Plasma pTau 217/β-amyloid 1–42 ratio for enhanced accuracy and reduced uncertainty in detecting amyloid pathology

Abstract

Blood biomarkers have the potential to revolutionize Alzheimer's disease diagnosis, offering advantages over CSF and PET due to their accessibility, scalability and cost-effectiveness. This study evaluated the effectiveness of individual plasma biomarkers, such as phosphorylated Tau (pTau) 217, as well as biomarker combinations, with a focus on the pTau 217/β-Amyloid (Aβ) 1–42 ratio to predict amyloid positivity. To improve clinical utility, a dual threshold approach was applied to maximize predictive values and positive likelihood ratios while minimizing the proportion of indeterminate results.

Plasma samples from 208 participants (including seven with subjective cognitive decline, 150 with mild cognitive impairment, 12 with Alzheimer's disease dementia and 39 with other cognitive conditions) from three cohorts (BioFINDER2, BIOCARD and MissionAD) were analysed to measure Aβ_1–42, Aβ_1–40 and pTau217 levels using the Fujirebio LUMIPULSE^® G1200 platform. Amyloid status was determined by FDA-cleared PET imaging and/or CSF biomarker ratios. Logistic regression modelling evaluated biomarkers either individually or in combination to identify those that best distinguished amyloid positivity. Clinically applicable thresholds were established through likelihood ratio analysis and further evaluated based on predictive values.

When assessing the ability of individual plasma biomarkers to differentiate between amyloid-positive and amyloid-negative participants, plasma pTau217 (P < 0.001) and plasma Aβ_1–42 (P = 0.0056) demonstrated significant discriminative power, whereas Aβ_1–40 (P = 0.30) did not. Notably, the integration of these biomarkers into the plasma pTau217/Aβ_1–42 ratio demonstrated enhanced classification performance (P < 0.001). Using a two-threshold approach based on positive and negative likelihood ratios (PLR/NLR) targets of 14/20, respectively, the plasma pTau217/Aβ_1–42 ratio achieved a positive predictive value (PPV) of 94.44% and negative predictive value (NPV) of 94.28%, in the parametric model, comparable to plasma pTau217 alone (PPV: 94.44%, NPV: 94.28%), but yielded fewer indeterminate results (26.5% versus 38.6%). Using a non-parametric model, the plasma ratio achieved a PPV and NPV of 94.62% and 91.78%, respectively, while plasma pTau217 alone achieved 92.41% and 92.86%; the ratio once again reduced the proportion of indeterminate results (20.2% versus 35.1%).

The plasma pTau217/Aβ_1–42 ratio demonstrated superior performance in identifying amyloid pathology and reduced the frequency of indeterminate results compared to plasma pTau217 alone. These findings support the evaluation of the clinical utility of the plasma pTau217/Aβ_1–42 ratio as a tool for identifying amyloid pathology in patients presenting with cognitive complaints.

Benina et al. evaluated the plasma pTau217/Aβ 1–42 ratio as a blood-based marker of Alzheimer's disease pathology. The ratio improved diagnostic accuracy and reduced indeterminate results compared with pTau217 alone, supporting its clinical utility for identifying amyloid positivity.

Introduction

Alzheimer's disease (AD), the most common cause of dementia worldwide, is a neurodegenerative disease characterized by a progressive decline across multiple cognitive domains.¹ The primary neuropathological hallmarks include the accumulation of extracellular β-amyloid (Aβ)-containing plaques and intracellular tau-containing neurofibrillary tangles in the brain.^2,3 The downstream outcome of these pathological processes underlies the synaptic and neuronal loss that contributes to macro-scale cortical atrophy.^4,5 AD pathology develops decades before the appearance of cognitive symptoms,⁶ starting with a preclinical stage characterized by the abnormal accumulation of Aβ. In many individuals, this stage is accompanied by early tau alterations detectable via CSF or plasma biomarkers, or Aβ and tau PET, even in the absence of cognitive impairment.^7,8 As the disease progresses, tau pathology becomes widespread and is associated with neurodegeneration,⁹ marking the transition to the prodromal phase of mild cognitive impairment (MCI), which may eventually progress to clinical AD dementia.³

The recent introduction of disease-modifying treatments (DMTs) targeting AD pathophysiology^10,11 is expected to significantly increase the demand for patient assessments to determine treatment eligibility.¹² Since these therapies target Aβ aggregates, biomarker confirmation of amyloid pathology is required before DMT initiation,¹³ particularly in early-stage patients, where clinical diagnostic criteria are often inaccurate. Currently, amyloid pathology is confirmed by amyloid PET with US Food and Drug Administration (FDA)-approved tracers¹⁴ or validated CSF tests measuring concentrations of Aβ_1–42, Aβ_1–40 and tau phosphorylated at threonine-181 (pTau181).^15,16 However, widespread implementation of these modalities is limited. Amyloid PET scans are costly and not universally available.¹⁷ Additionally, the use of CSF biomarkers is limited, as many healthcare providers do not perform lumbar punctures (LPs), some patients perceive the procedure as invasive or risky,¹⁸ and others may be ineligible due to medical contraindications, such as coagulopathies or the use of anticoagulant medications.¹⁹ These limitations hinder implementation in memory clinics or other care settings to confirm amyloid pathology.¹⁷ There is an urgent clinical need for scalable, cost-effective tools to identify patients eligible for DMTs.

The use of blood biomarker (BBM) tests to predict amyloid pathology offers a promising alternative to CSF and PET due to their limited invasiveness, scalability, accessibility and cost-effectiveness,²⁰ while demonstrating high concordance with amyloid PET imaging and CSF ratio testing.^21-23 Owing to their simplicity, they can be implemented in primary and secondary care settings,²⁴ and have already been incorporated into clinical trials.^25,26

The Global CEO Initiative on AD recently published a consensus outlining acceptable performance criteria for BBM tests for amyloid pathology.²⁴ For BBMs used as triage tests, where positive results are followed by confirmatory amyloid PET or CSF AD testing, the recommended performance to predict amyloid PET status includes ≥90% sensitivity and 75% to 85% specificity [positive predictive value (PPV) 78%–86%, negative predictive value (NPV) 88%–89%, at 50% prevalence]. For BBM assays used as stand-alone confirmatory tests without subsequent amyloid PET or CSF results, performance should be equivalent to that of CSF tests, with 90% sensitivity and specificity for predicting amyloid PET status (PPV 90%, NPV 90%, at 50% prevalence). Additionally, multiple studies have suggested using a two-threshold approach to optimize overall accuracy.^27-29 This approach categorizes results as positive, negative or intermediate/indeterminate (hereafter referred to as the indeterminate category). The upper threshold optimizes test sensitivity, ensuring a high probability of amyloid PET and CSF positivity, while the low threshold identifies those with a low likelihood of amyloid pathology. As defined by the Global CEO Initiative, results between the thresholds (i.e. the indeterminate range) indicate uncertainty of amyloid status or intermediate levels of amyloid pathology and would require confirmatory follow-up evaluation, such as CSF or amyloid PET testing, or repeat BBM assessment.²⁴ The Global CEO Initiative recommends limiting this ‘indeterminate’ group to no more than 20% of tested individuals (Supplementary Fig. 1).^24,30

