Diagnostic Accuracy of Plasma p-tau217 as a Pre-Screening Tool for Amyloid-PET: A Decision Curve Analysis in the ADNI Cohort

Abstract

Background/Objectives: Amyloid-PET is accurate but costly and capacity-limited. Blood-based biomarkers (BBMs), particularly plasma p-tau217, offer a scalable alternative for triaging patients. Objective of this study is to evaluate the diagnostic accuracy and clinical utility of plasma p-tau217 as a pre-screening tool for amyloid-PET positivity, identifying operating thresholds for rule-out and rule-in strategies. Methods: We analyzed 1681 participants from the ADNI cohort with concurrent plasma biomarkers (Fujirebio Lumipulse assays) and ¹⁸F-florbetapir PET. The primary outcome was discrimination of amyloid-PET positivity (Centiloid > 20). We compared p-tau217 alone against multivariable models (including Aβ42/40, GFAP, NfL, APOE) using Area Under the Curve (AUC) and Decision Curve Analysis (DCA). Two clinical thresholds were defined: a high-sensitivity “rule-out” cut-off (≥95% sensitivity) and a Youden-optimal “balanced” cut-off. These were validated using stratified bootstrap resampling. Results: Of 1681 participants, 679 (40.4%) were amyloid-positive. In the full sample, plasma p-tau217 alone achieved an AUC of 0.902 (95% CI 0.885–0.918). Operating thresholds were derived on a development split and applied to an independent validation split (N = 505). In the validation cohort, the high-sensitivity threshold (0.106 pg/mL) yielded 94.6% sensitivity and 93.7% NPV, effectively ruling out amyloid pathology. The Youden threshold (0.177 pg/mL) offered 78.9% sensitivity and 86.0% specificity. DCA demonstrated net benefit for p-tau217 screening over “PET all” strategies across clinically relevant probability ranges. Conclusions: Plasma p-tau217 provides high discrimination and clinically interpretable operating points for prioritizing confirmatory PET. Implementing a p-tau217-first strategy could significantly reduce unnecessary imaging without compromising diagnostic safety.

1. Introduction

Alzheimer’s disease (AD) is increasingly conceptualized as a biological continuum rather than solely a clinical syndrome. This paradigm shift was formalized by the National Institute on Aging–Alzheimer’s Association (NIA-AA) research framework, which defines AD in vivo through the presence of β-amyloid (A), pathologic tau (T), and neurodegeneration (N) biomarkers—the AT(N) system [1]. This biological definition has become critical with the recent regulatory approval of anti-amyloid monoclonal antibodies, such as lecanemab and donanemab, which require confirmation of amyloid pathology for treatment eligibility [2,3]. Consequently, health systems worldwide are facing an unprecedented demand for accurate amyloid detection to triage eligible patients [4].

Currently, the reference standards for detecting amyloid pathology are cerebrospinal fluid (CSF) analysis and amyloid-positron emission tomography (PET). Both modalities demonstrate excellent diagnostic accuracy and are widely accepted as “gold standards” [5]. However, their widespread implementation faces significant obstacles. PET is costly and geographically restricted, while lumbar puncture is invasive and limited by contraindications or patient refusal [6]. These constraints create a diagnostic bottleneck, risking delays in accessing disease-modifying therapies.

To address these limitations, high-performance blood-based biomarkers (BBMs) have emerged as a scalable alternative [7,8]. Specifically, plasma phosphorylated tau at threonine-217 (p-tau217) has consistently shown the highest accuracy for discriminating AD pathology, recently receiving attention as a key component of the new diagnostic landscape [9]. While other markers such as plasma Aβ42/40, GFAP, and NfL provide valuable information, recent evidence suggests p-tau217 alone may suffice for detecting amyloid positivity in clinical workflows [10].

However, high diagnostic accuracy (AUC) alone is insufficient for clinical adoption. Clinicians require actionable “operating points” tailored to specific intents—such as a high-sensitivity threshold for safely ruling out pathology, or a balanced threshold for confirming likely cases. Furthermore, as highlighted by the recent International Working Group recommendations, the interpretation of these biomarkers must distinguish between “at-risk” status and clinical disease [11].

