Evaluation of a panel of plasma biomarkers for Alzheimer’s disease in a diverse research cohort

Abstract

Background:

Plasma biomarkers show significant promise for Alzheimer’s disease (AD) diagnostics and risk prediction, however, much less is known about how these assays perform in a diverse research cohort of older adults.

Objective:

To compare plasma biomarkers with clinical diagnoses and assess variability by demographic factors in a diverse research cohort.

Methods:

Among 331 University of Michigan Memory and Aging Project (UM-MAP) participants, plasma biomarkers (pTau-217, pTau-181, GFAP, NfL, Aβ₄₂, Aβ₄₀, t-Tau) were measured. Demographic information (age, sex, education, race) was self-reported. Clinical consensus phenotypes (dementia of the Alzheimer Type (DAT), mild cognitive impairment (MCI), cognitively unimpaired (CU) were based on neuropsychological assessments. Logistic regression with machine learning for model variable selection was used to compare participants by clinical phenotypes.

Results:

Comparing CU and DAT participants, areas under the curve (AUCs) from receiver operator characteristic curves of single biomarker models ranged from 0.74–0.89. Optimal performance (AUC 99.7) was observed from stepwise regression with backward selection, which identified pTau-217, GFAP, sex, education, APOE ϵ4 allele, and race as model variables. When comparing MCI and DAT participants, only pTau-217 differed significantly (AUC 0.80). pTau-181 and pTau-217 levels were higher in white participants than Black/African American participants across all clinical phenotypes.

Conclusions:

Plasma biomarkers demonstrate promise for improving diagnostic accuracy in diverse research cohorts. Incorporating demographic variables facilitates enhanced interpretability of biomarker levels and the development of reference ranges.

Introduction

Alzheimer’s disease (AD) is a debilitating neurodegenerative disorder defined neuropathologically by characteristic amyloid-β (Aβ) plaques and neurofibrillary tangle tau deposits,¹ the accumulation of which is clinically characterized by the deterioration of cognitive, functional, and behavioral abilities.² In 2016, a classification scheme was developed that allows for clinical staging (i.e., cognitively unimpaired, mild cognitive impairment (MCI) and dementia) that incorporated biomarkers. This scheme was adopted by the National Institute on Aging and Alzheimer’s Association in 2018^1,3 and is known as the Aβ/Tau/Neurodegeneration (A/T/(N)) framework. Biomarker-based confirmation of AD incorporates cerebrospinal fluid (CSF) and neuroimaging biomarkers from positron emission tomography (PET).³ However, limited access to lumbar punctures and PET imaging, methods on which this framework depends, creates a barrier to early diagnosis in many diverse community-based cohorts.⁴ Since establishing this framework, plasma-based biomarkers with excellent diagnostic performance were developed and are being clinically validated.⁵ Indeed, updated diagnostic criteria were proposed in 2024 that incorporate plasma biomarkers into criteria for disease diagnosis and staging.⁵ AD-related blood biomarkers hold promise for increasing the ease of biomarker-based diagnosis of cognitive disorders.

Multiple studies demonstrated that plasma levels of Aβ₄₂:Aβ₄₀, phosphorylated tau (pTau-181 and pTau-217), glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL) are associated with the neuropathological changes characteristic of AD.^6–12 Plasma Aβ levels were linked to Aβ plaque deposition in the AD brain using PET imaging.^13,14 Reduced levels of certain forms of plasma Aβ, such as the Aβ₄₂:Aβ₄₀ ratio, are associated with an increased risk of cognitive decline and progression to dementia of the Alzheimer’s type (DAT).^13,15,16 Levels of plasma pTau-181 and pTau-217 become abnormally elevated around the time individuals become Aβ PET positive and prior to increases in tau PET signal, suggesting that they may predict future PET imaging values.^4,17,18 Levels of plasma pTau-181 and pTau-217 increase with disease severity as demonstrated by Aβ and tau PET tracers and also correlate with cognitive decline.^19–22 Plasma pTau-217 routinely outperforms other biomarkers across analyses, even in diverse participant pools, and is currently considered a strong predictor of progression during early AD stages.^9,23–25 Levels of pTau-217 show high utility as a biomarker in CSF Aβ+ and/or Aβ PET + participants when compared to CSF Aβ− and/or Aβ PET− participants in clinical cohorts.^16,19,26 Plasma pTau-217 also correlates well with tau PET positivity and postmortem tau neuropathology.^27,28 Thus, it appears plasma pTau-217 concentration likely reflects underlying Aβ and tau pathologies and may enhance the diagnosis of AD across the clinical continuum.^23,26

Several other promising biomarkers have recently emerged. First, GFAP is a cytoskeletal protein found predominantly in astrocytes.^29,30 Traditionally, GFAP was used as a marker for astrocyte activation or injury in various neurological conditions.^31,32 More recent studies investigated the potential of GFAP as a blood biomarker for AD, since it is highly predictive of the presence of Aβ plaques, though it may be less predictive of tau aggregation.^33–36 NfL is another cytoskeletal protein released into the CSF and bloodstream following neuronal damage or degeneration.³⁷ Elevated levels of NfL and GFAP are associated with multiple neurodegenerative disorders, including AD.³⁶ Elevated NfL levels correlate with many features of AD, such as reduced hippocampal volume, impaired cognitive performance, and reduced fluorodeoxyglucose PET uptake.⁷ NfL may have value as a predictive biomarker in longitudinal studies.^{16,21,24,33,38,39} Biomarker combinations have demonstrated improved diagnostic and prognostic utility over single biomarker measurements. Fujirebio recently received the first United States Food and Drug Administration clearance for a blood-based in-vitro diagnostic test for patients being assessed for AD based on the ratio of pTau-217 and Aβ_42, a ratio that has proven effective for identifying Aβ PET status.^40,41

While development toward the prognostic and diagnostic use of plasma biomarkers holds promise, much of this work was performed in well-defined A/T/(N)-confirmed clinical research cohorts characterized by CSF biomarkers, neuroimaging biomarkers, and/or post-mortem diagnosis of participants. Additionally, many prior studies lack diversity in the participant cohort, enrolling a majority of non-Hispanic White (NHW) participants, and leaving the ability to extrapolate findings to racially diverse groups unknown.⁴² Indeed, some previous studies demonstrated race-associated differences in biomarker levels between different ethnoracial groups, though the reasons for these differences remained unclear.^43–46 Broad agreement exists that biomarkers need to be tested in diverse cohorts to better understand race-associated factors that affect plasma biomarker levels.^17,23,42,47

Additionally, there is a need to understand how common comorbidities may affect plasma biomarker levels.^48,49 Individuals being seen in community-based clinics often experience higher levels of medical comorbidities like chronic kidney disease (CKD), hypertension, diabetes, obesity, depression, and cancer than would be found in the study participants included in the development and testing of biomarkers.^8,42 CKD increases plasma levels of both pTau-217 and pTau-181, thus affecting the reliability and validity of these biomarkers.⁵⁰ Similarly, dyslipidemia and hypertension are associated with higher levels of CSF and plasma Aβ₄₀, Aβ₄₂, and total and phosphorylated tau, potentially due to increased diffusion of these proteins at a compromised blood-brain barrier.^48,51–53 Obesity, which is defined by body mass index (BMI), is not correlated with changes in CSF biomarkers but is negatively correlated with many blood biomarkers of neurodegeneration.^54–56 Many clinical research cohorts exclude people with comorbidities, thus, more research in such populations is needed to analyze the diagnostic and prognostic effectiveness of blood biomarkers of neurodegeneration in the presence of common comorbidities.