This study was designed to evaluate the performance of BBMs in a secondary care setting, focusing on individuals exhibiting signs or symptoms of cognitive decline. We assessed how well plasma pTau217 and the pTau217/Aβ_1–42 ratio estimate amyloid status, using PET and/or CSF as reference standards. As part of this evaluation, we aimed to establish clinically applicable thresholds that would achieve satisfactory positive and negative predictive values (PPV, NPV) and positive and negative likelihood ratios (PLR, NLR), while limiting the proportion of indeterminate results to align with the Global CEO Initiative consensus.²⁴ Biobanked plasma samples from observational and interventional research cohorts were analysed using Fujirebio's Lumipulse assays.

Materials and methods

Participants

This study included a total of 208 individuals from three well-characterized cohorts: the Swedish BioFINDER2 Study (n = 44; ClinicalTrials.gov Identifier: NCT03174938), the Biomarkers for Older Controls at Risk for Dementia (BIOCARD) cohort (n = 66; NIH grant number: 5U19AG033655), and the screening phase of the phase 3 elenbecestat clinical trials (MissionAD; n = 98; ClinicalTrials.gov Identifier: NCT02956486). All participants provided informed consent for future biomarker research, underwent amyloid PET imaging with an FDA-cleared tracer (flutemetamol, florbetapir and florbetaben) and/or FDA-authorized CSF biomarker ratio assessment (Lumipulse G β-Amyloid 1–42/1–40 Ratio or Elecsys^® pTau 181/Aβ 1–42 Ratio) and were investigational and DMT treatment-naïve. Across cohorts, the average time between plasma and reference standard collection was 7 days for PET (median 0 days; range 108 days before to 102 days after plasma collection) and 0 days for CSF (median 0 days; range 1 day before to the same day as plasma collection). The majority of participants in this study had a clinical diagnosis within the AD continuum, including subjective cognitive decline (SCD)/subjective memory concern, MCI or mild AD dementia. Clinical diagnosis was defined by the enrolling site. Participants outside the AD continuum but identified by the enrolling site as cognitively impaired were classified into a fourth group, referred to as ‘other cognitive conditions’. This group included vascular dementia, dementia not otherwise specified, and cognitive impairment not meeting criteria for MCI. Detailed descriptions of the diagnostic criteria used at each site have been previously published.^31-33 Exclusion criteria included a history of significant neurological disease, schizophrenia, and/or alcohol/substance abuse or dependence within the past 3 years.

The de-identified dipotassium ethylenediaminetetraacetic acid (K₂EDTA)-plasma samples from these participants were selected based on specific pre-analytical and storage inclusion criteria. Eligible samples were required to be collected in low-binding polypropylene tubes containing K₂EDTA, subjected to no more than one freeze/thaw cycle, free of any sign of haemolysis, and stored in an ultra-low temperature freezer (≤ −60°C) until analysis.

Consent statement

All human participants provided informed consent.

Blood samples and plasma testing

On the day of testing, the biobanked plasma samples were thawed for a minimum of 45 min at room temperature (23°C–28°C), vortexed for 10 s and centrifuged at 2000g for 5 min before being tested. Plasma testing was conducted on the Fujirebio LUMIPULSE G1200 automated immunoassay analyser using the Lumipulse^®  G β-Amyloid 1–42-N Plasma, Lumipulse^®  G β-Amyloid 1-40-N Plasma and Lumipulse^®  G p-Tau 217 assays. Plasma biomarker levels were quantified in singlicate. Sample aliquots were tested at Fujirebio Diagnostic, Inc. facilities (Malvern, PA, USA).

Amyloid status assessment

Given that both CSF and PET are FDA-cleared reference standards for amyloid pathology, we incorporated both into our analysis to reflect real-world clinical practice. This also enabled us to assess whether the plasma ratio performed consistently across modalities, a key consideration for ensuring its clinical utility.

Amyloid status for each participant was determined according to cohort-specific protocols. In the BioFINDER2 cohort, amyloid status was assessed in CSF using the Lumipulse G β-Amyloid 1–42/1–40 Ratio assay following manufacturer's specifications and with PET imaging using the flutemetamol tracer. PET data was unavailable for one participant. For the MissionAD cohort, amyloid status was assessed in CSF using the Elecsys pTau 181/Aβ 1–42 Ratio assay, while PET imaging employed FDA-approved tracers (florbetapir, florbetaben or flutemetamol). In the BIOCARD cohort, amyloid assessment relied solely on the CSF Lumipulse G β-Amyloid 1–42/1–40 Ratio, following the manufacturer's specifications.

Amyloid positivity was defined by a PET standardized uptake value ratio (SUVR) of ≥1.13 (flutemetamol, florbetapir or florbetaben),^34,35 uniformly applied across all study cohorts, a CSF Lumipulse G β-Amyloid 1–42/1–40 Ratio of <0.073 or a CSF Elecsys pTau 181/Aβ 1–42 Ratio of >0.023. In cases where the SUVR value and CSF biomarker testing results were unavailable, amyloid status was determined based on the site's visual PET scan interpretation. This approach was used for n = 6 of the 208 participants (2.9%). Additional evaluation of the uniform SUVR threshold relative to tracer-specific cut-offs is provided in Supplementary Figs 2 and 3.

For CSF biomarkers, the Elecsys pTau 181/Aβ 1–42 Ratio used the >0.023 threshold as specified in the manufacturer's instructions.³⁶ For the Lumipulse G β-Amyloid 1–42/1–40 Ratio, although the manufacturer’s instructions specify three interpretive results (≤0.058 as positive, 0.059–0.072 as likely positive and ≥0.073 as negative), a simplified threshold of <0.073 was used to combine positive and likely positive results into a single amyloid-positivity category. This approach has also been adopted in previous studies.^37,38

In this study, 89 participants had PET data, 67 had CSF data and 52 had CSF and PET data. Among these 52, nine cases showed discordant results: seven participants had positive CSF biomarkers but negative PET imaging, while two participants had a negative CSF result but positive PET findings. In cases of discordance between CSF and PET results, amyloid status was classified as positive if either the CSF test or PET imaging showed positive amyloid findings, ensuring detection of amyloid pathology, regardless of the method used.

