lumbar punctures, required for measuring AD biomarkers in the cerebrospinal fluid (CSF). 3Blood-based AD biomarkers have demonstrated great promise so far, and are particularly promising for scalable implementation due to their minimally invasive and cost-effective nature. 1,23][24][25][26] Furthermore, all of these biomarker candidates have been evaluated as potential surrogate endpoints in disease-modifying clinical trials in AD, with a recent example being reductions in plasma p-tau217 as early as after 12 weeks of treatment with a promising anti-Aβ monoclonal antibody, donanemab. 27vertheless, several research questions must be addressed before large-scale implementation of blood-based AD biomarkers. 28While most studies have focused on their diagnostic and prognostic properties, little is known about their biological variation (BV), a foundational concept in clinical chemistry, crucial to ensure the safe implementation of diagnostic markers and to minimize misclassification risks in laboratory medicine. 29BV refers to the variation observed in clinical laboratory measurements determined by patients' physiology, and a strict guideline-defined methodology must be followed by BV studies to ensure robust results. 30,31Such studies require the serial, tightly controlled collection of samples from healthy individuals with a regular sampling rate, and that analytes should be quantified, at least, in duplicate. 30,31The key BV components are the within-subject biological variation (CV I ), which informs how much the concentration of a biomarker fluctuates around each individual's homeostatic setpoint, and the between-subject biological variation (CV G ), which informs on the variability between the homeostatic setpoints between different individuals.These parameters, alongside known assay-dependent analytical variation (CV A ), can provide highly clinically useful information for biomarker implementation.These include the reference change value (RCV), 32,33 which enumerates the change needed between consecutive measurements to exceed biological and analytical variation; the analytical performance specifications (APS) that clinical-grade assays should meet; 34,35 and the index of individuality (II), which evaluates the utility of population based reference intervals. 33,36  European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group on Biological Variation. 37,38[40][41][42] 2 METHODS

Study participants and sample collection
In this study, we quantified biomarkers in plasma-citrate samples from a subset of 20 individuals aged between 40 and 60 years within the EuBIVAS, 37,38  protocol used in EuBIVAS have been previously reported. 37While there was a considerable gap between sample collection and biochemical analyses (≈ 7 years), the evaluated AD blood biomarkers are known to be stable under the storage conditions used for the samples tested in this project. 4,5,18,23e protocol for EuBIVAS received approval from the institutional ethical review board of San Raffaele Hospital, in compliance with the World Medical Association Declaration of Helsinki, as well as the ethical board/regional ethics committee for each participating center.

Biomarker quantification
Biomarker quantification was conducted using single molecule array (Simoa) HD-X Analyzers from Quanterix at the Clinical Neurochemistry Laboratory of Sahlgrenska University Hospital in Sweden.A commercially available assay (Quanterix Neurology-4 Plex E) was used to simultaneously quantify for Aβ42, Aβ40 (and Aβ42/Aβ40, consequently), NfL, and GFAP. 27P-tau231 and p-tau181 were analyzed using Simoa assays developed at the University of Gothenburg, which have been validated as described elsewhere. 5,6To measure p-tau217, a novel commercially available assay from ALZpath (ALZpathDX) was used. 43All samples from the same participant were analyzed in the same analytical run, and each sample was quantified in duplicate.Internal quality controls (iQC) at three different concentrations, for each measurand, were analyzed in duplicate in the beginning and end of each run.Before analysis, blood samples were thawed, vortexed, and centrifuged at 4000 × g for 10 minutes as suggested in recent studies. 44,45

