Pathologic and cognitive correlates of plasma biomarkers in neurodegenerative disease

Abstract

INTRODUCTION

We investigate pathological correlates of plasma phosphorylated tau 181 (p‐tau₁₈₁), glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL) across a clinically diverse spectrum of neurodegenerative disease, including normal cognition (NormCog) and impaired cognition (ImpCog).

METHODS

Participants were NormCog (n = 132) and ImpCog (n = 461), with confirmed β‐amyloid (Aβ+/‐) status (cerebrospinal fluid, positron emission tomography, autopsy) and single molecule array plasma measurements. Logistic regression and receiver operating characteristic (ROC) area under the curve (AUC) tested how combining plasma analytes discriminated Aβ+ from Aβ‐. Survival analyses tested time to clinical dementia rating (global CDR) progression.

RESULTS

Multivariable models (p‐tau+GFAP+NfL) had the best performance to detect Aβ+ in NormCog (ROCAUC = 0.87) and ImpCog (ROCAUC = 0.87). Survival analyses demonstrated that higher NfL best predicted faster CDR progression for both Aβ+ (hazard ratio [HR] = 2.94; p = 8.1e‐06) and Aβ‐ individuals (HR = 3.11; p = 2.6e‐09).

DISCUSSION

Combining plasma biomarkers can optimize detection of Alzheimer's disease (AD) pathology across cognitively normal and clinically diverse neurodegenerative disease.

Highlights

• Participants were clinically heterogeneous, with autopsy‐ or biomarker‐confirmed Aβ.

• Combining plasma p‐tau₁₈₁, GFAP, and NfL improved diagnostic accuracy for Aβ status.

• Diagnosis by plasma biomarkers is more accurate in amnestic AD than nonamnestic AD.

• Plasma analytes show independent associations with tau PET and post mortem Aβ/tau.

• Plasma NfL predicted longitudinal cognitive decline in both Aβ+ and Aβ‐ individuals.

1. BACKGROUND

The development of ultrasensitive immunoassays has made possible blood based biomarkers that can accurately detect Alzheimer's disease (AD) pathology. Among the top plasma analytes with established high utility in AD are tau phosphorylated at threonine 181 (p‐tau₁₈₁), glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL). While the majority of studies have tested these analytes individually, several have evaluated the combined potential of plasma biomarkers to improve diagnostic accuracy for AD. ¹ , ² , ³ , ⁴ Here, we expanded on prior work by testing combinations of p‐tau₁₈₁, GFAP, and NfL in a clinically diverse sample with autopsy/biomarker‐confirmed pathology. We performed head‐to‐head comparisons to evaluate biomarker diagnostic accuracy across the disease span of healthy participants with normal cognition (NormCog), and symptomatic individuals with impaired cognition (ImpCog) that includes amnestic and atypical forms of AD‐dementia. In addition to diagnostic evaluation, we test biomarker combinations to predict future cognitive decline.

Autopsy studies show that plasma p‐tau₁₈₁ is associated with both post mortem β‐amyloid (Aβ) and tau burden, and is sensitive to high AD neuropathologic change (ADNC). ⁴ , ⁵ However, plasma p‐tau₁₈₁ may be less sensitive to earlier AD compared to other biomarkers, ⁶ , ⁷ and also less sensitive to intermediate ADNC than high ADNC. ³ , ⁸ , ⁹ In this context, diagnostic accuracy for Aβ status by plasma p‐tau₁₈₁ may be augmented by the incorporation of other plasma biomarkers. ¹ , ¹⁰

Plasma GFAP is a biomarker of reactive astrogliosis, not specific to AD. ¹¹ Nonetheless, there is a robust astroglial response to Aβ pathology in the brain, ¹² , ¹³ and increases in blood GFAP are typically highest in AD compared to other neurodegenerative diseases, such as frontotemporal dementia (FTD) and dementia with Lewy bodies (DLB). ¹⁴ Moreover, plasma GFAP may be more sensitive to intermediate levels of ADNC than p‐tau₁₈₁, ⁸ showing elevation at pre‐clinical/mild cognitive impairment (MCI) stages earlier than p‐tau₁₈₁ in Aβ+ PET individuals. ⁶ , ¹⁵ , ¹⁶ Plasma GFAP also associates with cognition and progression from clinically normal/MCI to amnestic dementia, ¹³ , ¹⁷ but it is unclear if prognostic utility differs in AD versus non‐AD pathology, or how GFAP compares to other plasma biomarkers.

Plasma NfL is a non‐specific marker of axonal degeneration that is elevated across neurodegenerative diseases compared to healthy controls. ¹⁸ When compared to controls, plasma NfL increases in association with AD pathology. ¹⁹ However, because NfL is highly elevated in non‐AD, NfL associations with AD pathology are often null or inversely related in mixed neurodegenerative samples that include FTD and DLB syndromes. ⁴ , ⁷ , ²⁰ , ²¹ , ²² Thus, it is unclear whether plasma NfL combined with other plasma biomarkers adds diagnostic value to stratify Aβ+ from Aβ‐ neurodegeneration. In addition to diagnostic potential, higher plasma NfL is associated with faster cognitive decline in AD and non‐AD. ⁹ , ¹⁹

For maximum efficacy, AD biomarkers must be accurate across the heterogeneous clinical spectrum encountered in dementia clinics. This includes different levels of cognitive impairment (e.g., NormCog, ImpCog), clinically typical and atypical AD (e.g., amnestic, non‐amnestic), as well as to discriminate AD from various non‐AD neurodegenerative diseases. In this study, we evaluated plasma biomarkers individually and in combination for detection of Aβ pathology and prognosis of cognitive decline. First, we tested diagnostic accuracy and pathological correlates of plasma biomarkers by disease stage: in NormCog (part 1) and ImpCog (part 2). Consistent with the framework proposed by the Dementia Nomenclature Initiative, ²³ motivated by the imperfect mapping of clinical syndrome with pathology, we define cases by syndrome and use plasma biomarkers to link to underlying pathoetiology. We confirmed Aβ status of participants using autopsy, cerebrospinal fluid (CSF), or ¹⁸F‐florbetaben positron emission tomography (PET). Head‐to‐head comparisons tested plasma biomarkers individually and in combination to stratify Aβ+ from Aβ‐, and we tested if diagnostic accuracy differed by sex, race, or phenotype. We further tested plasma associations with pathological Aβ and tau severity, measured using PET and autopsy data. Second, we tested prognostic associations of plasma biomarkers in Aβ+ (part 3) and Aβ‐ individuals (part 4). Linear mixed models tested associations of plasma analytes at baseline with longitudinal cognitive decline, and survival analyses compared biomarkers individually and in combination to predict future cognitive decline.

2. METHODS

Participants were selected retrospectively from the University of Pennsylvania (Penn) Integrative Neurodegenerative Disease Database (INDD). ²⁴ , ²⁵ Participants were recruited for observational research at the Penn AD and FTD research centers, and blood was banked as part of ongoing clinical research programs. Written consent was obtained according to the Declaration of Helsinki and approved by the Penn Institutional Review Board.

2.1. Syndromic/neuropsychological diagnoses

Participants were classified based on clinical syndrome according to National Alzheimer's Coordinating Center (NACC) Uniform Dataset 3.0 criteria ²⁶ at the time of plasma collection. NormCog participants were without cognitive impairment (n = 132). ImpCog participants (n = 461) included mild cognitive impairment (MCI; n = 120), cognitively impaired without MCI (n = 10), amnestic multidomain dementia (amnestic; n = 61), posterior cortical atrophy (PCA; n = 16), primary progressive aphasia (PPA; n = 68), frontotemporal dementia (FTD; n = 64), Lewy body spectrum diseases (LBSD; n = 74), and non‐amnestic multidomain dementia (n = 48). In analyses, PCA, PPA, FTD, LBSD, and non‐amnestic multidomain dementia were all considered non‐amnestic syndromes (n = 270). Of the NormCog participants, six eventually progressed to a diagnosis of MCI, one to FTD, and one to vascular cognitive impairment.