APOE status determination

APOE status was determined by genotyping. Genotyping results were provided by the respective site and used for APOE status assignment in this study. Participants carrying at least one ε4 allele (ε2/ε4, ε3/ε4 or ε4/ε4) were classified as ε4 carriers (ε4+); all other genotypes were classified as non-carriers (ε4−).

Statistical analysis

Discriminative analysis of plasma biomarkers for amyloid positivity

The plasma biomarkers examined in this study to evaluate amyloid positivity were Aβ_1–42, Aβ_1–40 and pTau217. Logistic regression modelling was used to estimate the probability of amyloid positivity as a function of the natural logarithm-transformed plasma concentrations of Aβ_1–40, Aβ_1–42 and pTau217. Additionally, potential biomarker combinations, such as the ratio of plasma pTau217/Aβ_1–42, were also evaluated for their ability to predict amyloid positivity.

Based on results obtained from the logistic regression modelling, a receiver operating characteristic (ROC) curve analysis was conducted to evaluate the discriminative ability of pTau217 and the pTau217/Aβ_1–42 ratio to identify amyloid status. The area under the ROC curve (AUC) was used as a summary measure of classification performance, equivalent to the non-parametric Wilcoxon rank-sum (Mann–Whitney U). The Youden index was calculated to identify the point maximizing sensitivity and specificity.

Additionally, to assess whether demographic and genetic factors might influence the amyloid positivity classification accuracy, age, sex and APOE status (ε4 carrier versus non-carrier) were evaluated using a two-way ANOVA. For each biomarker (log-transformed pTau217 and pTau217/Aβ_1–42), the main effects of amyloid status and each covariate were tested, and an evaluation was carried out to assess whether the covariates added explanatory power beyond amyloid status. If covariates added no explanatory power beyond amyloid status, they were not included in the final classification model. To accommodate multiple comparisons, a P-value < 0.01 was considered statistically significant for all analyses.

Thresholds determination

A parametric statistical analysis was conducted to optimize threshold selection for classification of amyloid positivity while retaining flexibility to evaluate thresholds. For this, biomarker distributions were first transformed to approximate normality using the Box-Cox transformation, defined as:

$$\mathrm{Y}=({\mathrm{X}}^{\mathrm{c}}-1)/\mathrm{c}$$ (1)

where X represents the biomarker value, Y denotes the Box-Cox transformed value, and c is the Box-Cox power parameter. This method has been shown to identify classification thresholds efficiently.³⁹ The transformation and subsequent analyses were performed in R using the car⁴⁰ package.

In this initial evaluation stage, a parametric approach was chosen over a non-parametric direct search due to several advantages. Statistically, it allows for more efficient use of the data by summarizing it into a few key parameters: the Box-Cox power, the means and the standard deviations of the transformed distributions for the amyloid-positive and amyloid-negative groups. This method also provides greater flexibility, enabling the calculation of performance metrics for any threshold mathematically, rather than being restricted to observed data points.

The means and standard deviations of the transformed distributions were used to calculate sensitivity and specificity estimates across a finely spaced grid of potential threshold values that covered the entire range of each biomarker. These estimates were then used to calculate the corresponding PPV and NPV values, as well as PLR and NLR values for each threshold. It is important to note that while sensitivity, specificity, and likelihood ratios are independent of prevalence, predictive values are affected by it. In these calculations, the prevalence of amyloid positivity used was 54.8%, reflecting the actual prevalence observed in our dataset and aligning with the prevalence typically observed in secondary care settings.^24,41

PLR and NLR were included in the evaluation to provide a more comprehensive measure of each biomarker's ability to distinguish between amyloid-positive and amyloid-negative status. Specifically, likelihood ratios (LRs) represent a useful statistical tool for clinicians to interpret test results and calculate post-test probabilities of a disease or condition. To support a triage-based classification approach using two thresholds (one to rule in and one to rule out amyloid positivity), threshold values were selected from the diagnostic performance grid based on LR criteria. Specifically, grid values yielding positive LRs (PLRs) of 10, 14 and 19 were evaluated as candidates for the upper threshold, while the lower threshold was fixed at a negative LR (NLR) of 0.05 (the reciprocal of 20). For clarity, we refer to this NLR value as ‘20’ throughout the manuscript, reflecting its reciprocal relationship (1/0.05 = 20) to facilitate comparison with PLR values. These PLR and NLR targets were chosen based on established diagnostic interpretation guidelines,⁴² where a PLR ≥ 10 is considered to provide strong evidence for ruling in disease, and a NLR ≤ 0.1 (1/0.1 = 10) offers strong evidence for ruling out disease. PLR targets of 14 and 19 were included to explore more conservative rule-in options. Given that the dataset included both CSF and PET amyloid status for many participants, thresholds were evaluated separately against CSF alone, PET alone, and the combination of the two reference methods (CSF and PET).

Based on the highest overall performance observed in the above evaluation, the selected PLR/NLR targets of 14/20 derived from the parametric model were also evaluated non-parametrically (given the non-normal distribution of the biomarker values across cohorts) to further assess their performance in distinguishing amyloid-positive from amyloid-negative participants on an individual level. This was done by applying each threshold pair directly to the 208 participants in the dataset and categorizing their biomarker values accordingly. Sensitivity, specificity, PPV, NPV, PLR and NLR were then calculated based on the actual (untransformed) biomarker distributions. These non-parametric estimates do not rely on distributional assumptions and are subject to sample variability; therefore, some divergence from the parametric projections was expected. Nevertheless, the degree of agreement between the two approaches provides a meaningful validation and helps ensure the robustness and practical validity of the parametric model used for threshold selection.

Results

Participants and biomarker characteristics

A total of 208 participants were included in this study, including those with SCD (n = 7), MCI (n = 150), Alzheimer's disease dementia (AD, n = 12) and other cognitive conditions (other, n = 39) based on their clinical diagnosis. The age ranged from 51 to 88 years, with an average of 72.3 years. Approximately half of the participants in each group were male (51.0% overall, 57.1% among SCD, 52.0% among MCI, 50.0% among AD and 46.2% among those with other cognitive conditions). Disease prevalence, defined as having a positive Aβ PET scan and/or CSF biomarker test result, was 54.8% overall (71.4% in SCD, 54.7% in MCI, 83.3% in AD and 43.6% in those with other cognitive conditions) (Table 1 and Supplementary Table 1).

Table 1.