In this study, using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort, we evaluated plasma p-tau217 as a pre-screening test for amyloid-PET positivity. We aimed to: (1) compare the performance of p-tau217 alone versus multivariable models; (2) assess clinical utility using Decision Curve Analysis (DCA); and (3) define robust cut-offs for “rule-out” and “rule-in” strategies to optimize referrals to confirmatory imaging.

2. Materials and Methods

2.1. Study Design and Participants

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu). The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD) [12].

We conducted a cross-sectional diagnostic accuracy study adhering to the Standards for Reporting Diagnostic Accuracy (STARD 2015) guidelines [13]. Inclusion criteria for the present analysis were: (1) availability of valid plasma biomarker data for p-tau217 analyzed via the designated platform; (2) a concurrent ^18F-florbetapir amyloid-PET scan; and (3) a time interval between blood draw and PET scan of ≤180 days, to minimize the risk of biological progression. Participants with incomplete demographic data, technical failures in PET acquisition, or hemolyzed plasma samples were excluded. Written informed consent was obtained from all participants or their authorized representatives at each ADNI site.

2.2. Amyloid-PET Reference Standard

The reference standard for this study was global cortical amyloid-β load measured by ¹⁸F-florbetapir PET. PET data were acquired across multiple sites and scanners but were harmonized using standard ADNI preprocessing protocols. Images were processed according to the UC Berkeley pipeline, which creates a standardized uptake value ratio (SUVR) map using the whole cerebellum as the reference region.

To ensure interoperability and clinical translation, SUVR values were converted to the Centiloid (CL) scale using the specific transformation equation derived for ¹⁸F-florbetapir [14]. Consistent with established cut-offs validated against neuropathology, amyloid-PET positivity (A+) was defined as a Centiloid value > 20 CL. This threshold was chosen to detect established amyloid pathology while minimizing noise from early accumulation.

2.3. Plasma Biomarker Quantification

Plasma biomarkers were analyzed using the Lumipulse G1200 fully automated chemiluminescent enzyme immunoassay (CLEIA) platform (Fujirebio Europe, Ghent, Belgium). This platform was selected for its high analytical sensitivity, precision, and potential for widespread clinical implementation. Blood samples were collected in EDTA tubes, centrifuged, aliquoted, and stored at −80 °C at the ADNI Biomarker Core Laboratory following standardized pre-analytical protocols to minimize degradation.

The biomarker panel included plasma p-tau217, Aβ42, Aβ40, GFAP, and NfL. The plasma Aβ42/40 ratio was calculated from the individual concentrations. All assays were performed by technicians blinded to the participants’ clinical and amyloid-PET status. To address potential assay variability and batch effects, we utilized the Lumipulse G1200 platform, which is a fully automated system utilizing ready-to-use reagent cartridges. This automation eliminates manual pipetting steps and standardizes incubation times, thereby inherently mitigating the lot-to-lot and day-to-day batch effects commonly observed in manual assays. Quality control (QC) was strictly monitored using internal calibration standards run at the beginning of each analytical batch. Both intra-assay and inter-assay coefficients of variation (CV) were evaluated, and any samples with CVs exceeding the platform’s pre-defined acceptable limits were systematically re-analyzed or excluded from the final dataset. Plasma p-tau217 was pre-specified as the primary predictor of interest.

2.4. Statistical Analysis

2.4.1. Descriptive Statistics and Discrimination

Continuous variables were tested for normality using the ShapirozWilk test. Group comparisons were performed using the non-parametric Wilcoxon rank-sum test. We fitted a series of univariate and multivariable logistic regression models to predict binary amyloid-PET status: Model 1 (p-tau217 alone), Model 2 (M1 + Aβ42/40), Model 3 (M2 + GFAP + NfL), and Model 4 (M3 + APOE ε4). Discrimination was quantified using the Area Under the ROC Curve (AUC) and compared using DeLong’s test [15].

2.4.2. Threshold Derivation and Robust Internal Validation

To define clinically actionable cut-offs, we identified two operating points:

• High-Sensitivity (“Rule-Out”) Threshold: Derived to fix sensitivity at ≥95%, prioritizing the minimization of False Negatives.
• Youden-Optimal (“Rule-In”) Threshold: Calculated to maximize the Youden Index (J = Sensitivity + Specificity − 1).