2.2. Plasma collection and analysis

Whole blood samples were collected in EDTA tubes and spun down, and plasma aliquots were stored at −80°C according to standard procedures. ²⁵ Plasma samples were analyzed on the Quanterix single‐molecule array (Simoa) HD‐X automated immunoassay platform. Samples were analyzed in duplicate and quantified using the V2 Advantage kit reagents for p‐tau₁₈₁, ²⁷ the Discovery kit reagents for GFAP, ²⁸ and the NF‐Light Advantage kit reagents. ²⁹ For patients with more than one plasma timepoint, plasma sample was selected closest to biomarker confirmation (CSF or PET) or earliest/baseline timepoint.

2.2.1. Missing data

All participants had plasma p‐tau₁₈₁; 99 were missing plasma NfL (33 NormCog; 66 ImpCog). Two plasma outliers were excluded for measures > 5 standard deviations (SDs) above the mean: GFAP = 979.9 pg/mL from an ImpCog Aβ+ participant (8.9 SD), and plasma NfL = 386.8 pg/mL from an ImpCog Aβ‐ participant (17.3 SD).

RESEARCH IN CONTEXT

1. Systematic review: We reviewed literature (PubMed and Google Scholar) testing plasma biomarkers for diagnosis and prognosis of AD. Most studies evaluate plasma analytes individually and in amnestic AD; however biomarkers must be evaluated in a clinically diverse population. We expand prior work by testing plasma associations with AD neuropathology across the heterogeneous spectrum of cognitive impairment, and test for differences in diagnostic accuracy by phenotype, sex, and race.

2. Interpretation: Combining plasma p‐tau₁₈₁, GFAP, and NfL improved detection of Aβ+ in both pre‐clinical (NormCog) and symptomatic phases (ImpCog). In ImpCog, plasma p‐tau₁₈₁ and GFAP correlated with Aβ and tau (autopsy and PET), whereas plasma NfL was higher in non‐AD pathology. Plasma NfL well predicted future cognitive decline in both Aβ+ and Aβ‐ cases.

3. Future directions: Studies should test additional plasma biomarkers (e.g., p‐tau₂₁₇, Aβ₄₂/Aβ₄₀) and if combining biomarkers can predict progression to cognitive impairment in Aβ+ NormCog cases (i.e., pre‐clinical AD).

2.3. Aβ status

Aβ status (+/−) was determined by autopsy, or by closest available CSF/PET biomarkers when autopsy data were not available. Aβ+ was defined by Consortium to Establish a Registry for AD (CERAD) ≥ 2 ³⁰ for autopsy, CSF Aβ₄₂/Aβ₄₀ ≤ 0.07, ³¹ CSF Aβ₄₂ ≤ 192, ³² or ¹⁸F‐florbetaben mean neocortical standardized uptake value ratio (SUVR) ≥ 1.097. All other participants were classified as Aβ‐.Some participants had two or more Aβ biomarkers available (n = 115; 19%), and 21 (4%) had conflicting autopsy, CSF, or PET results (1 NormCog; 20 ImpCog). Cases with conflicting Aβ biomarkers are outlined in Table S1.

The ¹⁸F‐florbetaben SUVR cutoff was calculated as the optimal value (Youden's index) distinguishing positive versus negative visual reads performed by expert readers in a partially overlapping dataset. CSF or PET was within 1 year of plasma collection (mean = 1.8 months; SD = 2.9).

2.3.1. Neuropathological diagnoses and severity assessments

Brains (14 NormCog; 88 ImpCog) were autopsied and assessed for ADNC at the Penn Center for Neurodegenerative Disease Research according to standard procedures. ²⁵ , ³³ Neuritic plaques were scored according to the Consortium to Establish a Registry for AD (CERAD) protocol. ³⁰ Individuals with CERAD ≥ 2 were Aβ+ (5 NormCog; 48 ImpCog), and with CERAD ≤ 1 were Aβ‐ (9 NormCog; 40 ImpCog). ADNC was scored according to ABC criteria, ³⁴ and Braak stage determined tau spread. ³⁵ There was an average of 4.1 years between plasma collection and death (SD = 2.9).

In addition, brains were assessed for other proteinopathies, ²⁵ including frontotemporal lobar degeneration (FTLD) pathology ³⁶ , ³⁷ and α‐synuclein (αSyn) by DLB stage, ³⁸ and for vascular disease. ³⁹ Cases with clinically relevant levels of multiple pathologies (e.g., intermediate/high ADNC, brainstem/limbic/neocortical DLB stage) were considered mixed pathology; otherwise, copathology was considered negligible.

Brain tissue samples were stained using immunohistochemistry, ²⁵ and gross severity of pathological accumulations of post mortem Aβ and tau were scored using a semi‐quantitative scale (0 = none, 0.5 = rare, 1 = minimal, 2 = moderate, 3 = severe). Post mortem Aβ and tau burden scores were averaged across limbic and neocortical regions standardly sampled ³⁴ : amygdala, cingulate, CA1/subiculum, entorhinal cortex, middle frontal gyrus, angular gyrus, and superior/middle temporal gyrus. Hemisphere was randomized; if both hemispheres were sampled, the average was taken.

2.3.2. CSF

Aβ status was determined using CSF for 23 NormCog and 299 ImpCog individuals. CSF samples were collected following a standard lumbar puncture preceded by an overnight fast, then processed and stored at −80°C according to standard procedures. ²⁵ , ⁴⁰ Samples were assayed prospectively. Fujirebio Lumipulse platform using Lumipulse kits (Fujirebio, Ghent, Belgium) quantified CSF Aβ₄₂/Aβ₄₀ concentrations (n = 129), and xMAP Luminex platform (INNO‐BIA AlzBio3 for research‐only reagents; Innogenetics) quantified CSF Aβ₄₂ (n = 193).

2.3.3. Aβ PET

Aβ status was determined using ¹⁸F‐florbetaben amyloid PET for 95 NormCog and 74 ImpCog individuals. Participants received approximately 8.1 mCi of ¹⁸F‐florbetaben intravenously, and image data were acquired in four 5‐minute frames beginning approximately 90 min post‐injection. Signal attenuation correction was performed using a low‐dose CT scan, and corrections were applied for scatter and head motion. The four volumes were aligned and averaged, and this average image was spatially registered to the individual's T1‐weighted MR image using a rigid‐body transform computed using Advanced Normalization Tools (ANTs) ⁴¹ and resampled to the T1 image resolution with linear interpolation. This T1‐aligned mean image was divided by mean signal in a whole‐cerebellum reference region, eroded to prevent contamination from adjacent structures, to produce SUVR images. We then computed mean SUVR within the neocortex, defined by the ANTs six‐class tissue segmentation. ⁴²

2.4. Tau PET

A total of 200 participants had tau PET data (66 NormCog; 134 ImpCog). ¹⁸F‐flortaucipir tau PET data were acquired as previously described ⁴³ : participants received approximately 10 mCi of ¹⁸F‐flortaucipir intravenously and were imaged from 75 to 105 min post‐injection in six 5‐minute frames. Data were processed using the same pipeline as ¹⁸F‐florbetaben data (above); voxel‐wise SUVRs were calculated relative to an inferior cerebellar reference excluding the dentate nuclei. For each participant, we then calculated mean SUVR within a composite region of Brodmann area 35 (BA35) and entorhinal cortex, defined using the Automated Segmentation of Hippocampal Subfields T1‐weighted MRI pipeline. ⁴⁴

There was an average of 6.9 months between plasma collection and tau PET (SD = 9.5; max = 74.3).