Participant demographic characteristics for clinical groups stratified by Aβ PET and/or CSF

		Totals	Diagnostic groups				CSF biomarker ratio		SUVR or visual PET read		CSF ratio and/or SUVR or visual PET read
			AD	MCI	SCD	Other	Positive	Negative	Positive	Negative	Positive	Negative
Totals per category		208	12	150	7	39	66	53	73	68	114	94
Sex	Male	106 (51.0%)	6 (50.0%)	78 (52.0%)	4 (57.1%)	18 (46.2%)	36 (54.5%)	25 (47.2%)	39 (53.4%)	39 (57.4%)	61 (53.5%)	45 (47.9%)
	Female	102 (49.0%)	6 (50.0%)	72 (48.0%)	3 (42.9%)	21 (53.8%)	30 (45.5%)	28 (52.8%)	34 (46.6%)	29 (42.6%)	53 (46.5%)	49 (52.1%)
Age, years	51–65	38 (18.3%)	1 (8.3%)	26 (17.3%)	5 (71.4%)	6 (15.4%)	9 (13.6%)	12 (22.6%)	11 (15.1%)	19 (27.9%)	14 (12.3%)	24 (25.5%)
	66–70	40 (19.2%)	2 (16.7%)	35 (23.3%)	0 (0.0%)	3 (7.7%)	8 (12.1%)	8 (15.1%)	11 (15.1%)	22 (32.4%)	16 (14.0%)	24 (25.5%)
	71–75	52 (25.0%)	1 (8.3%)	37 (24.7%)	2 (28.6%)	12 (30.8%)	19 (28.8%)	17 (32.1%)	18 (24.7%)	13 (19.1%)	31 (27.2%)	21 (22.3%)
	76–80	56 (26.9%)	8 (66.7%)	35 (23.3%)	0 (0.0%)	13 (33.3%)	19 (28.8%)	13 (24.5%)	24 (32.9%)	10 (14.7%)	37 (32.5%)	19 (20.2%)
	≥81	22 (10.6%)	0 (0.0%)	17 (11.3%)	0 (0.0%)	5 (12.8%)	11 (16.7%)	3 (5.7%)	9 (12.3%)	4 (5.9%)	16 (14.0%)	6 (6.4%)
	Average	72	74	72	62	74	74	71	73	69	74	70
	Median	74	76	74	60	75	75	73	75	68	75	70
	Range	51–88	63–80	54–88	51–73	53–88	51–88	53–88	51–86	55–85	51–88	53–88
Race	White	197 (94.7%)	12 (100.0%)	141 (94.0%)	7 (100.0%)	37 (94.9%)	65 (98.5%)	51 (96.2%)	71 (97.3%)	62 (91.2%)	111 (97.4%)	86 (91.5%)
	Asian	2 (1.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	2 (5.1%)	1 (1.5%)	1 (1.9%)	0 (0.0%)	0 (0.0%)	1 (0.9%)	1 (1.1%)
	Black	7 (3.4%)	0 (0.0%)	7 (4.7%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	1 (1.9%)	1 (1.4%)	5 (7.4%)	1 (0.9%)	6 (6.4%)
	Other (Chinese)	2 (1.0%)	0 (0.0%)	2 (1.3%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	1 (1.4%)	1 (1.5%)	1 (0.9%)	1 (1.1%)
Ethnicity	Non-Hispanic	123 (59.1%)	8 (66.7%)	81 (54.0%)	0 (0.0%)	34 (87.2%)	37 (56.1%)	33 (62.3%)	40 (54.8%)	18 (26.5%)	74 (64.9%)	49 (52.1%)
	Hispanic	40 (19.2%)	1 (8.3%)	39 (26.0%)	0 (0.0%)	0 (0.0%)	4 (6.1%)	0 (0.0%)	11 (15.1%)	29 (42.6%)	13 (11.4%)	27 (28.7%)
	Unknown	45 (21.6%)	3 (25.0%)	30 (20.0%)	7 (100.0%)	5 (12.8%)	25 (37.9%)	20 (37.7%)	22 (30.1%)	21 (30.9%)	27 (23.7%)	18 (19.1%)
Years of education	≤13	66 (31.7%)	7 (58.3%)	52 (34.7%)	0 (0.0%)	7 (17.9%)	20 (30.3%)	16 (30.2%)	25 (34.2%)	32 (47.1%)	33 (28.9%)	33 (35.1%)
	>13	138 (66.3%)	5 (41.7%)	94 (62.7%)	7 (100.0%)	32 (82.1%)	45 (68.2%)	35 (66.0%)	47 (64.4%)	33 (48.5%)	80 (70.2%)	58 (61.7%)
	Unknown	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Clinical diagnosis group	AD	12 (5.8%)	12 (100.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	4 (6.1%)	0 (0.0%)	9 (12.3%)	3 (4.4%)	10 (8.8%)	2 (2.1%)
	MCI	150 (72.1%)	0 (0.0%)	150 (100.0%)	0 (0.0%)	0 (0.0%)	41 (62.1%)	28 (52.8%)	59 (80.8%)	58 (85.3%)	82 (71.9%)	68 (72.3%)
	SCD	7 (3.4%)	0 (0.0%)	0 (0.0%)	7 (100.0%)	0 (0.0%)	5 (7.6%)	2 (3.8%)	3 (4.1%)	4 (5.9%)	5 (4.4%)	2 (2.1%)
	Other	39 (18.8%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	39 (100.0%)	16 (24.2%)	23 (43.4%)	2 (2.7%)	3 (4.4%)	17 (14.9%)	22 (23.4%)

AD = Alzheimer's disease; MCI = mild cognitive impairment; SCD = subjective cognitive decline; SUVR = standardized uptake value ratio.

Best-performing biomarkers for amyloid assessment

Logistic regression analysis of the evaluated plasma biomarkers showed that Aβ_1–42 (P = 0.0056) and pTau217 (P < 0.001) were the most significant biomarkers in distinguishing clinical diagnosis, with Aβ_1–42 demonstrating a strong negative association and pTau217 a strong positive association with disease status based on the logistic coefficients (β = −1.905 and β = 2.883, respectively). In contrast, Aβ_1–40 alone did not achieve statistical significance (P = 0.30) (Supplementary Table 2). Furthermore, multivariate regression revealed that Aβ_1–40 (P = 0.34) added no additional predictive power when Aβ_1–42 and pTau217 were included in the model and was thus excluded from further analysis (Supplementary Table 3). Notably, in the multivariate model, the coefficients for Aβ_1–42 and pTau217 were approximately equal in magnitude but opposite in direction (−4.50 and 3.05, respectively), suggesting that their combined effect could be captured more effectively through a ratio. A subsequent regression analysis confirmed that the pTau217/Aβ_1–42 ratio provided a strong classification power (β = −3.17, P < 0.001) (Supplementary Table 4). Accordingly, all subsequent modelling focused on plasma pTau217 levels alone and the pTau217/Aβ_1–42 ratio.