Statistical analysis
Our statistical analyses followed a series of well-established and guideline-defined steps for deriving BV data, as set out by the Biological Variation Data Critical Appraisal Checklist (BIVAC), a standard for the executing and reporting of BV studies. 31Outlier detection procedures were performed on three levels, including analytical (between replicates), within-subject (among 10 collections for CV I calculation), 7][48][49][50] For obtaining CV I and CV A estimates, we initially performed CV-transformation of the data in which each person's data are "normalized" by dividing by that person's mean value, so as to later perform analysis of variance (ANOVA) on these CV-transformed values. 51After CV transformation, we performed outlier identification and removal on the analytical levels (between replicates) by assessing the homogeneity of CV A with the Bartlett test.In case of heterogeneity for the analytical component, we first excluded the replicate value of the measurement that most deviated from that participant's mean.If the heterogeneity persisted, we then also excluded the second measurement result of the time point showing abnormal analytical variation.After ensuring analytical homogeneity, we evaluated the presence of outliers on the within-individual variation level by assessing the homogeneity of the within-individual CV I with the Cochran test.Then, we evaluated for each biomarker whether the results were consistent with steady state (i.e., no trend for increase or decrease during study) by fitting a linear regression model with the mean blood drawing value (pooled mean of the duplicate concentration measurements of each participant) as the dependent variable, with blood drawing number (from 1 to 10) as the independent variable.Individuals were considered in a steady state if the 95% confidence interval (CI) of the blood draw term (i.e., the slope) included 0. 52 Finally, the CV I was estimated with CV-ANOVA, the "Røraas method," a validated and recommended ANOVA method for estimating CV I and CV A . 51,53To calculate the between-subject biological variation (CV G ), we first applied the Dixon Q test to detect outliers in mean biomarker concentrations between subjects, and the Shapiro-Wilk test to verify the normality assumption on mean concentrations.7][48][49][50] First, CV I and CV G estimates were calculated for the whole study population, and also secondarily separately for males and females, for all measurands.
Confidence intervals for BV estimates were calculated as previously described, 54 and the lack of overlap of the 95% CI of estimates was used to indicate significant differences between subgroups.
Other relevant metrics were computed based on the abovementioned BV estimates calculated as follows.Desirable APS were calculated for imprecision (CV APS = 0.5 x CV I ) and for bias ).The RCV was calculated at a 95% bidirectional alpha (z = 1.65) as RCV = 100*(exp[± z x √2 x σ]-1), where , with σ 2 CVi = ln(CVi 2 + 1) and σ 2 CVa = ln(CVa 2 + 1).The II was calculated as the ratio of CV I and CV G for each biomarker, indicating whether population-based reference intervals can be useful for evaluating results. 33,36We also calculated the number of samples needed to be collected to estimate an individual's homeostatic point (NHSP) with a "D" absolute percentage proximity to the individual's true value with the equation n Note: The table summarizes key demographic information for the included participants.Data are described as mean (SD) or n (%).Abbreviations: BMI, body mass index; SD, standard deviation.

Participant characteristics
We included data analyzed from a total of 196 plasma samples, collected weekly over 10 weeks from 20 participants, with a mean number of 9.8 samples per participant.Key demographic information is described in Table 1.The age range of the included participants was   Note: The table displays the overall number of samples included and biomarker results produced, as well as the results of the homogeneity analyses carried to detect the presence of outliers on the replicate, sample, and subject levels.Abbreviations: Aβ, amyloid beta; CV I , within-individual biological variation; GFAP, glial fibrillary acidic protein; NfL, neurofilament light; p-tau, phosphorylated tau.surement were identified for Aβ40, Aβ42/Aβ40, GFAP, and NfL, while a few replicates were excluded for Aβ42, p-tau181, p-tau217, and p-tau231.When assessing the variance homogeneity for within-subject variation, outlier time points were identified for all biomarkers, but no subject had to be fully excluded.For the total study population, the mean percentage of results identified as outliers at the homogeneity analyses was 3.56% (range, 1.0% to 6.8%), which left a mean of 369 results (range, 332 to 381) used per biomarker to estimate the CV I .

Homogeneity analyses and outliers
In Figure 1 the 10-week biological variation, in concentrations, of each plasma biomarker, stratified by sex and ordered by increasing age, is displayed.In a separate outlier detection procedure before the CV G estimation, the Dixon Q test identified one outlier subject for NfL, and two outlier subjects for Aβ42/Aβ40 (indicated in Figure 1).No trend was identified for any of the included biomarkers in the overall study population or in male or female subgroups.No biomarker measurement for any analyte was below the lower limit of detection or the lower limit of quantification.

Between-subject biological variation (CV G )
Figure  3 indicates the CV G s separately for males and females.No differences in CV G estimates were found for the evaluated biomarkers, with overlapping 95% CIs between males and females for all measurands.Table 3 also shows the mean concentrations and their 95% CIs for males and females separately, with slightly higher concentrations in females observed for plasma GFAP and NfL, and slightly higher concentrations in males for plasma p-tau181 and p-tau231.