Regarding missing data, participants lacking concurrent PET status (N = 11) constituted a negligible fraction (<1%) of the initial screened cohort (N = 1692). Given this extremely low proportion, missingness was assumed to be completely at random, and data were handled using a complete-case analysis approach (listwise deletion) without employing multiple imputation, thereby avoiding the introduction of synthetic variance. For threshold derivation and validation, the dataset was partitioned into a development split and an independent validation split (N = 505). In the development split, we performed a rigorous internal validation using stratified bootstrap resampling (B = 2000 bootstrap samples) to derive the operating points, which were subsequently tested on the validation split. The thresholds were re-derived in each iteration, and performance metrics (Sensitivity, Specificity, PPV, NPV, Likelihood Ratios) were calculated. We report the bias-corrected 95% percentile confidence intervals (CIs).

2.5. Clinical Utility

We performed Decision Curve Analysis (DCA) to evaluate the net benefit of the p-tau217-first strategy compared to “PET all” and “PET none” strategies across a range of threshold probabilities. All analyses were performed using R 4.5.2. ROC curve analysis and comparison were performed using the ‘pROC’ package. Decision curve analysis was conducted using the ‘dcurves’ (or ‘rmda’) package. Bootstrap resampling was performed using the ‘boot’ package. Statistical significance was set at p < 0.05 (two-sided).

3. Results

3.1. Study Participants and Descriptive Statistics

We identified 1692 rows after linkage of plasma biomarker and PET files by ADNI Research ID; 11 lacked PET status and were excluded, yielding 1681 participants with both plasma biomarkers and amyloid-PET (Figure 1).

The blood-draw-to-PET interval had a median of 8 days (IQR 1–28; range 0–180). Overall, 679/1681 (40.4%) were amyloid-PET positive (Centiloid > 20). Demographic and biomarker characteristics are summarized in

Table 1. Demographic and biomarker characteristics of the study population.

Variable	Negative (N = 1002) 1	Positive (N = 679) 1	p-Value 2
pt217F	0.10 [0.07, 0.15]	0.36 [0.20, 0.61]	<0.001
AB42	27.67 [24.46, 31.58]	23.87 [21.25, 27.03]	<0.001
AB40	303.79 [270.94, 336.18]	310.48 [273.12, 354.11]	0.011
AB42/40	0.09 [0.08, 0.10]	0.08 [0.07, 0.08]	<0.001
pt217F/AB42	0.00 [0.00, 0.01]	0.01 [0.01, 0.02]	<0.001
NfL	15.50 [11.30, 21.70]	21.30 [15.50, 28.90]	<0.001
GFAP	127.55 [91.39, 180.90]	211.00 [145.90, 291.30]	<0.001
APOE			<0.001
- Non-carrier	78.2% (784.0)	35.9% (244.0)
- Carrier	21.8% (218.0)	64.1% (435.0)
gender			0.484
- Male	45.5% (453.0)	47.3% (318.0)
- Female	54.5% (543.0)	52.7% (355.0)
age	72.00 [67.00, 78.00]	76.00 [71.00, 82.00]	<0.001
pt217F	0.10 [0.07, 0.15]	0.36 [0.20, 0.61]	<0.001

¹ Median [Q1, Q3]; % (n). ² Wilcoxon rank sum test; Fisher’s exact test. Continuous variables are reported as median [IQR]; categorical variables as % (n).

. As expected, amyloid-positive (A+) participants were older, more likely to be APOE ε4 carriers (64.1% vs. 21.8%), and had significantly higher plasma p-tau217, GFAP, and NfL levels compared to amyloid-negative (A−) participants (p < 0.001 for all). Although plasma Aβ42/40 was significantly lower in the A+ group in univariate comparison (p < 0.001), the overlap between groups was substantial compared to p-tau217 (Figure 2).

3.2. Univariate and Multivariable Discrimination Models

In univariate logistic regression (odds ratios [OR] per IQR for continuous predictors), plasma p-tau217 showed a strong association with PET positivity (full estimates in Table S1; univariate forest plot in Figure 3).

Multivariable models were then fitted to assess incremental discrimination: M1 (p-tau217), M2 (+Aβ42/40), M3 (+GFAP, NfL), M4 (+APOE). Model summaries include adjusted ORs (per IQR), 95% CIs and p-values, AUC with DeLong 95% CIs, Brier score, McFadden’s R², ΔAIC, and calibration (intercept/slope and plots) (Tables S2 and S3; Figure 4). Comparative ROC curves are shown in Figure 5.