To evaluate the discriminatory performance of plasma pTau217 levels alone and the pTau217/Aβ_1–42 ratio, ROC curves were constructed for both pTau217 and the pTau217/Aβ_1–42 ratio (Fig. 1). The ROC AUC was 0.918 (95% CI 0.880–0.956) for pTau217 and 0.932 (95% CI 0.896–0.967) for the pTau217/Aβ_1–42 ratio, demonstrating strong discriminative ability for both markers. At the Youden optimal threshold, the pTau217 model (at a threshold = >0.173) achieved 81% sensitivity and 93% specificity (Youden's J = 0.74), while the pTau217/Aβ_1–42 ratio model (at a threshold = >0.00642) achieved 84% sensitivity and 95% specificity (Youden's J = 0.79) (Fig. 1). The ROC AUC is equivalent to the Wilcoxon rank-sum (Mann–Whitney U) test, and thus also serves as a non-parametric test to determine whether the biomarker distributions differ significantly between amyloid-positive and amyloid-negative individuals. In both cases, the evidence was statistically significant (P < 0.001 for both pTau217 and the pTau217/Aβ_1–42 ratio), confirming the high predictive value of both markers, with the ratio-based metric providing slightly superior performance.

Figure 1.

Discriminatory performance of plasma pTau217 and the pTau217/Aβ_1–42 ratio for prediction of amyloid PET + CSF positivity. Receiver operating characteristic (ROC) curves for plasma pTau217 (green) and the plasma pTau217/Aβ1–42 ratio (blue) to predict amyloid PET + CSF positivity. Aβ = amyloid-beta; FPF = false positive fraction; p-tau = phosphorylated tau; TPF = true positive fraction.

Additionally, covariates such as age, sex, and APOE ε4 carriership were evaluated for their potential influence on biomarker values and their interaction with amyloid status. Overall, neither age nor sex showed significant main effects or interactions, suggesting they are unlikely to impact biomarker performance. Although APOE ε4 status was found to have a significant main effect on the pTau217/Aβ_1–42 ratio (P = 0.009), the interaction with amyloid status, defined as positive amyloid PET or CSF test, was not significant (P = 0.2807). This indicates that the ratio's ability to distinguish amyloid positivity remains consistent across different APOE statuses (Supplementary Table 5). Furthermore, distribution plots showed only minor shifts in overall values, reinforcing the minimal classification relevance of this association (Supplementary Fig. 4). As a result, no demographic or genetic covariates were included in the final classification models.

Clinical thresholds for identification of CSF and PET amyloid positivity

To prepare biomarker data for parametric modelling, Box-Cox transformations were applied to approximate normal distributions. The optimal transformation parameters were found to be c = −0.25 for plasma pTau217 levels alone and c = −0.30 for the pTau217/Aβ_1–42 plasma ratio. The transformed distributions were then used to estimate diagnostic performance across a fine grid of threshold values. A two-threshold approach was subsequently implemented for both plasma pTau217 and the pTau217/Aβ_1–42 plasma ratio using predefined PLR/NLR targets (10/20, 14/20 and 19/20). These threshold combinations were initially evaluated parametrically against the combined results from CSF and PET to avoid bias towards either reference standard while estimating performance. Parametric calculations of the sensitivity, specificity, predictive values, and likelihood ratios are shown in Figs 2 and 3. As the PLR target increased, the PPV improved for both biomarkers, reaching 95.8% for pTau217 and the pTau217/Aβ_1–42 ratio at the 19/20 PLR/NLR targets. In all models, the lower threshold was fixed to achieve an NLR of 20 and ensure a high level of rule-out confidence, with PPVs in the negative zone consistently near 5.7%. These predictive values were calculated based on an amyloid positivity prevalence of 54.8%. As expected, stricter PLR targets resulted in larger indeterminate zones, reflecting the trade-off between diagnostic certainty and the proportion of inconclusive results (Tables 2 and 3).

Figure 2.

Distributional validation and parametric estimation of diagnostic performance for plasma pTau217. Quantile–quantile plots of plasma pTau217 values in amyloid-positive (A) and amyloid-negative (B) individuals. Parametric estimates of diagnostic performance across the range of plasma pTau217 concentrations, including sensitivity and specificity (C), predictive values (D) and likelihood ratios (E). Dotted horizontal lines indicate performance targets for pTau217 alone (90% for sensitivity and specificity, 85% for predictive value, and 10 for the likelihood ratio). NLR = negative likelihood ratio; NPV = negative predictive value; PLR = positive likelihood ratio; PPV = positive predictive value; Sens = sensitivity; Spec = specificity.

Figure 3.

Distributional validation and parametric estimation of diagnostic performance for plasma pTau217/Aβ_1–42 ratio. Quantile–quantile plots of plasma pTau217/Aβ_1–42 ratio values in amyloid-positive (A) and amyloid-negative (B) individuals. Parametric estimates of diagnostic performance across the range of pTau217/Aβ_1–42 ratio values, including sensitivity and specificity (C), predictive values (D) and likelihood ratios (E). Dotted horizontal lines indicate performance targets for the pTau217/Aβ_1–42 ratio (90% for sensitivity and specificity, 85% for predictive value, and 10 for the likelihood ratio). NLR = negative likelihood ratio; NPV = negative predictive value; PLR = positive likelihood ratio; PPV = positive predictive value; Sens = sensitivity; spec = specificity.

Table 2.

Parametric performance of pTau217 for detecting amyloid positivity based on CSF + PET combined at different PLR/NLR target thresholds

PLR/NLR	Thresholds	Test result	True positives	True negatives	Total	LR	PV (%)
10/20	≥0.19973	Positive	39.0	3.2	42.2	10	92.38
	0.08681 < Result < 0.19973	Indeterminate	14.4	18.7	33.1	0.637	43.58
	≤0.08681	Negative	1.4	23.3	24.7	0.050	5.72
14/20	≥0.23211	Positive	34.6	2.0	36.7	14	94.44
	0.08681 < Result < 0.23211	Indeterminate	18.8	19.9	38.6	0.780	48.62
	≤0.08681	Negative	1.4	23.3	24.7	0.050	5.72
19/20	≥0.26599	Positive	30.5	1.3	31.8	19	95.84
	0.08681 < Result < 0.26599	Indeterminate	22.9	20.6	43.5	0.914	52.66
	≤0.08681	Negative	1.4	23.3	24.7	0.050	5.72

Classification of test results for pTau217 using assay thresholds determined at PLR/NLR targets of 10/20, 14/20, and 19/20, evaluated against amyloid positivity as determined by CSF or PET imaging. Reported metrics include true positives, true negatives, likelihood ratios, and predictive values, stratified by test result category (positive, indeterminate, and negative). The numbers in the ‘true positive’, ‘true negative’, and ‘Total’ columns represent the predicted percentage breakdown in the study population. LR = likelihood ratio; NLR = negative likelihood ratio; PLR = positive likelihood ratio; PV = predictive value.