The University of Michigan Memory and Aging Project (UM-MAP) is a prospective observational, longitudinal study, enrolling older adults across the cognitive spectrum, and particularly from the racially diverse metro Detroit area.⁵⁷ The UM-MAP cohort is part of the National Institute on Aging-funded P30 Michigan Alzheimer’s Disease Research Center (MADRC). Study participants engage in annual visits, which include a neurological exam, a history and symptom survey, a blood draw, and neuropsychological testing. We set out to determine if the existing plasma biomarkers for AD could provide useful information in a subset of this cohort without CSF or PET characterization. We analyzed plasma levels of AD biomarkers, clinical and self-reported demographic information, and common medical comorbidities.

Methods

Study participants

Participants (N = 331) were members of the UM-MAP cohort of the MADRC, which recruits individuals from multiple sources, including the University of Michigan Health System clinics, Wayne State University Health Black Elders Center, and community outreach activities. Enrollment in UM-MAP also includes designation of a study partner. Participants or their surrogates provided written informed consent (and assent as appropriate). Participants were evaluated by a clinician who ascertained the presence of cognitive complaints and performed a neurologic exam, and underwent neuropsychological testing with a trained technician, blood draws, and neuroimaging. Blood draws used for biomarker measures and cognitive assessments were conducted within 12 months (median 0 months, mean 1.0 months, range 0–12 months) of each other. UM-MAP participants have annual clinical assessments, but data from only the first visit that included blood biomarker measures were analyzed in this work. UM-MAP and the work described here complied with federal and state laws and regulations. They were approved by the Institutional Human Use Review Board (IRB) of the University of Michigan Health System (HUM00000382), and participants were compensated as approved by the IRB.

Measurements of cognitive function

Measurements of cognitive function were collected in accordance with the Uniform Data Set (UDS) version 3 from the National Alzheimer’s Coordinating Center (NACC). The Clinical Dementia Rating scale (CDR) and the Montreal Cognitive Assessment (MoCA) were used for further analysis. Clinical phenotypes were established through a consensus conference that included at least three MADRC clinicians (neurologists/neuropsychologists) after review of all materials. Eligible participants for this study were those clinically categorized as one of the following phenotypes: (1) cognitively unimpaired (CU); (2) non-amnestic or amnestic, single-domain or multi-domain MCI of any suspected etiology, which is referred to collectively as MCI; or (3) amnestic multi-domain dementia syndrome felt due to AD, which is referred to as DAT. Our primary cognitive outcome variable was the clinical phenotype of either CU, MCI, or DAT. Of those categorized as MCI, 26% were clinically categorized as non-amnestic and 74% as amnestic MCI. These subgroups did not differ significantly with respect to plasma levels of measured biomarkers (Supplemental Table 1) and were therefore combined for further analyses.

We examined global cognition using the total MoCA score (range: 0–30) since it is a screening tool for cognitive impairment that assesses various aspects of cognitive functioning including visuospatial abilities, executive function, short-term memory, language, and orientation to time and place.^58,59 Higher scores indicate better cognitive functioning. The CDR is a clinical rating scale that gathers information from the participant and study partner to evaluate six aspects of cognitive and behavioral functioning (memory, orientation, judgment and problem-solving, community affairs, home and hobbies, and personal care). Two summary scores are calculated from the CDR.^60,61 We used the CDR Sum of Boxes (CDR-SB; range 0–18) score, since it provides more information and is better able to assess dementia gradations than the Global Score.⁶² Lower scores indicate better cognitive functioning. All cognitive measures were performed by trained and certified Michigan ADRC staff.

Plasma biomarker measurements

Participants’ blood was collected into 10 mL K2-ethylenediaminetetraacetic acid (EDTA) tubes. Plasma, buffy coat, and packed red cells were then processed and stored at −80°C. Plasma Aβ₄₀, Aβ₄₂, pTau-181, pTau-217, t-Tau, GFAP and NfL were measured on the Quanterix HD-1 or HD-X analyzer using the Simoa^® Neurology 3-Plex A Advantage (#101995 t-Tau, Aβ₄₂, Aβ₄₀), Simoa^® Neurology 2-Plex B Advantage (#103520 GFAP, NfL), ALZpath Simoa^® pTau-217 V2 (#104371), Simoa^® pTau-181 V2 (#103714) and Simoa^® pTau-181 V2.1 (#104112) kits. After thawing and mixing, plasma samples were centrifuged at 4°C for 5 min at 10,000×g. Samples were diluted according to manufacturer’s instructions using the instrument’s onboard dilution protocol and run in duplicate from a single well each on a 96-well plate. Eight-point calibration curves and sample measurements were determined on Simoa Analyzer software using a weighting factor 1/Y² and a four-parameter logistic curve fitting algorithm. Two levels of quality control material were included in each batch. Two bridge samples were included with each batch to monitor reproducibility. Intra-assay variance was <15% for all sample duplicates. Internal studies of inter-assay variance using pooled CU and DAT bridge samples can be found in Supplemental Table 2. We calculated the ratios of Aβ₄₂:Aβ₄₀ and pTau-217: Aβ₄₂ to compare with previous studies. For visualization and descriptive purposes, biomarker assays were used on their original scales. For regression modeling, biomarkers levels were z-score standardized.

APOE genotyping

APOE genotyping was performed at the National Centralized Repository for Alzheimer’s Disease and Related Dementias by DNA isolation and analysis of two APOE single nucleotide polymorphisms (rs429358 and rs7412). We operationalized this as the presence (one or two copies) versus absence of the APOE ϵ4 allele.

Covariate measures

Participant demographics (age in years, sex (male/female), race and ethnicity (NHW or non-Hispanic Black/African American, B/AA), educational attainment (high school or less, college graduate, professional degree), and medical history were ascertained at the clinical examination by self-report. Self-reported medical history measures included history of myocardial infarction, cancer, stroke, depression (in the last two years), diabetes, hypertension, hypercholesterolemia, presence of sleep apnea, and atrial fibrillation. Height (cm) and weight (kg) for each participant were measured and used to calculate body mass index (BMI; kg/m²).

Descriptive analyses

GraphPad Prism for macOS (version 10.2.3) and R Statistical Software (version 4.4.1) were used for generating graphs and statistical comparisons. Q-Q plots and the Shapiro-Wilk tests were used to determine the normality of data. Due to a lack of normal distributions within the data sets we described the distributions of continuous variables using median and quartiles. We described distributions of categorical covariates using count and frequency. Differences in demographic and clinical data and biomarker levels by cognitive status, sex, or race and ethnicity were tested with Chi-square and Mann-Whitney tests. We used scatter plots with median and interquartile range whiskers to visualize differences in biomarker levels between cognitive status. Kruskall-Wallis tests followed by Dunn’s test for multiple comparisons were conducted to examine differences in biomarker levels across groups stratified by clinical phenotype. Correlations between continuous variables were calculated using Spearman correlation. For these descriptive tests, we reported p values and Bonferroni multiple comparison-adjusted p values. A p value less than 0.05 was considered statistically significant.