2.5. Demographics and neuropsychological assessments

Age, sex, and race were collected by self‐report. Global cognition and functioning were assessed using the Mini‐Mental State Examination (MMSE) and Clinical Dementia Rating (CDR) Scale (both global and sum of boxes [CDR‐SB]). ⁴⁵ , ⁴⁶

2.6. Statistical analyses

Biomarkers and demographic variables were not normally distributed; Mann‐Whitney‐Wilcoxon and Kruskal‐Wallis tests compared continuous variables and chi‐squared compared categorical variables. Plasma analytes were log‐transformed in parametric models with covariates; unadjusted Mann‐Whitney‐Wilcoxon comparisons and Spearman's rho are summarized in figures. Missing data were dropped listwise from models. The β‐estimates, 95% confidence intervals (95% CIs), and p‐values are reported from models. Statistical tests were performed with a significance threshold of α = 0.05. Analyses were conducted using R version 4.3.1 (2023‐06‐16) software and the cutpointr, ⁴⁷ survival, ⁴⁸ leaps, ⁴⁹ lmerTest, ⁵⁰ and effectsize ⁵¹ packages.

2.6.1. Part 1: Plasma biomarkers to detect Aβ in NormCog

Part 1 evaluates plasma biomarkers in NormCog participants. Linear models compared biomarkers across Aβ status (Aβ+, Aβ‐). Models covaried for factors that might affect plasma concentrations, age and sex, and residual error (ε) (Equation 1). Effect sizes with 95% CI were calculated using generalized η ^{2 51} (η ² _G interpretation: ≥0.01 small, ≥0.06 medium, ≥0.14 large ⁵² ).

$$\begin{array}{c}\hfill \log \left(Analyte\right)={\beta }_0+{\beta }_{1}A\beta status+{\beta }_{2}AgeatPlasma+{\beta }_{3}Sex+\epsilon \end{array}$$ (1)

To discriminate Aβ+ from Aβ‐, optimal multivariable biomarker combinations were determined using data‐driven exhaustive selection ⁴⁹ and k‐folds cross‐validation (k = 5). Mallow's C _p and Bayesian information criterion (BIC) compared model fitness, and multivariable models with the lowest C _p were tested. Receiver operating characteristic (ROC) analyses with bootstrapping (2000 iterations) tested diagnostic accuracy ⁴⁷ ; area under the curve (AUC) with 5%−95% CI were reported. Youden's index determined the threshold that maximized sensitivity and specificity. To reduce sample bias in comparisons of diagnostic accuracy, ROC analyses were restricted to individuals with no missing plasma biomarker data (complete data for p‐tau₁₈₁, GFAP, and NfL). To test if diagnostic accuracy differed by phenotype, sex, or race, chi‐squared tests compared correct classifications (true positives/negatives) and errors (false positives/negatives) of multivariable models and the best performing analyte using the Youden's index cutpoint. Accuracy was calculated as (Total—Errors)/Total × 100.

Next, linear models tested tau PET SUVR as a function of plasma analytes, covarying for age, interval between collection, and sex (Equation 2). We also tested for interactions between analyte concentrations and Aβ status (Equation 3).

$$\begin{array}{ccc}\mathrm{Tau}\mathrm{PET}& =& {\beta }_0+{\beta }_1\log (\mathrm{Analyte})+{\beta }_2\mathrm{Age}\mathrm{at}\mathrm{PET}\\ & & {+\beta }_3\mathrm{Plasma}\mathrm{to}\mathrm{PET}+{\beta }_4\mathrm{Sex}+\epsilon \end{array}$$ (2)

$$\begin{array}{ccc}\mathrm{Tau}\mathrm{PET}& =& {\beta }_0+{\beta }_1\log \left(\mathrm{Anal}\mathrm{yte}\right)\times {\beta }_{2}A\beta \mathrm{Status}+{\beta }_3\mathrm{Age}\mathrm{at}\mathrm{PET}\\ & & {+\beta }_4\mathrm{Plasma}\mathrm{to}\mathrm{PET}+{\beta }_5\mathrm{Sex}+\epsilon \end{array}$$ (3)

2.6.2. Part 2: Plasma biomarkers to detect Aβ in ImpCog

Part 2 repeated analyses from Part 1 in ImpCog individuals, including comparing plasma concentrations by Aβ status (Equation 1), ROC analyses to stratify Aβ+ from Aβ‐, and associations of plasma analytes with tau PET (Equations 2 and 3). We also tested associations with post mortem Aβ burden (Equation 4) and tau burden (Equation 5).

$$\begin{array}{ccc}\mathrm{Post}\mathrm{mort}\mathrm{em}A\beta & =& {\beta }_0+{\beta }_1\log \left(\mathrm{Analyte}\right)+{\beta }_2\mathrm{Age}\mathrm{at}\mathrm{Death}\\ & & +{\beta }_3\mathrm{Plasma}\mathrm{to}\mathrm{Death}+{\beta }_4\mathrm{Sex}+\epsilon \end{array}$$ (4)

$$\begin{array}{ccc}\mathrm{Postmortem}\mathrm{Tau}& =& {\beta }_0+{\beta }_1\log (\mathrm{Analyte})+{\beta }_2\mathrm{Age}\mathrm{at}\mathrm{Death}\\ & & {+\beta }_3\mathrm{Plasma}\mathrm{to}\mathrm{Death}+{\beta }_4\mathrm{Sex}+\epsilon \end{array}$$ (5)

To confirm generalizability of biomarker performance, cross‐validation split the ImpCog cohort into training and test sets (50%/50% split). The multivariable model was trained and ROC analyses determined best thresholds using Youden's index in the training set. Models and thresholds were then applied to the independent test set to test performance.

2.6.3. Part 3: Plasma biomarkers to predict cognitive decline in Aβ+ individuals

For Aβ+ individuals with longitudinal MMSE (n = 159) and CDR‐SB (n = 231), part 3 tests how plasma biomarkers related to future cognitive decline in Aβ+ individuals (NormCog and ImpCog). Linear mixed effects models tested how global cognition and function (MMSE Equation 6; CDR‐SB Equation 7) changed over time by plasma analyte concentration at baseline (interaction term: Analyte × Years), with a random intercept for individual (γ). Models covaried for age at baseline/plasma collection, sex, years of education, APOE ε4 (dominant: 0 vs. 1+ alleles), and MMSE/CDR‐SB score at baseline. In longitudinal analyses, all individuals had 2 or more timepoints. To reduce potential bias and high leverage of individuals with slow progression/extended survival, longitudinal cognitive assessments were within 5 years of baseline plasma.

$$\begin{array}{ccc}& & MMSE={\beta }_0+{\beta }_{1}Years\times {\beta }_2\log \left(Analyte\right)+{\beta }_{3}Age+{\beta }_{4}Sex\hfill \\ & & +{\beta }_{5}Education+{\beta }_{6}APOE\epsilon 4+{\beta }_{7}MMSEatBaseline+\gamma +\epsilon \hfill \end{array}$$ (6)

$$\begin{array}{ccc}& & CDR-SB={\beta }_0+{\beta }_{1}Years\times {\beta }_2\log \left(Analyte\right)+{\beta }_{3}Age+{\beta }_{4}Sex\hfill \\ & & +{\beta }_{5}Education+{\beta }_{6}APOE\epsilon 4+{\beta }_{7}CDR-SBatBaseline+\gamma +\epsilon \hfill \end{array}$$ (7)

Survival analyses and Cox proportional‐hazards regression tested how plasma biomarkers at baseline predicted future clinical decline, with survival defined as years to global CDR progression (i.e., earliest increase in Global CDR score: 0 > 0.5 > 1 > 2 > 3). Optimal multivariable biomarker combinations were determined using exhaustive selection and five‐folds cross‐validation. Mallow's C _p and Bayesian information criterion (BIC) were used to compare model fitness. Cox proportional‐hazards models tested how each analyte and the multivariable model predicted time to CDR progression (t), based on baseline hazard [h₀(t)], and covarying for age at baseline, sex, and education. Individuals who never experienced the event (no increase in Global CDR score at follow‐ups) were right‐censored at last follow‐up. Hazard ratios (HRs), 95% CIs, and p‐values are reported. For significant analytes/models, we verified that proportional hazards assumptions were met (p > 0.05).