3.3. Clinical Utility and Decision Curve Analysis

Discrimination and threshold-based performance of M1 were broadly comparable to those of more complex models, with limited incremental gains in sensitivity/specificity or predictive values. Decision-curve analysis showed net benefit for p-tau217 alone overlapping with, or not meaningfully inferior to, that of extended models (Figure 6); risk-distribution plots confirmed clear separation of predicted probabilities between PET− and PET+ for M1, with only modest additional separation after adding other biomarkers (Figure 7).

3.4. Primary Outcome and Threshold Validation

As a continuous test for PET positivity, plasma p-tau217 achieved an AUC of 0.902 (95% CI 0.885–0.918). We then evaluated two prespecified operating thresholds derived on the development split (performance detailed in Table S4) and quantified by stratified non-parametric bootstrap (B = 2000; percentile 95% CIs); the resulting cut-offs (~0.106 for high-sensitivity and ~0.177 for Youden) were frozen and applied unchanged to the validation split. When applied to the independent validation split (N = 505), the high-sensitivity threshold (0.106 pg/mL) yielded a Sensitivity of 94.6%, Specificity of 54.2%, PPV of 58.3%, and NPV of 93.7%; absolute counts in the validation cohort were TP 193, FP 138, TN 163, FN 11. At the Youden-optimal threshold (0.177 pg/mL), performance in the validation split was: Sensitivity 78.9%, Specificity of 86.0%, PPV of 79.3%, and NPV of 85.8%; absolute counts were TP 161, FP 42, TN 259, FN 43 (Table S5, Figure 8).

4. Discussion

In this diagnostic accuracy study leveraging the large ADNI cohort, we demonstrate that plasma p-tau217, measured via a widely accessible automated immunoassay, offers high discrimination for amyloid-PET positivity. Our findings indicate that a parsimonious single-marker model performs equivalently to more complex panels including plasma Aβ42/40, GFAP, NfL, and APOE genotype. Furthermore, by applying Decision Curve Analysis and deriving robust “rule-out” and “rule-in” thresholds, we provide a clinically actionable framework. This approach aligns with the emerging need to streamline the diagnostic pathway for Alzheimer’s disease (AD), potentially reducing the need for confirmatory PET scans by over 50% without compromising diagnostic safety.

4.1. Comparison with Emerging Frameworks and Biomarker Dynamics

Our results consolidate the consensus that p-tau217 is the premier blood-based biomarker for amyloid pathology, a position recently reinforced by major reviews of the diagnostic landscape [10,16]. Consistent with these reports, we found that adding plasma Aβ42/40 provided negligible incremental diagnostic value. While CSF Aβ42/40 remains a reference standard, plasma Aβ42/40 measured by immunoassays (such as the Lumipulse platform used here) often exhibits a narrower dynamic range compared to Mass Spectrometry [17]. However, the Lumipulse platform itself has been recognized for its automation and potential for widespread clinical adoption [16]. Our data suggest that in this automated context, p-tau217 is sufficiently robust to serve as a standalone “gatekeeper,” and the addition of the plasma Aβ42/40 ratio as a separate covariate provides negligible incremental diagnostic value for initial triage.

4.2. Clinical Utility: Navigating the “At-Risk” State

A key contribution of this work is the translation of statistical discrimination (AUC) into operational strategies that respect the latest nosological recommendations. We defined a “rule-in” threshold (0.177 pg/mL) to identify individuals with a high probability of amyloid pathology.