Statistical modeling

To classify the three clinical phenotypes (CU, MCI, and DAT) we first performed single or dual plasma biomarker models (unadjusted multinomial logistic regression) in the full sample. We evaluated the classification performance of these models using receiver operating characteristic curve (ROC) area under the curve (AUC) analyses.

Next, we examined classification in models considering multiple predictor variables comparing the CU and DAT groups. To assess the performance of the classification model, the analytic sample was randomly divided into training data (80% of the whole dataset) and testing data (20% of the whole dataset), stratified by clinical phenotype. We compared the distributions of variables between the training and testing datasets to ensure even distribution. Eligible predictor variables for the classification models were based on prior research and they included age, sex, race, educational attainment, pTau-217, pTau-181, Aβ₄₂:Aβ₄₀ ratio, GFAP, NfL, pTau-217: Aβ₄₂ ratio, APOE alleles, and BMI. We used multiple machine-learning techniques for variable selection, informed by approaches used in previous research.^63,64 Specifically, among eligible features in the training data, we used stepwise logistic regression (backward selection) and penalized logistic regressions (lasso regression, ridge regression, and elastic net). The optimal model was determined based on metrics of best fit, including accuracy and AUC, comparing the CU and DAT groups. To assess the robustness of the stepwise regression model, we also used five-fold cross validation. We reported odds ratios and 95% confidence intervals for adjusted associations between biomarkers and cognitive status (CU versus DAT, and CU versus MCI)

We used the Youden index to calculate binary biomarker cutpoints between clinical diagnosis groups (CU versus DAT).^65,66 We evaluated the performance of the clinical cutpoints using sensitivity and specificity.

Results

Participant characteristics

The characteristics of the 331 participants are shown by clinical diagnosis in Table 1 and 48.6% were CU, 33.5% had MCI, 17.8% had DAT. Median age differed by clinical phenotype group, with age increasing from CU to MCI to DAT. The proportion of males differed by clinical phenotype group, with a larger proportion of male participants having a DAT diagnosis (Table 1). The majority (65.0%) of participants self-reported their race as NHW and 35.0% as B/AA. Though each clinical phenotype group included participants self-identifying as NHW or B/AA, the proportions were not evenly distributed across clinical phenotype groups (Table 1, Supplemental Table 3). The presence of an APOE ϵ4 allele, a known risk factor for dementia, was more prevalent in MCI and DAT groups than in CU participants (Table 1). APOE ϵ4 positivity did not differ between NHW (44.2%) and B/AA (38.3%) participants. Education levels differed by clinical phenotype (Table 1). While the median education level was 16 years for both NHW and B/AA participants, the mean years of education differed between the groups (16.2 versus 15.3 respectively, p = 0.005, Supplemental Table 3). BMI was lower in DAT than MCI (p = 0.01) and CU (p = 0.002) individuals. BMI was higher in B/AA participants than NHW individuals (p = 0.002, Supplemental Table 3). Significantly more NHW participants reported depression in the previous two years than B/AA participants (22.3% versus 3.4%, p < 0.001). Approximately half of the participants (51.9%) had hypertension and approximately half of the participants (52.9%) had hypercholesterolemia, with hypertension more prevalent in B/AA participants (p < 0.001). A prior cancer diagnosis (excluding that of the basal cell type) was reported by 22.7% of participants, and this prevalence was not different between NHW and B/AA racial groups. Sleep apnea was reported in 23.6% of participants, with a slightly higher prevalence in NHW participants (27.4% versus 16.4%, p = 0.030). Sleep apnea is considered a risk factor for the development of dementia, but the proportion of sleep apnea-positive participants did not differ among phenotype groups in this cohort.⁶⁷ The proportion reporting diabetes was higher in B/AA participants (19.8%) than in NHW participants (11.6%, p = 0.0501). Less than 5% of participants reported a history of stroke, myocardial infarction, and/or atrial fibrillation, so these measures were excluded from further analyses.

Table 1.

Study participant characteristics by clinical phenotype.

Characteristics	Overall (N=331) Median (IQR/N%)	CU (N=161) Median (IQR/N%)	MCI (N=111) Median (IQR/N%)	DAT (N=59) Median (IQR/N%)	p	Adjusted p
Age (y)	71.8 (67.7, 78.1)	70.5 (67.2, 76.3)	72.3 (67.5, 78.4)	75.3 (68.6, 82.5)	0.011*	0.033*
Sex					0.015*	0.045*
Male	112 (33.8%)	42 (26.1%)	45 (40.5%)	25 (42.4%)
Education					0.010*	0.030*
High School or less	60 (18.1%)	25 (15.5%)	20 (18.0%)	15 (25.4%)
College Graduate	136 (41.1%)	55 (34.2%)	55 (49.6%)	26 (44.1%)
Professional Degree	135 (40.8%)	81 (50.3%)	36 (32.4%)	18 (30.5%)
APOE ϵ4 allele (any)	139 (42.5%)	48 (30.2%)	53 (47.7%)	38 (66.7%)	<0.001***	<0.001***
Missing	4	2	0	2
Race					<0.001***	<0.001***
White	215 (65.0%)	108 (67.1%)	58 (52.3%)	49 (83.1%)
Black/African American	116 (35.0%)	53 (32.9%)	53 (47.7%)	10 (16.9%)
BMI (kg/m2)	27.9 (24.3, 31.3)	28.2 (24.8, 31.3)	28.3 (24.1, 31.9)	25.4 (23.0, 29.5)	0.013*	0.039*
pTau-181 (pg/mL)	2.3 (1.6, 3.4)	1.9 (1.5, 2.8)	2.3 (1.5, 3.3)	3.5 (2.8, 4.3)	<0.001***	<0.001***
Missing	1	0	1	0
pTau-217 (pg/mL)	0.47 (0.30, 0.83)	0.37 (0.27, 0.56)	0.48 (0.31, 0.83)	1.14 (0.81, 1.57)	<0.001***	<0.001***
Missing	5	1	2	2
pTau-217:Aβ42	0.060 (0.040, 0.115)	0.047 (0.035, 0.073)	0.061 (0.040, 0.113)	0.159 (0.109, 0.220)	<0.001***	<0.001***
Missing	5	1	2	2
Aβ42:Aβ40	0.041 (0.036, 0.046)	0.043 (0.038, 0.048)	0.041 (0.036, 0.045)	0.036 (0.033, 0.039)	<0.001***	<0.001***
GFAP (pg/mL)	171.6 (119.4, 240.2)	144.4 (113.8, 193.3)	175.0 (117.1, 238.3)	270.4 (201.1, 401)	<0.001***	<0.001***
NfL (pg/mL)	17.9 (12.8, 23.4)	16.1 (12.8, 21.8)	16.2 (11.1, 22.6)	23.8 (18.7, 33.0)	<0.001***	<0.001***
t-Tau (pg/mL)	3.7 (2.8, 4.9)	3.7 (3.0, 4.8)	3.5 (2.7, 4.8)	4.2 (3.0, 5.1)	0.113	0.339
CDR-Sum	0.5 (0, 1.5)	0 (0, 0.5)	0.5 (0.5, 1.0)	3.5 (2.0, 5.0)	<0.001***	<0.001***
MoCA Total	25 (21, 28)	27 (26, 28)	23 (21, 25)	16 (13, 19)	<0.001***	<0.001***
Missing	3	1	0	2
Health conditions (Self-report Yes)
Myocardial Infarction	11 (3.3%)	3 (1.9%)	5 (4.5%)	3 (5.1%)	^0.329	^0.987
Cancer	75 (22.7%)	31 (19.3%)	28 (25.2%)	16 (27.1%)	0.341	0.723
Stroke	10 (3.02%)	1 (0.62%)	4 (3.60%)	5 (8.48%)	^0.008	^0.024
Depression (in last 2yrs)	52 (15.7%)	22 (13.7%)	14 (12.6%)	16 (27.1%)	0.029*	0.087
Diabetes	48 (14.5%)	22 (13.7%)	16 (14.4%)	10 (16.9%)	0.828	1
Hypertension	172 (52.0%)	76 (47.2%)	65 (58.6%)	31 (52.5%)	0.182	0.546
Hypercholesterolemia	175 (52.9%)	79 (49.1%)	57 (51.4%)	39 (66.1%)	0.075	0.225
Sleep Apnea	78 (23.6%)	37 (23.0%)	27 (24.3%)	14 (23.7%)	0.967	1
Atrial Fibrillation	16 (4.8%)	9 (5.6%)	6 (5.4%)	1 (1.7%)	^0.500	^1