Table 3.

Parametric performance of pTau217/Aβ_1–42 ratio for detecting amyloid positivity based on CSF + PET combined at different PLR/NLR target thresholds

PLR/NLR	Thresholds	Test result	True positives	True negatives	Total	LR	PV (%)
10/20	≥0.00654	Positive	44.3	3.7	48.0	10	92.38
	0.00370 < Result < 0.00654	Indeterminate	8.7	13.5	22.2	0.533	39.26
	≤0.00370	Negative	1.7	28.0	29.8	0.050	5.724
14/20	≥0.00738	Positive	41.3	2.4	43.7	14	94.44
	0.00371 < Result < 0.00737	Indeterminate	11.8	14.7	26.5	0.660	44.46
	≤0.00370	Negative	1.7	28.0	29.8	0.050	5.724
19/20	≥0.00822	Positive	38.4	1.7	40.0	19	95.84
	0.00370 < Result < 0.00822	Indeterminate	14.7	15.5	30.2	0.785	48.77
	≤0.00370	Negative	1.7	28.0	29.8	0.050	5.724

Classification of test results for the pTau217/Aβ_1–42 ratio using assay thresholds determined at PLR/NLR targets of 10/20, 14/20 and 19/20, evaluated using amyloid positivity as determined by CSF or PET imaging. Reported metrics include true positives, true negatives, likelihood ratios and predictive values, stratified by test result category (positive, indeterminate and negative). The numbers in the ‘true positive’, ‘true negative’ and ‘Total’ columns represent the predicted percentage breakdown in the study population. LR = likelihood ratio; NLR = negative likelihood ratio; PLR = positive likelihood ratio; PV = predictive value.

When evaluated against the combined CSF and PET results as the reference standard and targeting a PLR of 10 and NLR of 20, plasma pTau217 alone achieved a PPV of 92.4% and a NPV of 94.3%, with 33.1% of the study's population falling into the indeterminate zone. Increasing the PLR target to 14, while maintaining the NLR at 20, and consequently keeping the NPV at 94.3%, resulted in an increase in PPV to 94.4%, while the indeterminate zone expanded to 38.6%. Finally, with the PLR and NLR targets set to 19 and 20, respectively, plasma pTau217 achieved a PPV of 95.8%, with the indeterminate zone further increasing to 43.5% (Table 2).

For the pTau217/Aβ_1–42 ratio, targeting a PLR of 10 and NLR of 20 resulted in a PPV of 92.4% and a NPV of 94.3%, with 22.2% of participants falling into the indeterminate zone. Increasing the PLR to 14 and maintaining the NLR of 20 increased the PPV to 94.4%, while the indeterminate zone reached 26.5%. Finally, targeting a PLR of 19 and NLR of 20, the PPV increased to 95.8%, with the indeterminate zone further increasing to 30.2% (Table 3).

Within all the tested conditions, pTau217 yielded a higher proportion of results in the indeterminate zone compared to the pTau217/Aβ_1–42 ratio (Tables 2 and 3), highlighting improved classification when the ratio was used.

To select optimal thresholds, we aimed to achieve acceptable PPV and PLR values while limiting the proportion of indeterminate test results, thereby balancing performance with the proportion of conclusive test results. While increasing the PLR target from 10 to 14 and 19 led to progressively higher PPV values for both pTau217 and the pTau217/Aβ_1–42 ratio, this gain came at the cost of a growing indeterminate zone. Notably, the 14/20 PLR/NLR target-derived thresholds produced higher PPVs than the 10/20 PLR/NLR target-derived thresholds for both biomarkers, indicating improved confidence in positive test results, while maintaining the indeterminate test results close to the 20% target. At the 19/20 PLR/NLR target-derived thresholds, both biomarkers achieved their highest PPVs, but the proportions of indeterminate results also reached their highest values, limiting the clinical utility of this stricter classification. Importantly, for both ptau217 and the pTau217/Aβ_1–42 ratio, the 14/20 PLR/NLR target-derived thresholds used to predict the combined CSF and PET reference standard were positioned more centrally between the CSF-only and PET-only thresholds. In contrast, the 19/20 PLR/NLR target-derived thresholds were skewed closer to the CSF-based values (Supplementary Tables 6 and 7). This balanced positioning of the 14/20 PLR/NLR target-derived thresholds for both biomarkers helps prevent bias towards either reference standard and supports broader clinical applicability. Thus, the 14/20 PLR/NLR target-derived thresholds offered a favourable balance between diagnostic certainty and classification performance. The PPVs at this level were close to those observed at the 19/20 PLR/NLR target-derived thresholds, yet the proportion of indeterminate results was smaller. Therefore, the 14/20 PLR/NLR target-derived threshold combination was chosen as the preferred option for further evaluation.

To assess the generalizability and robustness of this selected threshold derived from the parametric model, the 14/20 PLR/NLR target-derived thresholds were evaluated non-parametrically to categorize the 208 participants in the dataset. Non-parametric performance metrics are not expected to match parametric estimates exactly, as they are influenced by the empirical distribution of observed biomarker values. However, the degree of concordance between the two approaches is an important check on the validity and robustness of the parametric model. An additional advantage of the non-parametric resubstitution method is that it allows calculation of confidence intervals around performance metrics, thereby providing a more comprehensive evaluation of classification accuracy.

Non-parametric reclassification of participants using the 14/20 PLR/NLR target-derived threshold pair in both biomarkers confirmed the comparative advantages of the pTau217/Aβ_1–42 ratio over pTau217 alone. Specifically, the pTau217/Aβ_1–42 ratio achieved a slightly higher PPV (94.6% versus 92.4%) and a higher PLR (14.512 versus 10.032) compared to pTau217, indicating greater confidence in a positive classification. It is important to note that since the thresholds were computed from the parametric model with targets of 10, 14, and 19 PLRs and 20 NLR, the actual observed PLRs in the non-parametric model resulted in slight deviations from these target values, leading to minor differences in computed PLRs and NLRs. Although the NPV was slightly higher for pTau217 alone (92.9% versus 91.8%), the ratio once again resulted in a smaller proportion of indeterminate test results (20.2%) compared to pTau217 (35.1%), supporting its utility in minimizing diagnostic uncertainty (Tables 4 and 5). Additionally, the close concordance across methods strengthens confidence in the validity of the 14/20 PLR/NLR target-derived thresholds and their potential clinical utility. Non-parametric performance for pTau217 and the pTau217/Aβ_1–42 ratio by clinical diagnostic group are presented in Supplementary Tables 8 and 9.