DISCUSSION
We report BV estimates for AD plasma biomarkers generated based on a high number of weekly samples per individual, for a comprehensive biomarker panel measured within the same participants.6][57] We found that within-and between-subject biological variation can be considerably different for AD biomarker classes, which may impact biomarkers differently according to each application context.
Beyond improving the interpretation of laboratory tests, reliable BV data enable the determination of the APS needed for each biomarker.
Assay imprecision (i.e., CV A ) should be considerably lower than the biomarker's CV I , with the desirable analytical performance being that CV A ≤ CV I /2. 58Here, CV A s were slightly higher than desired for most biomarkers.Plasma Aβ42 and Aβ40 (but not Aβ42/Aβ40) and p-tau181 were within the desirable range, with p-tau217 showing a very close to desirable analytical performance (CV A = 5.7%; CV APS ≤ 5.2%).
Nevertheless, it is important to emphasize that BV estimates are not standalone criteria to determine APS for assays, but rather a complementary tool to refine the determination of analytical goals given

Note:
The table displays the biological variation estimates for each biomarker and mean concentrations for the whole participant population in sex-stratified sub-groups and their 95% CIs.Abbreviations: Aβ, amyloid beta; CI, confidence interval; CV A , analytical variation; CV G , between-individual biological variation; CV I , within-individual biological variation; GFAP, glial fibrillary acidic protein; NfL, neurofilament light; p-tau, phosphorylated tau. each analyte's clinical application context.For instance, the higherthan-desirable observed CV A s are likely not a cause of concern in light of the main clinical applications of NfL and p-tau variants, whereas Aβ42/Aβ40 might be more affected, as discussed below.
Plasma p-tau is an AD-specific biomarker envisioned to be implemented as a screening tool to classify patients seeking medical advice for cognitive symptoms into high, intermediate, and low risk of having AD pathology. 59,60In our study, CV A was remarkably similar for all three plasma p-tau variants, but they demonstrated considerably different CV I s.Interestingly, p-tau231 demonstrated the lowest CV I (6.3%, 95% CI: 5.5-7.4%),followed by p-tau217 (10.3%, 95% CI: 9.2-11.7%)and p-tau181 (16.7%, 95% CI: 12.5%-23.6%),and this could suggest possible differences in release, clearance, or transportation of plasma p-tau species.Recent head-to-head comparisons of plasma ptau variants showed the most promising candidates were increased between 100% and 360% in the presence of AD pathology. 9,10Considering the magnitude of these increases, currently available p-tau assays demonstrate satisfactory analytical performance for clinical applica-tions, making their diagnostic ability less vulnerable to biological and analytical variation.
Plasma NfL has been successfully introduced in some clinical routine laboratories, being useful in a range of neurodegenerative diseases and acute neurological conditions. 22,24,26,61Plasma NfL showed a relatively low CV I (7.4%, 95% CI: 6.4-8.5%), and a higher CV G (21.2%, 95% CI: 15.1-29.8%).This is in accordance with the ≈ 10% CV I and CV G s reported in a previous BV study evaluating serum NfL in a Turkish cohort with 10 weekly collections. 57Our RCVs for NfL (increase: +30.7; decrease: −23.4%) also closely agreed with those in that study (increase: +32.7; decrease: −24.7%).A study in a Danish cohort reported a lower CV I for serum NfL with non-overlapping CIs (CV I = 3%, 95% CI: 1.2-5.0%),and lower RCVs (increase: +24.3; decrease: −19.5%). 56These lower estimates could be attributed to the shorter sampling period (three consecutive days vs. 10 weeks), which may underestimate BV. 62 The relatively high CV G seen for plasma NfL  such as AD, where the diagnostic utility of NfL remains limited.However, in certain clinical scenarios, the magnitude of NfL increases is much larger compared to this higher between-subject variation, including differentiating primary psychiatric disorders from frontotemporal dementia, 63 or in prognostic evaluation of acute conditions such as cardiac arrest, stroke, and traumatic brain injury. 24,26,64ile it is not yet clear what GFAP in the blood reflects (showing unexpected differences against CSF GFAP), 65 it has been associated with Aβ pathology, showing promising diagnostic performance, 18,20 even though increased GFAP levels have been reported in Aβ-negative neurodegenerative conditions. 66We observed a relatively low CV I for plasma GFAP (8.0%, 95% CI: studies.The high between-individual variability of plasma GFAP, alongside its poorly understood clinical meaning, may pose difficulties for its individual-level interpretation. Plasma Aβ42/Aβ40 is a widely investigated biomarker associated with Aβ pathology currently in clinical use to support an AD diagnosis. 12,67,68[16] Our BV findings further support that plasma Aβ42/Aβ40 will likely face long-term difficulties if introduced in clinical practice.Plasma Aβ42/Aβ40 demonstrated a considerably low CV I (3.3%, 95% CI: 2.9-3.9), and a relatively low CV G (6.6%, 95% CI: 4.6-9.3).The low CV I introduces a very high demand on analytical performance, because, desirably, the CV A should be less than half of the CV I (CV APS = 1.7%;Table 4), and, optimally, less or equal to a quarter of CV I (CV I /4 = 0.7%).
In contrast, plasma p-tau presents larger disease-related increases, making the ≈ 5.5% CV A s acceptable.Additionally, while there are several different plasma Aβ assays currently in use, they present similarities in their biochemical target and in small disease changes. 13nsidering previous BV studies showing different assay versions for the same analyte present indistinguishable CV I s, it is expected that the CV I for plasma Aβ42/Aβ40 would be similar across assays. 47,52Further, the desirable bias is also considerably low for plasma Aβ42/Aβ40 (B APS = 3.7%; Table 4), and it is unlikely that batch-to-batch variations could be kept low enough to meet this APS.