Numeric variables are summarized using median (25th percentile, 75th percentile). Categorical variables are summarized using number (percent). p values for continuous variables are from Kruskal-Wallis rank sum tests p value and for categorical variables from Pearson’s Chi-squared test and ^Fisher’s test and adjusted p value with Bonferroni corrections for multiple comparisons. Missing values: APOE ϵ4, 4; BMI, 1; MoCA, 3; pTau-217, 5; pTau-181, 1. BMI: body mass index; CDR-Sum: Clinical Dementia Rating-Sum of Boxes; MoCA: Montreal Cognitive Assessment; Aβ₄₀, amyloid-β 40; Aβ₄₂: amyloid-β 42; CU: cognitively unimpaired; MCI: mild cognitive impairment; DAT: Dementia of the Alzheimer type; pTau-217: phosphorylated tau 217; pTau-181: phosphorylated tau 181; GFAP: glial fibrillary acidic protein, NfL: neurofilament light chain; t-Tau: total tau protein

Individual plasma biomarkers by cognitive status, cognitive scores, and demographics

We compared individual biomarker levels and biomarker ratios by clinical phenotype groups (Figure 1).

Figure 1.

Plasma biomarker levels across clinical phenotype groups. a-f) The levels of (a) pTau-217 (b) pTau-181 (c) glial fibrillary acidic protein (GFAP) (d) Amyloid-β 42 to 40 ratio (Aβ₄₂:Aβ₄₀) (e) neurofilament light chain (NfL) (f) total Tau (t-Tau) and (g) pTau-217: Amyloid-β 42 ratio (pTau-217:Aβ42) were measured in plasma from 161 cognitively unimpaired (CU), 111 mild cognitive impairment (MCI) and 59 Dementia of the Alzheimer type (DAT) participants. Bars are medians and whiskers represent the interquartile ranges. p value is from Dunn’s test: *p < 0.05; **p < 0.01; ***p < 0.001.

Plasma pTau-217.

pTau-217 levels were elevated in DAT compared to MCI (p < 0.001) and CU (p < 0.001). In addition, the MCI group had higher pTau-217 levels compared to CU that bordered on significance (p = 0.050) (Figure 1(a)). The pTau-217 AUC comparing CU to DAT was 0.891 (95% CI: 0.84,0.94; p < 0.001, Figure 2(a) and Supplemental Table 4), CU to MCI was 0.589 (95% CI: 0.52,0.66; p = 0.013, Supplemental Table 4) and MCI to DAT was 0.803 (95% CI: 0.75,0.88; p < 0.0001, Supplemental Table 4). pTau-217 was significantly correlated with age (r = 0.32, p < 0.001, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = 0.45, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and negatively correlated with MoCA score (r = −0.38, p < 0.001, Supplemental Figure 3 and Supplemental Table 5). pTau-217 was negatively correlated with BMI (r = −0.23, p < 0.001, Supplemental Figure 4 and Supplemental Table 5). To determine the extent to which BMI might account for clinical phenotype-associated differences in pTau-217 levels, linear regression models were examined, using clinical phenotype group, BMI, and race as predictors of pTau-217 levels. Both clinical phenotype group and BMI were significant predictors, suggesting independent effects on pTau-217 (Supplemental Table 6). To assess the extent to which sex relates to pTau-217, female and male participants within each clinical phenotype group and across clinical phenotype groups were compared. Females had significantly lower levels of pTau-217 than males only in the CU group (p = 0.005, Supplemental Figure 5). To assess the extent to which self-identified race relates to pTau-217, NHW and B/AA participants’ pTau-217 levels were compared and were significantly different (p < 0.001) with B/AA participants having significantly lower pTau-217 levels. This difference was also significant within each clinical phenotype (Figure 3). Notably, within groups, race-associated CDR-SB differences were not significant (Supplemental Figure 6). Race was a significant predictor for pTau-217 levels in a linear regression model adjusted for clinical phenotype group and BMI, suggesting that BMI and race have independent effects on biomarker levels (Supplemental Table 6).

Figure 2.

Receiver operating characteristic (ROC) curves for assessed plasma biomarkers. a) Cognitively unimpaired (CU) versus dementia of the Alzheimer’s type (DAT) ROC curves for pTau-217 [navy, AUC = 0.89 (95% CI: 0.84,0.94), p < 0.001], the pTau-217: Amyloid-β 42 ratio [blue, AUC = 0.90 (95% CI: 0.85,0.94), p < 0.001], pTau-181 [teal, AUC = 0.80 (95% CI: 0.74,0.87), p < 0.001], glial fibrillary acidic protein [red, GFAP; AUC = 0.84 (95% CI: 0.77,0.90), p < 0.001], amyloid-β 42 to 40 ratio [orange, Aβ₄₂:Aβ₄₀; AUC = 0.74 (95% CI: 0.67,0.81), p < 0.001], neurofilament light chain [purple, NfL; AUC = 0.78 (95% CI: 0.72,0.85), p < 0.001], and total tau [turquoise, t-Tau; AUC = 0.57 (95% CI: 0.48,0.65), p = 0.129]. b) ROC curves for stepwise regression with backward selection model comparing CU participants and participants diagnosed with DAT [solid; AUC = 0.98 (95% CI: 0.95,1.00)], CU participants and participants diagnosed with MCI [dotted; AUC = 0.67 (95% CI: 0.52,0.82)], and MCI participants and participants diagnosed with DAT [dashed; AUC = 0.90 (95%CI: 0.78,1.00)].

Figure 3.

Self-reported race and levels of plasma biomarkers. a-b) The levels of plasma pTau-217 (a) and pTau-181 (b) were compared between self-identified non-Hispanic white (NHW) and Black/African American (B/AA) participants within the CU (gray), MCI (blue) or DAT (red) clinical phenotype groups. Bars are medians and whiskers represent the interquartile ranges. p values are from Mann-Whitney test comparing NHW to B/AA participants within each clinical phenotype group: *p < 0.05; **p < 0.01; ***p < 0.001.