Table 4.

Non-parametric performance of pTau217 for detecting amyloid positivity based on CSF + PET combined using 14/20 PLR/NLR target thresholds

PLR/NLR	Thresholds	Test result	True positives	True negatives	Total	Frequency (%)	LR (95% CI)	PV, % (95% CI)
14/20	≥0.23211	Positive	73	6	79	38.0	10.032 (5.045, 24.965)	92.41 (84.4, 96.5)
	0.08681 < Result < 0.23211	Indeterminate	37	36	73	35.1	0.847 (0.585, 1.229)	50.69 (39.5, 61.8)
	≤0.08681	Negative	4	52	56	26.9	0.063 (0.020, 0.147)	7.14 (2.8, 17.0)

Classification of test results for pTau217 using assay thresholds determined at PLR/NLR target of 14/20, evaluated against amyloid positivity as determined by CSF or PET imaging. Reported metrics include true positives, true negatives, frequencies for each test result category, likelihood ratios and predictive values. CI = confidence interval; LR = likelihood ratio; NLR = negative likelihood ratio; PLR = positive likelihood ratio; PV = predictive value.

Table 5.

Non-parametric performance of pTau217/Aβ_1–42 ratio for detecting amyloid positivity based on CSF + PET combined using 14/20 PLR/NLR target thresholds

PLR/NLR	Thresholds	Test result	True positives	True negatives	Total	Frequency (%)	LR (95% CI)	PV, % (95% CI)
14/20	≥0.00738	Positive	88	5	93	44.7	14.512 (6.929, 39.796)	94.62 (88.0, 97.7)
	0.00371 < Result < 0.00737	Indeterminate	20	22	42	20.2	0.750 (0.432, 1.289)	47.62 (33.4, 62.3)
	≤0.00370	Negative	6	67	73	35.1	0.074 (0.030, 0.147)	8.22 (3.8, 16.8)

Classification of test results for the pTau217/Aβ_1–42 ratio using assay thresholds determined at PLR/NLR target of 14/20, evaluated using amyloid positivity as determined by CSF or PET imaging. Reported metrics include true positives, true negatives, frequencies for each test result category, likelihood ratios and predictive values. CI = confidence interval; LR = likelihood ratio; NLR = negative likelihood ratio; PLR = positive likelihood ratio; PV = predictive value.

Lastly, to evaluate whether the 14/20 PLR/NLR target-derived threshold pair from the combined CSF ratio and PET model was applicable to individual reference standards, we assessed the performance of the pTau217/Aβ_1–42 ratio, given its superior performance compared to pTau217 alone, using samples with either CSF ratio only or PET results only under both parametric and non-parametric frameworks. In the parametric model, when the 14/20 PLR/NLR target-derived thresholds were applied to participants with CSF ratio results, the data yielded a PPV of 98.7% and a PLR of 58.72, with 24.1% of the results classified as indeterminate. When applied to participants with PET results, these thresholds yielded a PPV of 87.0%, a PLR of 6.24, and an indeterminate rate of 24.1% (Supplementary Tables 10 and 11). In the non-parametric model, for participants with CSF ratio results, the 14/20 thresholds yielded a PPV of 97.8%, and a PLR of 35.33, with 20.2% of results classified as indeterminate. For participants with PET results, the PPV reached 88.9%, the PLR was 7.45, and 20.6% of results were classified as indeterminate (Supplementary Tables 12 and 13). Supplementary Fig. 5 shows that plasma-indeterminate cases were generally CSF-positive but PET-negative, consistent with the expected temporal sequence of amyloid biomarker changes.

Discussion

In this study, we assessed the performance of AD plasma biomarkers, including pTau217, Aβ_1–42, Aβ_1–40 and the combination of two biomarkers through the pTau217/Aβ_1–42 ratio, as predictors of amyloid pathology associated with AD. We also determined thresholds that could be implemented in assessing symptomatic individuals on a spectrum from SCD to AD in speciality care centres (intended use population).

The initial multivariate analysis showed that plasma levels of both Aβ_1–42 and pTau217 were individually linked to disease status. On the other hand, Aβ_1–40 plasma levels did not provide any additional discriminatory predictive value and were excluded from further evaluation. Although evaluation of covariates in our study indicated that age and sex were not significantly associated with plasma pTau217 levels or the pTau217/Aβ_1–42 ratio, prior studies have reported demographic influences on plasma pTau217 alone. In a recent multi-cohort analysis using the Lumipulse platform, plasma pTau217 levels were significantly higher in women than in men and in younger (<73 years) versus older participants with AD pathology.⁴¹ Conversely, another study found that male sex was associated with higher plasma pTau217 concentrations.⁴³ These differences may reflect variations in cohort composition, sample size, or methodology, highlighting the importance of considering demographic influences when interpreting plasma biomarker levels. Although APOE ε4 status was associated with higher pTau217/Aβ_1–42 ratio values in our analysis, the accuracy of identifying AD pathology using the ratio was not affected. This may be explained by the fact that the shift in pTau217/Aβ_1–42 ratios associated with APOE ε4 genotype occurred proportionally in both amyloid-positive and amyloid-negative individuals, thereby preserving the separation between groups and maintaining classification accuracy.

Notably, the concentrations of pTau217 and Aβ_1–42 changed in opposite directions with disease progression. In a multivariate model, their coefficients were of similar magnitude but of opposite sign, supporting the rationale for combining them into a ratio. The resulting pTau217/Aβ_1–42 ratio demonstrated better classification accuracy than pTau217 levels alone, as evidenced by a higher AUC and improved sensitivity and specificity at the optimal Youden index threshold.

By recognizing the limitations of single-threshold approaches, and to enhance overall test accuracy and reduce the incidence of false positives, we also implemented a two-threshold (three-category) strategy, as previously recommended^26,27 for both plasma pTau217 and the pTau217/Aβ_1–42 plasma ratio.

The two-threshold approach enabled classification of participants into negative, indeterminate, and positive groups. While a positive result confirms amyloid positivity and a negative result rules it out, the indeterminate category identified individuals whose amyloid status remained uncertain. Importantly, this approach aimed to keep the proportion of indeterminate results at or below 20%, providing clinicians with clear guidance for appropriate follow-up evaluations, such as CSF or amyloid PET testing, or repeat BBM assessment, depending on clinical context.²⁴ Additionally, in the present study, as part of this approach, we incorporated likelihood ratio targets to optimize the post-test probability of amyloid pathology following a positive or negative test result. All performance metrics used to guide threshold selection were derived from a parametric model. For both biomarkers, the thresholds derived from the 14/20 PLR/NLR targets provided the best balance between a higher PPV and a manageable indeterminate zone compared with the low (10/20) and higher (19/20) PLR/NLR targets.