We report RCVs for AD plasma biomarkers, which can be potentially clinically valuable when monitoring individuals over time by enumerating the change that can be explained by biological and analytical variation, becoming especially relevant with novel anti-Aβ immunotherapies such as lecanemab and donanemab. 27,69These drugs substantially reduced plasma p-tau217 levels, on a group level, as early as at 12 weeks of treatment (with our RCVs derived in a similar time frame), during the initial phase of Aβ-plaque reduction. 27RCVs could be potentially used to identify whether a reduction in plasma p-tau217 after treatment initiation could indeed be related to a positive treatment response.On the other hand, this class of drugs can cause F I G U R E 3 Example of 10-week variability in plasma p-tau181 and p-tau217 in two study subjects.The figure shows the variability in plasma p-tau181 and p-tau217 levels over 10 weeks in two study subjects.Dots correspond to the mean concentration of the two duplicate measurements, and all of these subjects' measurements demonstrated acceptable agreement between replicates, indicating that any deviation observed does not come from analytical imprecision.Dashed lines represent previously published cut-offs for plasma p-tau181 and p-tau217 as illustrative examples of the potential impacts of these fluctuations over decision making.Of note, the outlier data points were excluded from CV I calculations in the homogeneity analyses as they do not reflect the expected homeostatic fluctuation.It is likely that these outlier data points instead correspond to a yet unknown factor that affects biomarker readings.CV I , within-individual biological variation; p-tau, phosphorylated tau.
amyloid-related imaging abnormalities (ARIA) of the hemorrhage or edema type, which can be very harmful. 70If NfL proves capable of tracking such changes, RCVs could also be potentially useful to monitor ARIA emergence.However, it is important to consider that our RCVs were obtained using the presumably healthy cognitively unimpaired sample and assays herein described, and, for this reason, cannot be considered universal values, and each laboratory has to determine their own RCVs based on their CV A estimates.
Our study provides a unique opportunity to evaluate shorter term fluctuations of AD blood biomarkers.Most published studies have collected samples either cross-sectionally or over longer periods of time (e.g., 6 months, yearly), and little is known about their shorter term variability.Figure 3 shows plasma p-tau181 and p-tau217 for two male subjects with similar ages.Subject "A" shows minimal fluctuation around the homeostatic point, while subject "B" experiences two spikes in plasma p-tau.These fluctuations were not analytical outliers, and because a chronic disease like AD is unlikely to manifest an oscillatory progression from week to week in middle-aged adults, there might be yet uncharacterized factors influencing biomarker readings.
Clinical decisions made on a single sample collected on a p-tau "spike" day could erroneously classify patients as "abnormal," as exemplified by two previously described cut-offs. 43,71Such high-value outliers are not uncommon in Aβ-negative groups when examining data points from recent cross-sectional studies, [4][5][6] and we recommend caution for researchers and clinicians when interpreting AD blood biomarker results from a single sample.While this has also happened to the other biomarkers in our study, we chose to highlight this phenomenon for p-tau to demonstrate that even the most promising biomarker classes may be subject to unexpected variations that need to be characterized.
Additionally, knowledge of CV I enables calculating the number of samples needed to estimate the individual's homeostatic point (HSP) within a certain proximity of the true value.To estimate the true HSPs of all analytes with a deviation of ± 20% (α < 0.05), a reasonable margin for most analytes here (given clinical contexts discussed above), one sample suffices for all biomarkers but p-tau181 (NHSP = 3).Reducing this deviation to 5% (likely needed for plasma Aβ42/Aβ40 but not for others), three samples would be required (Table 4).The II evaluates the utility of reference intervals (RIs).For analytes with pronounced individuality and a relatively low CV I compared to CV G (II < 0.6), RCVs are more useful than RIs for accurate interpretation of sequential results, with each individual serving as the optimal reference point for assessing serial results.However, RIs remain suitable for analytes with high II (particularly when II > 1.4). 33,36Here, all II values were below 0.6, except for p-tau181 and Aβ42/Aβ40, which had slightly higher values, indicating marked individuality for these analytes. 33,36,72is study has limitations.Although well powered in individuallevel serial sampling, the number of participants was relatively small, possibly affecting CV G more than CV I estimates.We found some concentration differences between males and females, and an unexpected sex difference in CV I for p-tau181, warranting further studies.The relatively younger population studied here may not capture biomarker fluctuations related to factors such as co-morbidities and medication use. 73Lack of confirmatory CSF or PET biomarkers prevented us from evaluating the effects of AD pathology over BV estimates.Further evaluation on ethnically diverse populations is also needed.The use of citrate plasma does not affect the obtained CV values, but biomarker concentrations (such as those in Figure 1) should be interpreted bearing in mind the matrix type and collection tube.For most of the evaluated AD biomarkers, citrate plasma has shown very similar concentrations compared to paired ethylenediaminetetraacetic acid, and slightly lower levels in citrate plasma for NfL and GFAP. 44Our study has a number of strengths, involving quantifying a comprehensive panel of AD biomarkers in a dataset following all EFLM recommendations for BV studies, 30,31 which has generated reliable BV data for many other analytes. 37,38