Plasma pTau-181.

pTau-181 differed between CU or MCI individuals and those with DAT (both p < 0.001) (Figure 1(b)), but pTau-181 levels did not differ between CU and MCI (p = 0.88). The AUC when comparing CU to DAT was 0.804 (95% CI: 0.74,0.87; p < 0.001, Figure 2(a) and Supplemental Table 4) and MCI to DAT was 0.750 (95% CI: 0.69,0.83; p < 0.001, Supplemental Table 4), but CU to MCI was 0.537 (95% CI: 0.46,0.61; p = 0.293, Supplemental Table 4). pTau-181 was correlated with age (r = 0.26, p = 0.001, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = 0.35, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and was negatively correlated with MoCA score (r = −0.29, p < 0.001, Supplemental Figure 3 and Supplemental Table 5) and BMI (r = −0.15, p = 0.004, Supplemental Figure 4 and Supplemental Table 5). In this cohort, overall pTau-181 plasma concentration was lower in females than males (p = 0.003), driven by a difference within the CU group (p = 0.003, Supplemental Figure 5). The underlying reasons for this difference were not clear, as there was a lower CDR-SB average in CU females compared to males (p = 0.011), but MoCA total scores did not differ by sex within clinical phenotype groups. pTau-181 concentration was lower in B/AA participants when compared to NHW participants across clinical phenotype groups (p < 0.001) and within groups (Figure 3). Race was a significant predictor of pTau-181 levels in a linear regression model adjusted for clinical phenotype group and BMI, suggesting that race affects pTau-181 levels independent of race-associated BMI differences (Supplemental Table 6).

Plasma GFAP.

Comparisons by clinical phenotype groups show that GFAP levels were significantly elevated in DAT compared to MCI (p < 0.001) and CU (p < 0.001), but were not different between CU and MCI (p = 0.184, Figure 1(c)). The AUC when comparing CU to DAT was 0.836 (95% CI: 0.77,0.90; p < 0.001, Figure 2(a) and Supplemental Table 4), CU to MCI was 0.574 (95% CI: 0.46,0.61; p = 0.038, Supplemental Table 4) and MCI to DAT was 0.753 (95% CI: 0.69,0.83; p < 0.001, Supplemental Table 4). GFAP was correlated with age (r = 0.48, p < 0.001, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = 0.39, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and negatively correlated with MoCA score (r = −0.36, p < 0.0001, Supplemental Figure 3 and Supplemental Table 5) and BMI (r = −0.24, p < 0.0001, Supplemental Figure 4 and Supplemental Table 5). GFAP levels were higher in females than males only within the MCI group (Supplemental Figure 5). GFAP levels were lower in B/AA participants compared to NHW participants only within the CU group (p = 0.043), though MCI and DAT groups trend similarly (p = 0.26, p = 0.29 respectively); Supplemental Figure 7). Race was not a significant predictor in a linear regression model adjusted for clinical phenotype group and BMI, suggesting that race-associated differences observed in GFAP levels may be largely explained by race-associated BMI differences (Supplemental Table 6).

Plasma Aβ₄₂:Aβ₄₀ ratio.

Measurements of individual Aβ species (Aβ₄₂ or Aβ₄₀) demonstrated limited differences across the clinical phenotypes (Supplemental Figure 8). This is consistent with prior work and drove us to use the ratio of Aβ₄₂ to Aβ₄₀. We found that Aβ₄₂:Aβ₄₀ differed between CU an DAT individuals (p < 0.001) and between MCI and DAT individuals (p = 0.001), but not between CU and MCI individuals (p = 0.108, Figure 1(d)). The AUC, when comparing CU to DAT, was 0.742 (95% CI: 0.67,0.81; p < 0.001, Figure 2(a) and Supplemental Table 4), MCI to DAT was 0.655 (95% CI: 0.60,0.77; p = 0.001, Supplemental Table 4), and CU to MCI was 0.590 (95% CI: 0.51,0.65; p = 0.011, Supplemental Table 4). Aβ₄₂:Aβ₄₀ was negatively correlated with age (r = −0.15, p = 0.006, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = −0.224, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and positively correlated with MoCA score (r = 0.185, p = 0.001, Supplemental Figure 3 and Supplemental Table 5) and BMI (r = 0.198, p < 0.001, Supplemental Figure 4 and Supplemental Table 5). When normalized to BMI, group differences were no longer significant (Supplemental Table 6). No significant sex-associated differences were identified in the Aβ₄₂:Aβ₄₀ ratio (Supplemental Figure 5). The Aβ₄₂:Aβ₄₀ ratio was higher in B/AA participants when compared to NHW MCI (p = 0.001) and CU participants (p = 0.004), but this difference is not seen in the DAT group (p = 0.25); Supplemental Figure 7). Race was not a significant predictor of Aβ₄₂:Aβ₄₀ ratio in a linear regression model adjusted for clinical phenotype group and BMI, suggesting that race-associated differences observed in Aβ₄₂:Aβ₄₀ ratio may be largely explained by race-associated BMI differences (data not shown).

Plasma NfL.

NfL differed between individuals with a clinical phenotype of DAT and CU or MCI individuals (p < 0.001 for both) but levels between CU and MCI individuals were not significantly different (p > 0.999, Figure 1(e)). The AUC when comparing CU to DAT was 0.785 (95% CI: 0.72,0.85; p < 0.001, Figure 2(a) and Supplemental Table 4), MCI to DAT was 0.765 (95% CI: 0.70,0.84; p < 0.001, Supplemental Table 4) and CU to MCI was 0.516 (95% CI: 0.45,0.59; p = 0.656, Supplemental Table 4). NfL was correlated with age (r = 0.490, p < 0.001, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = 0.304, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and negatively correlated with MoCA score (r = −0.215, p < 0.001, Supplemental Figure 3 and Supplemental Table 5) and BMI (r = −0.200, p < 0.001, Supplemental Figure 4 and Supplemental Table 5). Within the CU group, the NfL plasma concentration was lower in females than in males (p = 0.042) (Supplemental Figure 5). To assess the extent to which self-identified race related to NfL concentration, NHW and B/AA participants’ NfL levels were compared by clinical phenotype groups. A difference was seen within the CU group (p = 0.030, Supplemental Figure 7), but not MCI and DAT groups (p = 0.09, p = 0.10, respectively). Race was not a significant predictor in a linear regression model when the model was adjusted for clinical phenotype group and BMI, suggesting that race-associated differences observed in NfL levels may be largely explained by race-associated BMI differences (Supplemental Table 6).

Plasma t-tau.

Total tau levels did not differ between clinical phenotype groups in this cohort (Figure 1(f) and Supplemental Table 4). Total tau levels were not correlated with age, CDR-SB, MoCA score, or BMI (Supplemental Figures 1–4, Supplemental Table 5). No significant sex-associated differences were identified in t-Tau levels (Supplemental Figure 5). The total tau plasma concentration was lower in CU B/AA participants when compared to CU NHW participants (p = 0.037, Supplemental Figure 7), but levels in NHW and B/AA participants were not different in MCI and DAT groups (p = 0.84, p = 0.09, respectively). Race was a significant predictor of t-Tau concentration in a linear regression model adjusted for clinical phenotype group and BMI, while clinical phenotype group was not significant (Supplemental Table 6).