$$\begin{array}{ccc}\hfill h\left(t\right)& =& h_0\left(t\right)\times exp(b_{1}log\left(Analyte\right)+b_{2}Age+b_{3}Sex+b_{4}Education\hfill \\ & & +b_{5}GlobalCDRatBaseline)\hfill \end{array}$$ (8)

2.6.4. Part 4: Plasma biomarkers to predict cognitive decline in Aβ‐ ImpCog individuals

Part 4 repeated analyses from Part 3 in ImpCog Aβ‐ individuals with longitudinal MMSE (n = 116) and CDR‐SB (n = 165); Aβ‐ NormCog individuals were excluded due to no evidence of neurodegenerative disease. Linear mixed effects models related changes in MMSE (Equation 6) and CDR‐SB (Equation 7) to plasma concentrations at baseline, and survival analyses tested how plasma analytes and multivariable Cox proportional‐hazards model predicted future CDR progression in ImpCog Aβ‐ individuals (Equation 8).

3. RESULTS

Table 1 summarizes the clinical and demographic characteristics of participants. NormCog participants were older at plasma collection (W = 39741; p = 7.9e‐08), had more females (χ ²(1) = 12; p = 0.00055), and fewer non‐Hispanic White participants than ImpCog (χ ²(5) = 25; p = 0.00011). Within NormCog, Aβ+ were older than Aβ‐ (W = 500.5; p = 2e‐04), but there was no difference in sex (p = 0.19) or racial distribution (p = 0.94) by Aβ status. Likewise within ImpCog, there was no difference in age (p = 0.34), sex (p = 0.98), or race (χ ²(5) = 11; p = 0.059) by Aβ status. Table S2 details the available neuropathological data for autopsied ImpCog.

TABLE 1.

Demographics of normal cognition (NormCog) and impaired cognition (ImpCog) participants by Aβ status.

Parameter	NormCog:Aβ‐	NormCog:Aβ+	ImpCog:Aβ‐	ImpCog:Aβ+	p‐Value
N	113	19	233	228
Age at plasma (years)	70.0 [67.0, 75.0]	77.0 [73.5, 85.5]	67.0 [60.0, 72.0]	65.0 [59.0, 72.0]	<0.001
Global CDR (0–3)	0.0 [0.0, 0.0]	0.0 [0.0, 0.0]	0.5 [0.5, 1.0]	0.5 [0.5, 1.0]	<0.001
APOE ε4 (%)					<0.001
0	89 (78.8%)	10 (52.6%)	164 (75.2%)	119 (52.7%)
1	18 (15.9%)	7 (36.8%)	54 (24.8%)	81 (35.8%)
2	6 (5.3%)	2 (10.5%)	0 (0.0%)	26 (11.5%)
NACC clinical diagnosis (%)					<0.001
Normal	113 (100.0%)	19 (100.0%)	0 (0.0%)	0 (0.0%)
Impaired without MCI	0 (0.0%)	0 (0.0%)	8 (3.4%)	2 (0.9%)
MCI	0 (0.0%)	0 (0.0%)	48 (20.6%)	72 (31.6%)
Amnestic dementia	0 (0.0%)	0 (0.0%)	3 (1.3%)	58 (25.4%)
PCA	0 (0.0%)	0 (0.0%)	2 (0.9%)	14 (6.1%)
FTD	0 (0.0%)	0 (0.0%)	44 (18.9%)	20 (8.8%)
PPA	0 (0.0%)	0 (0.0%)	47 (20.2%)	21 (9.2%)
LBSD	0 (0.0%)	0 (0.0%)	51 (21.9%)	23 (10.1%)
Nonamnestic dementia	0 (0.0%)	0 (0.0%)	30 (12.9%)	18 (7.9%)
Sex: Male (%)	45 (39.8%)	4 (21.1%)	128 (54.9%)	124 (54.4%)	0.002
Self‐reported Race (%)					0.002
American Indian/Alaska Native	0 (0.0%)	0 (0.0%)	1 (0.4%)	0 (0.0%)
Asian	1 (0.9%)	0 (0.0%)	5 (2.1%)	4 (1.8%)
Black/African American	20 (17.7%)	3 (15.8%)	16 (6.9%)	5 (2.2%)
Hispanic White	1 (0.9%)	0 (0.0%)	3 (1.3%)	0 (0.0%)
More than one race	2 (1.8%)	0 (0.0%)	4 (1.7%)	2 (0.9%)
White	89 (78.8%)	16 (84.2%)	204 (87.6%)	214 (95.1%)

Note: For continuous variables, median and interquartile range (IQR) are reported; Wilcoxon tests performed group comparisons. For categorical variables, count (percentage [%]) are provided; chi‐squared tests performed frequency comparisons. p‐values are reported for group comparisons.

Abbreviations: APOE, apolipoprotein E; Global CDR, Clinical Dementia Rating; FTD, frontotemporal dementia; MCI, mild cognitive impairment; NACC, National Alzheimer's Coordinating Center; PCA, posterior cortical atrophy; PPA, primary progressive aphasia; LBSD, Lewy body spectrum diseases.

3.1. Part 1: Plasma biomarkers to detect Aβ in NormCog

3.1.1. Stratification of Aβ+ from Aβ‐

We tested how each plasma analyte associated with Aβ status in NormCog; unadjusted comparisons are summarized in Figure 1A. Linear models confirmed differences in NormCog individuals after covarying for age and sex: Aβ+ was associated with higher plasma GFAP (β = 0.29; 95% CI: [0.076,0.5]; p = 0.0082) with large effect (η ² _G  = 0.20), and with higher p‐tau₁₈₁ (β = 0.23; 95% CI: [0.024,0.44]; p = 0.029) with medium effect (η ² _G  = 0.098) compared to Aβ‐; there was no difference in NfL levels by Aβ status (p = 0.75).

FIGURE 1.

Plasma biomarkers by Aβ status: normal cognition. (A) Boxplots show median, interquartile range (IQR), and outliers for analytes. Color indicates Aβ status. Asterisks represent p‐values from Wilcoxon pairwise comparisons (*** p < 0.001, **** p < 0.0001, and not significant [ns]). (B) ROC curves discriminating Aβ+ from Aβ‐. Color indicates biomarker; line indicates multivariable (solid) or univariable (broken) model type. ROC, receiver operator characteristic.

Examining covariates, older age was significantly associated with higher GFAP (β = 0.024; 95% CI: [0.018,0.03]; p = 4.5e‐13), p‐tau₁₈₁ (β = 0.015; 95% CI: [0.0091,0.021]; p = 1.1e‐06), and NfL (β = 0.033; 95% CI: [0.027,0.038]; p = 1.9e‐19). Males had lower GFAP than females (β = −0.22; 95% CI: [−0.37,−0.068]; p = 0.0046); sex had no significant association for p‐tau₁₈₁ (p = 0.27) or NfL (p = 0.21).

ROC analyses were performed in participants with complete plasma data (NormCog n = 99). Model comparison using five‐folds cross‐validation showed that a model comprising all three biomarkers had the best fit based on lowest Mallow's C _p (C _p = 4.0) compared to models with either two (C _p = 6.8) or one biomarkers (C _p = 8.4), although models with one biomarker had the lowest BIC (−1.4) compared to two (BIC = −0.5) or three biomarkers (BIC = −1.0). In NormCog, Multivariable logistic regression demonstrated that Aβ+ was associated with higher p‐tau₁₈₁ (β = 3.5; 95% CI: [1.1,6.3]; p = 0.0084) and GFAP (β = 3.5; 95% CI: [1.6,5.9]; p = 0.0013), and lower NfL (β = −2.5; 95% CI: [−4.9,−0.58]; p = 0.019).