3 .
Future directions: Future studies should evaluate the potential clinical impact of the application of BV knowledge upon clinical and research interpretation of AD plasma biomarkers, especially in disease monitoring and in the evaluation of safety and efficacy of novel therapeutic interventions.

40 to 60
years, with a mean (standard deviation [SD]) age of 46.4 years (6.20) for the whole study population.Half of the participants were female, and key demographic characteristics were generally similar between sexes.The study population came from five centers in four European countries (Italy [n = 7], Netherlands [n = 5], Norway [n = 5], Spain [n = 3]), and all participants were White.Participants were healthy, with a mean (SD) body mass index of 23.3 kg/m 2 (2.85 kg/m 2 ), did not have hypertension, and the majority (55%) engaged in physical activity for more than 3 hours per week.Only one participant was a smoker (5%), and 11 reported consuming 1 to 2 units of alcohol per week.

(
reflecting higher inter-individual variability) may pose a challenge to its clinical interpretation in conditions with modest NfL fold changes TA B L E 4 Metrics derived from BV estimates.

Table 2
Homogeneity analyses and number of results included for calculation of biological variation estimates.
displays results of the homogeneity analyses for outlier detection and the final number of results included for each of the biomarkers.TA B L E 2 Biological variation estimates for the whole study population and according to sex.
TA B L E 3 The table displays each biomarker's desirable analytical performance specifications (APS) for replicate precision and bias.Also, the table displays the reference change values (RCV), which indicate the percentage change needed between two consecutive measurements so that such an increase or decrease significantly overcomes analytical and biological variation.The index of individuality and number of samples needed to estimate the homeostatic point with a given proximity are also shown.