Plasma pTau-217:Aβ₄₂ ratio.

The pTau-217:Aβ₄₂ ratio differed between CU or MCI individuals and those with DAT (both p < 0.001) (Figure 1(g)), and between CU and MCI individuals (p = 0.031). The AUC when comparing CU to DAT was 0.896 (95% CI: 0.85,0.94; p < 0.001, Figure 2(a)) and MCI to DAT was 0.803 (95% CI: 0.76,0.89; p < 0.001, Supplemental Table 4), but CU to MCI was 0.596 (95% CI 0.52,0.67; p = 0.008, Supplemental Table 4). pTau-217:Aβ₄₂ was correlated with age (r = 0.27, p < 0.001, Supplemental Figure 1 and Supplemental Table 5) and CDR-SB (r = 0.46, p < 0.001, Supplemental Figure 2 and Supplemental Table 5) and was negatively correlated with MoCA score (r = −0.38, p < 0.001, Supplemental Figure 3 and Supplemental Table 5) and BMI (r = −0.29, p < 0.001, Supplemental Figure 4 and Supplemental Table 5). Similar to other biomarkers, overall pTau-217:Aβ₄₂ was lower in females than males within the CU group (p = 0.019, Supplemental Figure 5) pTau-217:Aβ₄₂ was lower in B/AA participants when compared to NHW participants across the CU and MCI clinical phenotype groups (p = 0.030 and p < 0.001, respectively) (Supplemental Figure 7). Race was a significant predictor of pTau-217:Aβ₄₂ in a linear regression model adjusted for clinical phenotype group and BMI, suggesting that BMI and race have independent effects on biomarker levels (Supplemental Table 6).

Multivariable modeling of the associations between biomarkers and cognitive status

In our testing and training sets, we assessed multiple types of variable selection procedures (stepwise regression with backwards selection, lasso, ridge, and elastic net) (Table 2). We assessed models using AUC analysis and testing and training accuracy and determined that optimum classification of clinical phenotype groups was obtained with stepwise regression.

Table 2.

Comparison of accuracy in classification of clinical phenotypes between the testing and training datasets, using multiple machine learning methods.

Models	Train Accuracy	Test Accuracy	Diff	AUC on Test Data (95% CI)
				CU-DAT	MCI-DAT	CU-MCI
Stepwise Multinomial Logistic Regression (Backward Selection)	61.0%	65.1%	4.1%	98.2% (95.2, 100)	89.6% (78.4, 100)	67.4% (52.4, 82.4)
Stepwise Multinomial Logistic Regression (Backward Selection) without pTau-181	60.0%	65.0%	5.0%	99.7% (98.9, 100)	91.3% (81.9, 100)	69.0% (54.1, 83.9)
Lasso Regression	55.0%	70.3%	15.3%	90.0% (76.9, 100)	76.7% (58.5, 94.8)	66.7% (55.5, 77.9)
Ridge Regression	58.1%	71.9%	13.8%	93.3% (82.5, 100)	79.1% (61.7, 96.6)	73.1% (60.2, 86.1)
Elastic Net Regression	56.6%	70.3%	13.7%	94.4% (84.4, 100)	79.7% (63.2, 96.3)	68.3% (56.3, 80.2)

DAT: Dementia of the Alzheimer Type; CU: Cognitively Unimpaired; MCI: Mild Cognitive Impairment; Diff: Difference between the testing and training accuracy; AUC: Area Under the Curve; CI: Confidence Interval

Our optimum model, the stepwise logistic regression model with backward selection, was trained with an initial set of predictor variables, including age, sex, race, years of education, pTau-217, pTau-181, Aβ₄₂: Aβ₄₀ ratio, GFAP, NfL, APOE alleles, and BMI. Through iterative and model selection procedures, the final stepwise logistic regression model was generated. The significant predictor variables were sex, race, years of education, pTau-217, pTau-181, GFAP, and APOE. The variance inflation factor for pTau-217 (VIF = 3.29), pTau-181 (VIF = 2.44), and GFAP (VIF = 1.62) were all less than 5, indicating relatively low multicollinearity among these variables. After training the model through cross-validation, the model’s performance was evaluated with the testing dataset. The stepwise model’s overall testing accuracy was 65%, differentiating between CU, MCI, and DAT. For distinguishing between CU and DAT, the model’s AUC was 98.2% (95% CI: 95.2,100), and for differentiating between MCI and DAT, the model’s AUC was 89.6% (95% CI: 78.4,100). The performance of classification between CU and MCI had an AUC of 67.4% (95% CI: 52.4,82.4) (Table 2).

Our model identified significant biomarker predictors for DAT (Supplemental Table 7). A one-standard deviation increase in pTau-217 was associated with 4.30 times greater odds of DAT (95% CI: 1.97, 9.37; p < 0.001) relative to CU. Similarly, a one-standard deviation increase in GFAP was associated with 2.12 times greater odds of DAT (95% CI: 1.23, 3.67; p = 0.007), compared to CU. Similar findings were observed comparing CU and MCI, though the magnitudes of association were attenuated. Unexpectedly, a one-standard deviation increase in pTau-181 was marginally associated with 0.52 lower odds of DAT (0.26, 1.03, p = 0.062), compared to CU. As pTau-217 and pTau-181 are typically correlated and pTau-217 explained most of the DAT-related variance, pTau-217 likely acted as a suppressor by removing irrelevant variance in pTau-181, thus causing the odds ratio of pTau-181 to reverse.⁶³

As a sensitivity analysis, this model was run while omitting pTau-181, to reduce biological redundancy. In the model lacking pTau-181, overall testing accuracy was 65%, similar to the model including pTau-181. For distinguishing between CU and DAT, the model’s AUC was 99.7% (95% CI: 98.9,100), and for differentiating between MCI and DAT, the model’s AUC was 91.3% (95% CI: 81.9,100). The performance of classification between CU and MCI had an AUC of 69.0% (95% CI: 54.1,83.9) (Table 2). In the model without pTau-181, one-standard deviation increase in pTau-217 was associated with 2.52 times greater odds of DAT (95% CI: 1.49, 4.25; p < 0.001) relative to CU and a one-standard deviation increase in GFAP was associated with 2.24 times greater odds of DAT (95% CI: 1.31, 3.84; p = 0.003), compared to CU. These results suggest that pTau-217 explains much of the DAT-related variance from phosphorylated tau (Table 2).

Biomarkers and common comorbidities

We sought to assess the biomarker variables included in our stepwise logistic regression model with backward selection (pTau-217, pTau-181, and GFAP) in the context of common comorbidities. We investigated associations between these biomarkers and self-reported comorbidities. Myocardial infarction, stroke, and atrial fibrillation were reported by fewer than 5% of participants and were therefore excluded from these analyses. Examined comorbidities were not significantly associated with plasma levels of pTau-217 or pTau-181 (Figure 4). GFAP had a positive association with cancer and hypercholesterolemia, suggesting that these comorbidities may affect plasma levels of GFAP (Figure 4).

Figure 4.