ROC analyses in NormCog tested performance of each plasma analyte and multivariable model to stratify Aβ+ from Aβ‐ (Figure 1B, Table 2A), and bootstrapping tested for AUC differences between the multivariable model and individual analytes. The multivariable model had the highest AUC (0.87), although this did not reach significance compared to GFAP (AUC = 0.81; p = 0.063), or p‐tau₁₈₁ (AUC = 0.78; p = 0.11), due in part to the small number of Aβ+ NormCog cases. The multivariable model AUC was significantly higher than NfL (AUC = 0.60; p = 6.3e‐04).

TABLE 2.

Receiver operating characteristic (ROC) analyses to stratify Aβ+. from Aβ‐ using biomarkers.

A. Normal cognition	AUC	CI (5%)	CI (95%)	Threshold	Sensitivity	Specificity	Accuracy
Multivariable: p‐tau+GFAP+NfL	0.87	0.80	0.94	0.17	0.87	0.79	0.80
Plasma GFAP	0.81	0.71	0.89	148.69	0.76	0.73	0.73
Plasma p‐tau	0.78	0.68	0.86	2.20	0.85	0.63	0.66
Plasma NfL	0.60	0.49	0.71	16.02	0.73	0.47	0.50

B. Impaired cognition	AUC	CI (5%)	CI (95%)	Threshold	Sensitivity	Specificity	Accuracy
Multivariable: p‐tau+GFAP+NfL	0.87	0.84	0.90	0.49	0.76	0.83	0.80
Plasma p‐tau	0.84	0.81	0.87	2.64	0.76	0.78	0.77
Plasma GFAP	0.75	0.71	0.79	138.87	0.74	0.64	0.68
Plasma NfL	0.55	0.50	0.60	33.79	0.82	0.42	0.61

Note: ROC analyses in (A) NormCog and (B) ImpCog. ROC metrics are calculated using bootstrapping with 2000 iterations. Analytes and multivariable models are listed in descending order of area under the curve (AUC) and the 5%−95% confidence interval (CI) for AUC. Best threshold (determined by Youden's index) and associated sensitivity, specificity, and accuracy are reported.

Abbreviations: AUC, area under the curve; CI, confidence interval; GFAP, glial fibrillary acidic protein; Nfl, neurofilament light chain; p‐tau, plasma phosphorylated tau.

To test if diagnostic errors differed by patient characteristics, we compared multivariable model and GFAP (next highest AUC) accuracy applying Youden's index cutpoints (Table 2) by sex and race. Chi‐squared tests indicated that proportion of correct classifications (true positives/negatives) to errors (false positives/negatives) did not significantly differ by sex for either the multivariable model (p = 0.24) (Accuracy: males = 91%; females = 82%) or GFAP (p = 0.82) (Accuracy: males = 74%; females = 71%). The proportion of correct classification to errors did not significantly differ by race for either the multivariable model (p = 0.50) (Accuracy: Hispanic/non‐White = 90%; non‐Hispanic Whites = 83%) or GFAP (p = 0.78) (Accuracy: Hispanic/non‐White = 76%; non‐Hispanic Whites = 71%).

3.1.2. Tau PET correlations with plasma biomarkers

We tested associations of plasma analytes with tau PET SUVR in NormCog (Figure 2). In unadjusted models, the only significant correlation was between plasma NfL and tau PET SUVR (Figure 2), but this did not remain significant after linear models covaried for age, plasma‐to‐PET interval, and sex (p = 0.23).

FIGURE 2.

Correlations of plasma with tau PET: normal cognition. Scatterplots of each plasma analyte by tau PET SUVR. Lines plot least squares regression by Aβ status for tau PET. Color indicates Aβ+ (red) or Aβ‐ (blue). Spearman's rho across all participants (Aβ+ and Aβ‐) and p‐values are reported. PET, positron emmission tomography; SUVR, standard uptake value ratio.

Models with tau PET were repeated testing interactions with Aβ status. The association of plasma p‐tau₁₈₁ with tau PET SUVR was significantly stronger in Aβ+ than Aβ‐ (β = 2.5; 95% CI: [0.3,4.7]; p = 0.027); neither GFAP nor NfL had a significant interaction (both p > 0.32).

3.2. Part 2: Plasma biomarkers to detect Aβ in ImpCog

3.2.1. Stratification of Aβ+ from Aβ‐

We tested how each plasma analyte associated with Aβ status in ImpCog participants; unadjusted comparisons are summarized in Figure 3A. Linear models confirmed differences after covarying for age and sex. In Aβ+ individuals, plasma p‐tau₁₈₁ was higher than Aβ‐ (β = 0.62; 95% CI: [0.54,0.70]; p = 1.6e‐40) with large effect (η ² _G  = 0.32), as was GFAP (β = 0.45; 95% CI: [0.37,0.54]; p = 1.3e‐22) with large effect (η ² _G  = 0.19). NfL was not significantly different by Aβ status (β = −0.099; 95% CI = [−0.21,0.013] p = 0.084).

FIGURE 3.

Plasma biomarkers by Aβ status: impaired cognition. (A) Boxplots show median, interquartile range (IQR), and outliers for analytes. Color indicates clinical severity (Global CDR). Asterisks represent p‐values from Wilcoxon pairwise comparisons (**** p < 0.0001 and not significant [ns]). (B) ROC curves discriminating Aβ+ from Aβ‐. Color indicates biomarker; line indicates multivariable (solid) or univariable (broken) model type. CDR, Clinical Dementia Rating; ROC, receiver operator characteristic.

Examining covariates, older age was significantly associated with higher NfL (β = 0.013; 95% CI: [0.0061,0.019]; p = 0.00016), GFAP (β = 0.01; 95% CI: [0.0054,0.015]; p = 3.7e‐05), and p‐tau₁₈₁ (β = 0.0053; 95% CI: [0.00061,0.0099]; p = 0.027). Males had lower GFAP (β = −0.18; 95% CI: [−0.26,−0.089]; p = 7.7e‐05) and NfL than females (β = −0.16; 95% CI: [−0.28,−0.051]; p = 0.0044); sex had no significant association for p‐tau₁₈₁ (p = 0.85). 

FIGURE 4.

Correlations of plasma biomarkers with tau and Aβ: impaired cognition. Scatterplots of each plasma analyte by (A) tau PET SUVR, (B) post mortem tau burden, (C) post mortem Braak Stage (0‐6), and (D) post mortem Aβ burden. Lines plot least squares regression by Aβ status for tau PET, and overall for post mortem burden for Aβ and tau (black). Color indicates Aβ+ (red) or Aβ‐ (blue). Spearman's rho and p‐values across all participants (Aβ+ and Aβ‐) are reported. PET, positron emmission tomography; SUVR, standard uptake value ratio.

Because plasma levels may be affected by co‐morbidities, we tested for differences between subgroups of ImpCog participants with available autopsy data (see Table S2). While we saw no evidence of plasma differences by vascular disease (Figure S1) or by other concomitant pathologies (Figure S2) in this cohort, we note the limited sample size.

ROC analyses were performed in participants with complete plasma data (ImpCog n = 394). Model comparion using five‐folds cross‐validation showed that a model with all three biomarkers (C _p = 4.0; BIC = −136.7) had consistently better fit than models with either one (C _p = 36.7; BIC = −113.1) or two biomarkers (C _p = 22.3; BIC = −122.6). In ImpCog, multivariable logistic regression demonstrated that Aβ+ was associated with higher p‐tau₁₈₁ (β = 2.6; 95% CI: [2.0,3.2]; p = 1.7e‐15) and GFAP (β = 1.6; 95% CI: [1.0,2.3]; p = 5.7e‐07), and lower NfL (β = −1.1; 95% CI: [−1.6,−0.63]; p = 9.6e‐06).