However, as strongly emphasized by the recent International Working Group (IWG) recommendations [11], the interpretation of these biomarkers requires strict clinical contextualization, particularly in asymptomatic individuals. A positive p-tau217 result in a cognitively unimpaired person strictly denotes an “at-risk” biological state—confirming the presence of AD-related pathophysiological changes—but it must never be conflated with a clinical diagnosis of Alzheimer’s disease. The temporal lag between amyloid/tau accumulation and the onset of cognitive decline can span decades. Furthermore, the trajectory from biological positivity to clinical dementia is highly probabilistic, not deterministic. It is heavily modulated by individual resilience, cognitive reserve, advancing age, and the presence of non-AD co-pathologies (e.g., vascular disease). Consequently, deploying a “rule-in” biomarker strategy in asymptomatic populations raises profound ethical and clinical challenges. Disclosing a positive biomarker status without objective cognitive impairment risks inducing severe psychological burden and “biomarker-driven anxiety,” while offering limited prognostic certainty regarding if or when dementia will actually develop. Therefore, biomarker positivity should not be used for deterministic labeling in routine practice. Instead, our high-specificity threshold is best utilized within specialized settings to prioritize patients for rigorous longitudinal clinical monitoring, comprehensive neuropsychological evaluation, or to determine eligibility for secondary prevention trials and emerging disease-modifying therapies (DMTs). Ultimately, while plasma p-tau217 is a powerful tool for detecting the biological footprint of AD, the diagnosis of dementia remains a fundamentally clinical exercise.

Conversely, our high-sensitivity “rule-out” threshold (0.106 pg/mL) yielded a Negative Likelihood Ratio of 0.09. In a specialized memory clinic setting, this would allow clinicians to safely defer expensive imaging for most negative patients, focusing resources on those who truly need them.

4.3. Systemic Implications and Cost-Effectiveness

The implementation of such a triage strategy addresses a critical urgency. As highlighted in the recent Lancet Series on AD outlook [16], health systems are facing a “bottleneck” due to the approval of anti-amyloid monoclonal antibodies [4]. The demand for amyloid confirmation is expected to far exceed PET and CSF capacity. Adopting a p-tau217-first strategy, as supported by our Decision Curve Analysis, offers a sustainable solution: it reduces the burden on specialized memory clinics and reserves scarce PET slots for complex or equivocal cases. This biomarker-guided pathway is essential to ensure equitable access to diagnosis and potential treatment, mitigating the disparities often seen in high-cost diagnostic settings [18,19].

Despite the clear advantages of a p-tau217-first strategy, several practical implementation barriers must be overcome before widespread clinical adoption. First, laboratory standardization remains a critical hurdle; pre-analytical sample handling, centrifugation protocols, and storage conditions can significantly impact biomarker yield. Second, the field currently lacks harmonized cut-offs across different analytical platforms and laboratory environments. Establishing global reference materials is essential to ensure that a specific pg/mL value holds the same clinical weight universally. Finally, reimbursement pathways present a major bottleneck. Healthcare systems and insurance providers are currently grappling with the cost-effectiveness of reimbursing BBMs, particularly when ordered outside of specialized memory clinics or for patients with ambiguous cognitive profiles. Clear guidelines linking biomarker results to actionable therapeutic or diagnostic steps are urgently needed to secure broad reimbursement.

4.4. Limitations and Generalizability

Our findings must be interpreted within the context of several limitations. First, our data extraction focused strictly on the biological AT(N) continuum, meaning syndromic clinical labels (CN, MCI, AD dementia) were not included in the models. Pooling accuracy across this broad clinical spectrum introduces a spectrum bias. Second, the ADNI cohort is a selected research population with a high prevalence of amyloid positivity (~40%), substantially higher than in unselected primary care populations [20]. In a lower-prevalence primary care setting, the Positive Predictive Value (PPV) would substantially decrease, leading to more false positives if used as a standalone diagnostic tool. Conversely, the Negative Predictive Value (NPV) would increase, reinforcing the safety of the “rule-out” strategy. Therefore, our thresholds are best suited for specialized memory clinics rather than general primary care. Third, the ADNI cohort is predominantly Caucasian and highly educated; future studies must validate these specific p-tau217 thresholds in cohorts with greater ethnic, socioeconomic, and medical comorbidity diversity [18,19]. Finally, we focused on biological prediction of amyloid status; as noted by the IWG [11], the translation from amyloid positivity to clinical dementia is probabilistic, not deterministic, and clinical judgment remains paramount.

5. Conclusions

Plasma p-tau217 stands out as a robust, standalone tool for amyloid triage. By validating specific operating points for ruling out and ruling in pathology, this study supports a pragmatic workflow that integrates seamlessly with the new diagnostic and therapeutic landscape described by recent major guidelines. Implementing this p-tau217-first strategy represents a scalable, cost-effective path to manage the growing demand for AD diagnosis in specialized clinical settings.