Forest plots showing the relationships between self-reported comorbidities [cancer n = 75 (22.7% of the cohort), Depression n = 52 (15.7%), Diabetes n = 48 (14.5%), Hypertension n = 172 (52.0%), Hypercholesterolemia n = 175 (52.9%), Sleep apnea n = 78 (23.6%)] and levels of plasma biomarkers included in our stepwise logistic regression model with backward selection. Coefficients and confidence intervals were calculated for (A) pTau-217, (B) pTau-181, and C) GFAP.

Biomarker cutpoint calculations

To test the potential clinical utility of our analyses, we estimated a binary cutpoint (threshold) for each biomarker by comparing CU and DAT participants using the Youden index (Table 3). We used clinical phenotype as an endpoint in this study, rather than Aβ PET, which has been used as an endpoint in previous studies, to determine if plasma biomarker cutpoints could be calculated without CSF or PET characterization.^60,61

Table 3.

Thresholds for DAT positivity compared to CU for each biomarker in the whole sample and by race/ethnicity.

	All (N = 59 DAT, N = 161 CU)	NHW participants (N = 49 DAT, N = 108 CU)	B/AA participants (N = 10 DAT, N = 53 CU)
pTau-217
Threshold (pg/mL)	> 0.75	> 0.80	> 0.75
Sensitivity	80.7%	80.9%	70.0%
Specificity	89.4%	88.9%	96.2%
pTau-181
Threshold (pg/mL)	> 3.04	> 3.04	> 1.84
Sensitivity	72.9%	77.55%	90.0%
Specificity	80.1%	75%	67.9%
GFAP
Threshold (pg/mL)	> 180.20	> 215.32	> 168.91
Sensitivity	86.4%	79.6%	90.0%
Specificity	70.2%	81.5%	77.4%
NfL
Threshold (pg/mL)	> 18.07	> 27.91	> 18.04
Sensitivity	86.4%	53.1%	90.0%
Specificity	59.6%	91.7%	69.8%
Aβ42:Aβ40
Threshold (ratio)	< 0.038	< 0.038	< 0.040
Sensitivity	71.2%	69.4%	90.0%
Specificity	76.4%	75.9%	77.4%
pTau-217:Aβ42
Threshold (ratio)	> 0.107	> 0.107	> 0.094
Sensitivity	75.4%	78.7%	70.0%
Specificity	89.4%	86.1%	96.2%

Thresholds were estimated using the Youden index and evaluated with sensitivity and specificity. DAT: Dementia of the Alzheimer Type; CU: Cognitively Unimpaired

Discussion

Blood-based biomarkers demonstrate potential for the prediction and detection of neuropathological causes of clinical symptoms, having good concordance with more traditional biomarker measurements.^55,68,69 With the relative ease and low cost of blood acquisition and testing, blood-based biomarkers may soon enable earlier detection of AD (and other etiologies) and enable more frequent testing for monitoring disease progression. However, many validation studies utilize participant cohorts with clinical diagnoses confirmed by PET imaging and CSF biomarkers.^65,70,71 We aimed to investigate the utility of pTau-217, pTau-181, GFAP, NfL, Aβ₄₂:Aβ₄₀, and t-Tau plasma biomarkers in a clinical cohort that lacks the extensive characterization provided by more invasive and expensive procedures, such as CSF and PET biomarker analyses. We also examined the relationships between plasma biomarkers, demographic variables, comorbidities, and cognitive measures.

We first assessed the clinical phenotype classification accuracy of plasma biomarkers in UM-MAP participants. All biomarker levels assessed, except t-Tau, were significantly different between CU and DAT participants. pTau-217, pTau-181, GFAP and NfL levels and the ratios of Aβ₄₂:Aβ₄₀ and pTau-217:Aβ₄₂ were significantly different between MCI and DAT participants. Our analysis supports the use of pTau-217, pTau-181, GFAP, NfL, Aβ₄₂: Aβ₄₀ and pTau-217:Aβ₄₂ as measures with potential diagnostic value. These results agree with similar studies and demonstrate the power of these biomarkers to detect physiological processes associated with cognitive impairment, even within diverse research cohorts.^23,47,72 Of the biomarkers assessed, only pTau-217 and pTau-217:Aβ₄₂ differed significantly between CU and MCI individuals, with the single biomarker measurement outperforming the ratio. This is comparable to what was observed in studies of plasma biomarkers in well-characterized cohorts, where plasma pTau-217 had the best performance of biomarkers in discriminating between Aβ-positive and Aβ-negative individuals, determined by Aβ PET.^11,65 This may also reflect the mixed etiologies in our MCI participants. Those with MCI not due to underlying AD would not be expected to have elevated AD biomarkers. Collectively, our work and that of others demonstrate the potential ability of pTau-217 to identify those at risk of progression to dementia earlier than other established biomarkers in well-characterized^11,65 and diverse cohorts.⁸ Future longitudinal biomarker studies in the UM-MAP cohort will likely focus on the ability of these biomarkers to detect those most at risk of clinical decline.

All biomarkers assessed except t-Tau correlated with age, which is a known risk factor for dementia and is controlled for in most studies of AD plasma biomarkers.⁶ The correlation observed in this study may be partially due to the difference in median age among the clinical phenotype groups in this cohort. However, the association between plasma biomarker concentration and age was significant for most biomarkers (pTau-181, pTau-217, pTau-217: Aβ₄₂, GFAP, and NfL) even when limiting the analysis to the CU group. Since plasma phosphorylated tau, GFAP, and NfL were shown to increase over a decade prior to the onset of cognitive symptoms, this may also represent a portion of CU individuals with prodromal disease.^12,71 Levels of these biomarkers may increase with age in the absence of cognitive symptoms, but it is also possible that CU individuals with higher biomarker levels may represent pre-clinical dementia stages, as studies show that baseline biomarker levels in CU individuals can predict future cognitive decline.^10,24 However, age was not a predictor variable identified in our stepwise regression analysis. This is likely due to failure to increase clinical phenotype category discrimination more than what is accounted for by the biomarkers associated with aging.

All biomarkers assessed except t-Tau were negatively correlated with BMI, where participants with a higher BMI were more likely to have lower plasma biomarker levels. This may result from an increase in total plasma volume for those with a higher BMI, thus diluting proteins originating in the brain that are indicative of pathological processes.^56,73 Alternatively, the correlation between lower BMI and higher plasma biomarker concentrations could be due to weight loss associated with dementia, as BMI also correlates with CSF biomarkers of AD.^74,75 However, this is unlikely to explain the correlation observed across all of the clinical phenotypes. Despite the potential influences of BMI on the plasma biomarkers assessed in this study, when we normalized biomarkers to BMI it did not decrease discriminatory accuracy for most biomarkers or the coefficient of variance in the sample set. This suggests that the variability in the data set was not due to variability in body size. The exception to this was plasma Aβ₄₂:Aβ₄₀, which no longer discriminated between groups after BMI normalization. Indeed, Hermesdorf et al. found a strong correlation between BMI and Aβ₄₂ plasma levels and Syrjanen found that individuals with high BMI also had high Aβ₄₂ and Aβ₄₀ levels.^53,76 Plasma Aβ levels may be affected by BMI more than other plasma biomarkers of AD. BMI was not a variable included in our logistic regression model with backward selection, supporting the assertion that it is not one of the main sources of variability in clinical phenotype group differences. Brickman et al. similarly found that including BMI as a covariate in their logistic regression model did not affect their modeled risk of developing DAT, nor was it predictive of outcome.²³ However, awareness of the potential for BMI to influence blood-based protein levels and potentially mask indications of pathology is important as plasma biomarkers move into clinical use.