ROC analyses tested performance of each plasma analyte and multivariable model (p‐tau₁₈₁ + GFAP + NfL) to stratify Aβ+ from Aβ‐ (Figure 3B, Table 2B). In ImpCog, the multivariable model had the best performance (AUC = 0.87), with significantly higher AUC than p‐tau₁₈₁ (AUC = 0.84; p = 0.014), GFAP (AUC = 0.75; p = 2.7e‐08), and NfL (AUC = 0.55; p = 4.2e‐22).

Using Youden's index cutpoints (Table 2), chi‐squared tests compared correct classifications to errors of the multivariable model and the next best analyte, p‐tau₁₈₁, across clinical diagnoses (amnestic, MCI/impaired without MCI, and all other non‐amnestic [non‐amnestic, PCA, FTD, LBSD, PPA]). Multivariable model classification differed by phenotype (χ ² = 7; p = 0.033), with highest accuracy in amnestic dementia (92%), followed by MCI/impaired without MCI (81%), and then non‐amnestic dementia (76%). Likewise, plasma p‐tau₁₈₁ performance significantly differed by phenotype (χ ² = 6.8; p = 0.041) with highest accuracy in amnestic dementia (90%), followed by MCI/impaired without MCI (80%), and then non‐amnestic dementia (74%). Classifications did not differ by sex for either the multivariable model (χ ² = 3.1, p = 0.079) (Accuracy: female = 83%; male = 76%) or plasma p‐tau₁₈₁ (χ ² = 3.1, p = 0.10) (Accuracy: female = 81%; male = 74%). Classifications did not significantly differ by self‐identified race for either the multivariable model (p = 0.13) (Accuracy: Hispanic/non‐White = 89%; non‐Hispanic White = 78%) or plasma p‐tau₁₈₁ (p = 0.33) (Accuracy: Hispanic/non‐White = 84%; non‐Hispanic White = 76%).

3.2.2. Cross‐validation in independent training and test sets

To increase rigor and ensure that our multivariable model performance is generalizable, we performed cross‐validation (50%/50% random split, stratified by Aβ status) of ImpCog cases into independent training and test sets. We trained the multivariable logistic regression in the training set. Consistent with results above, we observed that Aβ+ was associated with higher p‐tau₁₈₁ (β = 3.5; 95% CI: [2.5,4.7]; p = 1.2e‐09) and GFAP (β = 1.6; 95% CI: [0.63,2.7]; p = 0.0019), and lower NfL (β = −1; 95% CI: [−1.8,−0.37]; p = 0.0033). Examining performance in the independent test set, ROC analyses validated that the multivariable model had the best performance (AUC = 0.84), with significantly higher AUC than p‐tau₁₈₁ (AUC = 0.81; p = 0.019), GFAP (AUC = 0.75; p = 2.6e‐03), and NfL (AUC = 0.56; p = 9.7e‐09). See Table S3 for full ROC data.

3.2.3. Pathological correlates of plasma biomarkers

We tested associations of plasma analytes with tau PET SUVR in ImpCog (Figure 4). Linear models covarying for age at plasma and sex confirmed significant associations of tau PET SUVR (Figure 4A) with higher p‐tau₁₈₁ (β = 0.38, 95% CI: [0.30,0.46]; p = 1.8e‐16) and GFAP (β = 0.41; 95% CI: [0.32,0.50]; p = 3.2e‐15^‐15), and lower NfL (β = −0.11; 95% CI: [−0.22,−0.0017]; p = 0.046). Likewise, linear models covarying for plasma‐to‐death interval and sex confirmed significant associations of post mortem tau burden (Figure 4B) with higher p‐tau₁₈₁ (β = 0.66; 95% CI: [0.34,0.98]; p = 9.3e‐05) and GFAP (β = 0.72; 95% CI: [0.29,1.1]; p = 0.0013), but not NfL (p = 0.15). Post mortem Aβ plaque burden (Figure 4C) was also associated with higher p‐tau₁₈₁ (β = 1.0; 95% CI: [0.67,1.4]; p = 4e‐07) and GFAP (β = 0.79; 95% CI: [0.24,1.3]; p = 0.0053), and lower NfL (β = −0.63; 95% CI: [−1.1,−0.11] p = 0.017). Full models are reported in Tables S4 and S5.

Given significant interactions, we tested associations between tau PET SUVR and plasma analytes in Aβ‐ ImpCog cases only; none were significant (p‐tau₁₈₁: β = 0.055, 95% CI: [−7e‐04,0.11], p = 0.053; GFAP: β = 0.066; 95% CI [−0.011,0.14], p = 0.091; NfL: p = 0.91). To disentangle independent contributions of plasma analytes to predict Aβ and tau burden, posthoc linear models included all three biomarkers as predictors. Tau PET had independent associations with all three biomarkers, post mortem tau with GFAP and NfL, and post mortem Aβ with plasma p‐tau₁₈₁ (Table S6).

3.3. Part 3: Plasma biomarkers to predict cognitive decline in Aβ+ individuals

Baseline, cross‐sectional correlations within Aβ+ showed that worse CDR‐SB was correlated with higher plasma p‐tau₁₈₁ (rho = 0.13; p = 0.04), GFAP (rho = 0.13; p = 0.045), and NfL (rho = 0.18; p = 0.011). Worse MMSE was correlated with plasma higher p‐tau₁₈₁ (rho = −0.24; p = 0.0038), GFAP (rho = −0.29; p = 0.00058), and NfL (rho = −0.19; p = 0.037).

In Aβ+ participants, longitudinal analyses tested how each plasma biomarker at baseline predicted future cognitive decline (Figure 5). More severe CDR‐SB increase (Interaction: time × plasma analyte) was significantly associated with higher p‐tau₁₈₁ (β = 0.66; 95% CI: [0.3,1.0]; p = 0.00033) and NfL (β = 0.86; 95% CI: [0.44,1.3]; p = 8.9e‐05), but not GFAP (p = 0.26). Faster decline in MMSE (Interaction: time × plasma analyte) was significantly associated with higher p‐tau₁₈₁ (β = −1.6; 95% CI: [−2.2,−0.92]; p = 3.5e‐06) and GFAP (β = −0.61; 95% CI: [−1.2,−0.058]; p = 0.032), but not NfL (β = −0.69; 95% CI: [−1.4,0.032]; p = 0.061). Full models are summarized in Table S7. Survival analyses tested biomarkers in Aβ+ cases at baseline to predict future CDR change/progression (any increase in global CDR). Of the 200 Aβ+ cases, 96 (48%) had experienced CDR progression (the event) by the last follow‐up.

FIGURE 5.

Longitudinal cognitive decline by plasma at baseline: Aβ+ individuals. Spaghetti plots showing trajectories for (A) CDR‐SB and (B) MMSE by time. Higher CDR‐SB and lower MMSE indicated worse cognitive functioning. For visualization, plasma analyte levels were stratified in tertiles: high (red), medium (pink), and low (blue). Thin lines connect timepoints from an individual. Thick lines indicate mixed effects models. CDR, clinical dementia rating; CDR sum of boxes; MMSE, Mini‐Mental State Examination.

Model comparison using five‐folds cross‐validation showed that Cox proportional hazards models with two plasma biomarkers (C _p = 2.4) had better fit than models with either one (C _p = 4.8) or three biomarkers (C _p = 4.0); although BIC was lowest for models with one biomarker (BIC = −2.2) compared to two (BIC = −1.6) or three (BIC = 3.1). Exhaustive selection determined NfL and p‐tau₁₈₁ were the optimal biomarker combination to predict survival until CDR/state change: faster CDR progression was significantly associated with higher plasma NfL (HR = 2.6; 95% CI = [1.6,4.4], p = 0.00017), but not p‐tau₁₈₁ (HR = 1.2; 95% CI = [0.71,2.0], p = 0.49).