pTau-181, pTau-217, and NfL levels were lower in CU females than CU males. Sex-associated difference between pTau-217, pTau-181, and t-Tau levels in CU participants was noted previously in other cohorts^6,8,72,77 but these findings were not consistent throughout other studies.^23,78,79 Tsiknia et al. suggested that the effect of sex may be more subtle than increasing biomarker levels, demonstrating a sex-biomarker interaction on change in cognition over time.⁷⁹ In this cohort, cognitive differences between CU females and males may be responsible for the sex-associated differences observed, as the CU female CDR-SB was lower than that of CU males. Additionally, the noted difference in the proportion of males in each clinical phenotype group may be due to selection bias.⁸⁰ It may also be the case that more males in this CU cohort are in the prodromal stage of DAT, and have higher phosphorylated tau levels that reflect amyloid pathology. The lack of a broad effect of sex on biomarker levels observed in this study does not negate the urgent need for additional studies related to the interplay of biological sex and risk factors of DAT.⁸¹

The concentrations of pTau-181 and pTau-217 were significantly lower in B/AA participants than NHW participants across all groups. This was reported previously for dementia-associated biomarker levels in CSF and plasma.^82–85 However, few studies have examined phosphorylated tau associations with race in plasma, and those that have show conflicting results.^{45,72,86–88} While adjusting for comorbidities ameliorated race-associated differences in previous work, the comorbidities examined in this study were not significantly associated with phospho-tau biomarker levels.^45,89,90 Chronic kidney disease (CKD) elevates plasma levels of Aβ₄₂, Aβ₄₀, pTau-217, pTau-181, and NfL,^{11,33,53,72,91} which may be due to decreased renal clearance.^92,93 We do not have access to clinical data to determine the prevalence of CKD in the UM-MAP cohort, and therefore, a higher incidence of CKD in our NHW participants could help explain the race-associated differences observed. CDR-SB is slightly lower in B/AA participants compared to NHW participants, but this difference is not sustained across clinical phenotype groups and is therefore unlikely to fully explain the differences observed. This race-associated difference may reflect group imbalances. For example, the MCI may be less often due to underlying AD etiology in UM-MAP B/AA participants. Genetics may also play a role, as people who carry two copies of the APOE ϵ4 and are of European descent have a higher level of AD risk than those of African descent.⁹⁴ Recent studies show a strong correlation between pTau-217 levels and Aβ PET in racially and ethnically underrepresented groups suggesting that race-related differences in biomarker levels reflect diagnostic practices and/or contextual factors rather than underlying pathology.⁹⁵ There are likely other socioeconomic or demographic variables that are not captured in our data that are contributing to these differences. Differences in neighborhood characteristics, social engagement, sleep habits, and educational characteristics not captured by our measurement of education years may all affect biomarker levels.^82,96,97 Race is likely a mediating factor for variables that have a direct effect on biomarker levels, rather than a causative factor.

This race-associated difference was also reflected in our calculations of biomarker cutoffs, with cutoffs higher for NHW participants than B/AA participants. While we were unable to identify explanatory variables for our race-associated difference in plasma phosphorylated tau concentrations, this may have importance as plasma biomarkers move into clinical use for diagnostic, prognostic, and clinical trial purposes. The binary biomarker cutoffs calculated from this cohort are similar to those published using more well-characterized cohorts.^65,66 Differences likely reflect that our biomarker cutpoint values were determined using clinical phenotype as an endpoint rather than Aβ PET positivity. It also should be noted that these cutpoint values are specific to the participant population associated with the MADRC UM-MAP cohort and to the assays, platforms, and pre-analytic methods used in this study. Importantly, the similarity of our cutpoints to published values highlights that plasma biomarkers can classify clinical phenotype in this cohort without CSF or PET characterization.

While it remains possible that a single plasma biomarker could be utilized for dementia diagnoses, it is far more likely that a panel of biomarkers will have a greater ability to discriminate between clinical phenotype groups. Using a stepwise regression model with backward selection, we determined that in this cohort a model including a combination of pTau-217 and GFAP plasma biomarkers and the demographic variables APOE alleles, education level, sex, and race discriminated between CU and DAT individuals more reliably than any single biomarker. pTau-181, though included in our initial stepwise regression model, did not improve classification performance of the model, providing support for the superiority of pTau-217 to predict clinical phenotype group. Self-reported comorbidities did not appear to be associated with the levels of the phosphorylated tau plasma biomarkers included in the model, suggesting that this model may retain validity regardless of the presence of the examined comorbidities. The additional biomarker included in the model, GFAP, was associated with two of the common medical comorbidities examined. The positive association between GFAP concentration and hypercholesterolemia was noted in a previous study.⁴⁹ Hypercholesterolemia was self-reported in our study, thus, we were unable to determine the strength of the correlation between cholesterol levels and GFAP levels. GFAP has long been considered a biomarker for the presence of glioblastomas, but its association with a history of other cancers should be investigated further. It should be noted that medical correction of comorbidities was not taken into account in this study, but could mask associations of comorbidities with biomarker levels.

Key strengths of the study include the racially diverse sample of study participants with rigorous clinical assessments and measurements of many established plasma biomarkers of neurodegeneration. We also identified a modeling approach including pTau-217, GFAP, and demographic variables that provided an AUC of 99.7% for distinguishing between CU and DAT and 91.3% for differentiating between MCI and DAT in a diverse research cohort. This cohort is likely more representative of the general population than that which might be found in a more specialized academic setting. However, the demographic variables and comorbidities associated with participants were not evenly distributed across our clinical phenotype categories, leaving open the possibility that some of the biomarker differences observed are driven by underlying differences unassociated with cognitive status. Additionally, a reliance on clinical consensus diagnosis as an endpoint in our analyses, rather than PET or CSF values, inherently introduces some misclassifications. Additional limitations include self-report of comorbidities, a lack of information about the presence of CKD in this cohort, and a lack of CSF and PET data regarding amyloid and tauopathy neuropathology status. The relatively small number of B/AA participants with a DAT diagnosis precludes robust assessment of how race/ethnicity correlates with plasma biomarkers specifically in individuals diagnosed with DAT. Our data highlight potential race/ethnicity differences in the association of levels of plasma biomarkers (particularly pTau-217 and pTau-181) with clinical diagnosis that may affect cutpoints when utilizing such measures in diverse populations. Though some studies suggest differences may be mitigated or eliminated when neuropathological biomarkers (i.e., PET imaging) are used as the reference as opposed to clinical diagnosis,^88,95,98 it remains critical to identify whether differences in performance exist as biofluid biomarkers are more widely validated across multiple populations. Finally, the cross-sectional findings presented here will need further validation in larger independent replication datasets with prospective prediction models. Despite the limitations, our work clearly indicates the diagnostic utility of many established plasma biomarkers of DAT, and the superiority of pTau-217 in discriminating between CU and impaired individuals in a diverse research cohort without CSF or PET biomarkers. This work also highlights the continued need to investigate diverse populations in the context of dementia and associated biomarkers.

Supplementary Material

Supplemental material for this article is available online.

Data availability statement

The data supporting the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.