Cox proportional hazards tested each plasma analyte individually and the multivariable model (NfL+p‐tau₁₈₁) in Aβ+ cases at baseline to predict future CDR change/progression covarying for age at baseline/plasma, sex, and education. Faster time to CDR progression was associated with higher plasma NfL (HR = 3.6; 95% CI: [2.2,5.8]; p = 2.8e‐07), multivariable survival model (HR = 3.5; 95% CI: [2.2,5.7]; p = 1.3e‐07), and p‐tau₁₈₁ (HR = 2.2; 95% CI: [1.3,3.7]; p = 0.0027); GFAP was not associated with CDR progression (p = 0.34). Lower baseline Global CDR was significantly associated with faster progression for all models (p < 0.0026). No other covariates were significant (age all p > 0.079; sex all p > 0.62; education all p > 0.49).

3.4. Part 4: Plasma biomarkers to predict cognitive decline in ImpCog Aβ‐ individuals

Baseline, cross‐sectional correlations in ImpCog Aβ‐ showed that CDR‐SB was positively correlated with plasma NfL (rho = 0.25; p = 0.00027), but not GFAP (p = 0.19) or p‐tau₁₈₁ (p = 0.18). MMSE was inversely correlated with plasma GFAP (rho = −0.29; p = 0.00079) and NfL (rho = −0.54; p = 2.0e‐10), but not p‐tau₁₈₁ (p = 0.88).

In Aβ‐ participants, longitudinal analyses tested how each plasma biomarker at baseline predicted future cognitive decline (Figure 6). More severe CDR‐SB increase (Interaction: time × plasma analyte) was significantly associated with higher NfL (β = 0.91; 95% CI: [0.6,1.2]; p = 8.6e‐09) and lower p‐tau₁₈₁ (β = −0.44; 95% CI: [−0.84,−0.045]; p = 0.03), but not GFAP (p = 0.12). In Aβ‐, faster decline in MMSE (Interaction: time × plasma analyte) was significantly associated with higher NfL (β = −1.7; 95% CI: [−2.4,−1.0]; p = 6.6e‐07), but not GFAP (β = −0.85; 95% CI: [−1.8,0.096]; p = 0.079), or p‐tau₁₈₁ (β = 0.75; 95% CI: [−0.11,1.6]; p = 0.091). Full models are summarized in Table S8.

FIGURE 6.

Longitudinal cognitive decline by plasma at baseline: Aβ‐ individuals. Spaghetti plots showing trajectories for (A) CDR‐SB and (B) MMSE by time. Higher CDR‐SB and lower MMSE indication worse cognitive functioning. For visualization, plasma analyte levels were stratified in tertiles: high (red), medium (pink), and low (blue). Thin lines connect timepoints from an individual. Thick lines indicate mixed effects models. CDR, clinical dementia rating; CDR sum of boxes; MMSE, Mini‐Mental State Examination.

Survival analyses tested biomarkers in Aβ‐ cases at baseline to predict future CDR change/progression. Of 293 Aβ‐ cases, 92 (31%) had experienced CDR progression (the event) by the last follow‐up. Model comparison using five‐folds cross‐validation showed that models with only one plasma biomarker (C _p = 2.2; BIC = −16.0) had consistently better fit than either two (C _p = 2.3; BIC = −12.8) or three biomarkers (C _p = 4.0; BIC = −8.0). Plasma NfL was always the analyte selected. Therefore, survival analyses tested all analytes individually in Aβ‐ ImpCog cases.

Cox proportional hazards tested each plasma analyte in ImpCog Aβ‐ cases at baseline to predict future CDR change/progression (increase in global CDR) covarying for age at baseline/plasma, sex, and years of education. Plasma NfL was associated with faster time to CDR progression (HR = 3.3; 95% CI: [2.3,4.8]; p = 1.6e‐10), as was plasma GFAP (HR = 2.3, 95% CI: [1.5,3.7]; p = 3e‐04); plasma p‐tau₁₈₁ was not significant (p = 0.58).

Examining covariates, higher baseline global CDR was associated with faster CDR progression for the p‐tau₁₈₁ model (HR = 1.6, 95% CI: [1.2,2.2]; p = 0.0034), but not other models (GFAP: p = 0.053; NfL: p = 0.65). No other covariates were significant (age all p > 0.15; sex all p > 0.075; education all p > 0.067).

4. DISCUSSION

A growing body of literature shows that plasma p‐tau₁₈₁, GFAP, and NfL have diagnostic utility in AD. Even so, each have their diagnostic blind spots: plasma p‐tau₁₈₁ performs less well in early AD, ⁶ plasma GFAP does not have sufficient specificity, ¹⁶ and plasma NfL is highest in non‐AD. ⁷ The clinical spectrum of AD is broad, and biomarker accuracy and optimal strategy may differ by patient factors like disease stage and clinical phenotype (amnestic vs. non‐amnestic). We hypothesized that combining biomarkers may maximize diagnostic and prognostic accuracy in heterogeneous cases, and our head‐to‐head comparisons robustly tested individual analytes and multivariable models in a clinically diverse cohort. Cross‐validation confirmed that multivariable models best discriminated Aβ+ from Aβ‐ in both NormCog and ImpCog. For prognosis, longitudinal mixed effects models and survival analyses consistently indicated that plasma NfL at baseline had high utility to predict future clinical decline in both Aβ+ and Aβ‐ cases.

Our finding that multivariable models combining p‐tau₁₈₁, GFAP, and NfL can optimize diagnosis of AD pathology is supported by previous work. ¹ , ⁴ While plasma p‐tau₁₈₁ alone shows excellent discrimination of high ADNC from not/low/intermediate ADNC, ⁵³ , ⁵⁴ , ⁵⁵ the lower sensitivity of plasma p‐tau₁₈₁ to intermediate ADNC ⁸ , ⁹ may help explain its lower performance in NormCog than ImpCog that we observe here. In NormCog, we see improved performance of the multivariable model compared to p‐tau₁₈₁ (AUC of 0.87 vs. 0.78), although this did not reach significance in part due to small number of Aβ+. In the larger ImpCog sample, the multivariable model showed significantly better performance (AUC = 0.87; Table 2) than p‐tau₁₈₁ alone (AUC = 0.84); this finding was confirmed by cross‐validation in independent training and test sets. Pathological correlations support the conclusion that each plasma analyte may have independent associations with AD pathology. In ImpCog, both plasma p‐tau₁₈₁ and GFAP showed positive correlations with tau PET SUVR, post mortem Aβ burden, and post mortem tau burden. While plasma NfL is elevated in AD, ¹⁹ here NfL showed inverse associations with tau PET SUVR and post mortem Aβ burden due to its high sensitivity to non‐AD neurodegeneration, especially FTLD. ⁷ , ⁵⁶ Thus, plasma NfL may augment multivariable models by helping to normalize for non‐AD influences on plasma GFAP or p‐tau₁₈₁. All three analytes demonstrated independent and significant associations with tau PET SUVR; for post mortem Aβ burden, only plasma p‐tau₁₈₁ showed a significant independent association, while both plasma GFAP and NfL demonstrated significant, independent associations with post mortem tau burden. Prior work finds independent associations of both Aβ and tau pathology to plasma p‐tau₁₈₁ levels, while GFAP was more strongly associated with tau than Aβ. ³ Still, many other autopsy and PET studies find that plasma GFAP is more strongly associated with Aβ pathology than tau, ⁸ , ¹² , ⁵⁷ or with both. ⁵⁸ Differences in findings may be due to differences in sample sizes, measurement modality (autopsy, PET), regions quantified, and/or relatively coarse semi‐quantitative histopathological measurements; pathological contributions to plasma concentrations will have to be further disentangled by future studies. Nonetheless, we and others ⁴ clearly find that plasma analytes demonstrate independent, complementary contributions to predicting AD pathology. In sum, plasma GFAP may augment detection of AD pathology with plasma p‐tau₁₈₁, especially in early AD, while plasma NfL may help normalize for alterations due to non‐AD pathology.

Beyond pathological associations of each biomarker, we tested how patient factors influenced diagnostic performance, given the unique clinical heterogeneity of our cohort. Considering the limited clinicopathological correlations due to clinical variation in AD, ⁵⁹ , ⁶⁰ , ⁶¹ plasma biomarkers must discriminate of AD from diverse non‐AD pathologies (e.g., FTLD, αSyn). In our ImpCog cohort, 42% of Aβ+ individuals had a non‐amnestic syndrome. This is in contrast to most other studies of plasma biomarkers, where biomarker‐confirmed AD are largely/all amnestic. ⁶² , ⁶³ Previously, we observed that CSF biomarkers, specifically CSF p‐tau₁₈₁, were less accurate to detect ADNC for non‐amnestic AD, compared to amnestic. ⁶⁴ Here too, we found that plasma biomarkers were less accurate to identify Aβ+ for non‐amnestic than amnestic or MCI subgroups. In addition to disease stage and clinical phenotype, we tested if biomarker accuracy differed across race and sex. Although studies have found similar plasma p‐tau₁₈₁ levels in females and males, they indicate that sex may impact interpretation of plasma p‐tau₁₈₁. ⁶⁵ Likewise, similar p‐tau₁₈₁ levels across racial groups have also been reported. ⁵³ , ⁶⁶ Higher levels of plasma GFAP have been consistently reported in females than males, ² , ¹⁶ , ⁵⁶ while findings are mixed for sex differences for plasma NfL. ² , ¹⁹ , ⁵⁶ Others have found no difference in plasma biomarker levels across racial groups, ⁶⁷ especially after accounting for differences in kidney renal dysfunction. ⁶⁸ When testing the accuracy of the multivariable models by sex and race, there was no diagnostic difference in NormCog; in ImpCog, the multivariable model was more accurate in females than males, but accuracy did not differ by race. Still, our sample was limited in racial/ethnic diversity and future follow‐up studies will need to evaluate model accuracy in different racial groups.

Regarding prognosis, plasma NfL at baseline was consistently associated with future cognitive decline, with higher NfL associated with worse longitudinal MMSE, CDR‐SB, and faster CDR‐progression in both Aβ+ and Aβ‐ cases. This is in line with prior findings that NfL associates with worse cognition in both AD and non‐AD neurodegenerative diseases. ⁹ , ¹⁹ , ⁶⁹ , ⁷⁰ , ⁷¹ Unlike diagnosis, we found little evidence that multivariable models improved survival estimates over NfL alone in either Aβ+ or Aβ‐. Exhaustive selection indicated that NfL+p‐tau₁₈₁ best predicted CDR progression in Aβ+, in line with another study which found p‐tau₁₈₁+NfL had high utility to predict progression from MCI to AD dementia. ¹⁰ Even so, the lowest BIC was NfL alone. In sum, plasma NfL was consistently associated with future cognitive decline in both Aβ+ and Aβ‐.

While NfL had the most consistent prognostic performance, higher plasma p‐tau₁₈₁ was also associated with worse longitudinal MMSE, CDR‐SB, and faster CDR‐progression in Aβ+ cases. These findings show that plasma p‐tau₁₈₁ may be a state sensitive biomarker in AD. For plasma GFAP, cognitive prognosis was more inconsistent, despite several reports in the literature that it is associated with cognition in AD and non‐AD. ⁸ , ¹³ , ⁷² , ⁷³ In Aβ‐ cases, we found plasma GFAP was associated with cross‐sectional and longitudinal MMSE and with faster CDR progression, but not with cross‐sectional or longitudinal CDR‐SB. In Aβ+ cases, cross‐sectional correlations of cognition and GFAP were significant, and plasma GFAP was associated with worse longitudinal MMSE. However in Aβ+, plasma GFAP was not associated with longitudinal CDR‐SB nor did it predict global CDR progression. Differences in GFAP findings may be related to the majority of studies testing cross‐sectional associations with cognition, whereas our longitudinal clinical analyses find less prognostic utility of GFAP in Aβ+. Another potential factor is that GFAP may show early elevation in AD and, thus, may have better prognostic utility early in disease course (e.g., pre‐symptomatic or MCI). ¹³ Accounting for confounding factors may also influence results; another study showed no association of plasma GFAP with cognition in AD after covarying for age and sex. ⁵⁷ Longitudinal biomarker studies are needed to test plasma GFAP trajectories across AD disease stage, if GFAP trajectories differ by AD versus non‐AD pathology, and if GFAP as a disease monitoring or prognostic biomarker differs by disease stage (early vs. late).

There are several caveats to consider when interpreting our findings. First, a major limitation to this study is that we did not measure other isoforms of p‐tau (e.g., 217, 231) that may be more sensitive to early AD than p‐tau₁₈₁. ³ , ⁶ It is unknown if these isoforms differ in diagnostic accuracy by clinical phenotype (amnestic vs. non‐amnestic) that we see here for plasma p‐tau₁₈₁; future studies should test if diagnostic accuracy of various p‐tau isoforms are augmented by plasma GFAP and NfL. Second, in NormCog we did not observe significant correlations of plasma biomarkers with tau PET, although there was a significant interaction by Aβ status for plasma p‐tau₁₈₁. This is likely due to the small number of Aβ+ cases in the NormCog cohort with available tau PET (n = 6). Future work will need to further test pathological correlations in early, pre‐clinical AD. The small number of Aβ+ cases in the NormCog cohort also precluded our ability to perform cross‐validation in independent training and test sets, as performed in ImpCog. Third, we note that some Aβ+ cases were positive for other pathologies (e.g., FTLD, αSyn, vascular disease), reflecting the complexity of pathological diagnosis and that mixed pathologies are the norm in AD. Understanding how mixed pathology affects biomarker levels and interpretation is critical. ⁸ , ⁵⁷ , ⁷⁴ Even so, autopsy data were not available for the majority of the cohort to test the effect of co‐pathologies, as well as vascular disease. While we found no differences by these factors, we may have been underpowered in the smaller autopsy cohort, and future research is needed. Fourth, our population was largely white (79.5% in NormCog; 91.3% in ImpCog), and representation for some groups was very small and underpowered to be examined. It will be important to validate the differences we observed in cohorts with increased diversity. Finally, we did not have body mass index (BMI) or renal dysfunction data in this cohort, which are known confounders of plasma biomarker levels. ⁷⁵ , ⁷⁶

To ensure maximum efficacy, we need biomarker strategies that are robust in the heterogeneous spectrum of neurodegenerative disease. This includes different levels of cognitive impairment, as well as amnestic and non‐amnestic AD. Here, we show that combining plasma p‐tau₁₈₁, GFAP, and NfL in a multivariable model can improve diagnostic accuracy in both NormCog and ImpCog. We also found that patient factors like clinical phenotype (amnestic vs. non‐amnestic) and sex may affect diagnostic performance, and should be considered in future studies. Finally, plasma NfL alone consistently predicted future cognitive decline in both Aβ+ and Aβ‐ cases.

CONFLICT OF INTEREST STATEMENT

I.M.N. reports past educational speaker (Biogen), and advisory board (Eisai) disclosures. D.A.W. has served as a paid consultant to Eli Lilly and Qynapse. He serves on a DSMB for Functional Neuromodulation and GSK. He is a site investigator for a clinical trial sponsored by Biogen. All other authors have report no conflicts of interest relevant to this study. Author disclosures are available in the supporting information.

CONSENT STATEMENT

Written consent was obtained according to the Declaration of Helsinki and approved by the Penn Institutional Review Board.

Supporting information

Supporting information

Supporting information

Cousins KAQ, Phillips JS, Das SR, et al. Pathologic and cognitive correlates of plasma biomarkers in neurodegenerative disease. Alzheimer's Dement. 2024;20:3889–3905. 10.1002/alz.13777