Given the non-normal distribution of biomarker values in this study, the non-parametric analysis was applied to the thresholds derived from the parametric model to verify their performance under empirical conditions. The results closely aligned with those from the parametric analysis, reinforcing the validity and clinical relevance of the 14/20 PLR/NLR thresholds. When comparing biomarkers, the pTau217/Aβ_1–42 ratio consistently demonstrated higher predictive accuracy and fewer indeterminate results than pTau217 alone. To further confirm the robustness of these thresholds, additional analyses were performed for the pTau217/Aβ_1–42 ratio using CSF and PET reference standards separately. The thresholds demonstrated similar strong performance when applied to CSF alone, while a slightly lower classification performance was observed when assessed against PET alone.

Other studies have reported similar advantages of the plasma pTau217/Aβ_1–42 ratio over pTau217 alone, particularly in reducing the proportion of indeterminate classifications and improving diagnostic certainty. For example, Martínez-Dubarbie et al.⁴⁴ analysed memory clinic patients and cognitively unimpaired volunteers using CSF as the reference standard, deriving plasma thresholds via ROC/Youden analysis with a two-threshold approach. In this study, both pTau217 and the pTau217/Aβ_1–42 ratio achieved 97% overall accuracy, but the ratio markedly reduced intermediate classifications (27.1% versus 42.6%). Wang et al.⁴⁵ studied Chinese clinical and community cohorts, employing Aβ PET with quantitative and validated visual reads, and applied both a single threshold (90% sensitivity) and a two-threshold strategy (95% sensitivity/95% specificity). In the clinical cohort, the pTau217/Aβ42 ratio achieved a lower percentage of intermediate cases than pTau217 alone (10.7% versus 13.0%), while maintaining comparable accuracy. In the community cohort, both biomarkers showed similarly high accuracy (93.8%), but the ratio again reduced the proportion of intermediate results (16.5% versus 31.4%). Palmqvist et al.⁴¹ conducted a multicentre European study across memory clinic and primary care cohorts using CSF and PET references to define one- and two-threshold strategies. In secondary care, accuracy was similar for pTau217 and the pTau217/Aβ42 ratio (≈90%), while the ratio was slightly lower in primary care (89% versus 92%). Nonetheless, the ratio reduced indeterminate results in both settings (7% versus 15% in secondary care; 10% versus 16% in primary care). Although these studies varied in reference standards and threshold selection methods, all used the Lumipulse automated platform for plasma measurements. Across these studies, the plasma pTau217/Aβ_1–42 ratio consistently reduced indeterminate results^41,44,45 compared to pTau217 alone. Notably, Martínez-Dubarbie et al.⁴⁴ and Wang et al.⁴⁵ included cognitively unimpaired participants in their analyses, whereas both Palmqvist et al.⁴¹ and our study focused exclusively on symptomatic individuals with cognitive concerns attending memory clinics, reflecting the current guidelines that limit the clinical use of blood-based biomarkers to patients with symptoms of cognitive decline. Taken together, these findings add to the growing body of evidence supporting the pTau217/Aβ_1–42 ratio as a clinically relevant composite biomarker that can aid diagnostic assessment and reduce the need for confirmatory testing.

Given that clinical diagnosis of AD dementia can be inaccurate, with amyloid positivity rates of 80%–90% in diagnosed cases and misdiagnosis rates as high as 30%, the incorporation of biomarker-based approaches is essential to improve diagnostic accuracy. In our cohort, amyloid pathology was detected in 54.8% of the total study population. Among participants clinically diagnosed with AD dementia, 83.3% tested positive for amyloid pathology by CSF/PET, a rate consistent with previous studies. For instance, the AIBL and ADNet cohorts reported amyloid positivity in 117 of 132 dementia cases (88.6%) despite the clinical diagnosis being made by multidisciplinary panels blinded to biomarker data.⁴⁶ Similarly, the BioFINDER study found that clinical evaluation alone achieved 58% accuracy in identifying underlying AD pathology, compared to 89%–90% when using plasma biomarkers.³² Misdiagnosis of dementia is well documented, with post-mortem studies reporting error rates as high as 30% and meta-analyses confirming that clinical diagnosis has limited sensitivity.⁴⁷ These observations suggest that the amyloid positivity rate seen in clinically diagnosed AD dementia cases in our cohort aligns with expected values and underscores the importance of incorporating biomarkers such as pTau217 and Aβ_1–42 to identify amyloid pathology and improve diagnostic accuracy.

A key strength of our study is the implementation of a two-threshold approach optimized to achieve high predictive values and a low proportion of indeterminate results. By achieving recommended performance metrics (PPV ≥90%, NPV ≥90%, and indeterminate results ≤20%) as outlined in the recent consensus statement from the Global CEO Initiative on Alzheimer's Disease,²⁴ we established clinically meaningful thresholds that balance diagnostic certainty with real-world practicality. The predictive values were calculated based on an amyloid positivity prevalence of 54.8%, consistent with the rate of amyloid positivity observed in both the study cohort and secondary care. This approach leverages the accessibility of blood-based testing while aligning with current standards of care, which typically rely on a CSF ratio measurement or amyloid PET imaging to confirm amyloid pathology. It is important to note that this study evaluated the performance of the Lumipulse pTau 217/Aβ1–42 Plasma Ratio only in symptomatic individuals assessed in secondary care, and not in the cognitively unimpaired individuals or those in the preclinical stage of AD. Within this population, however, our prevalence-independent analysis using likelihood ratios enhances the generalizability of our findings across different clinical settings where symptomatic patients are evaluated, regardless of the prevalence of disease, and thus provides additional information on the effectiveness of the diagnostic test.

Currently, both CSF and PET biomarkers are FDA-cleared and accepted as reference standards for amyloid pathology. To better reflect real-world clinical practice, where either modality may be used depending on availability, patient characteristics, or physician preference, we incorporated both in our approach. Threshold selection was therefore based on combined CSF biomarker ratio and PET test results, with participants classified as amyloid-positive if either test was positive, including discordant cases. This strategy helps reduce bias towards a single reference standard and promotes diagnostic consistency across modalities, which is particularly important given the known discrepancies between CSF biomarker ratios and PET-based amyloid positivity classification.^48,49 Specifically, these studies have demonstrated that CSF Aβ_1–42 levels may become abnormal earlier than amyloid PET, resulting in cases where individuals are classified as amyloid-positive by CSF analysis but negative by PET imaging. This highlights the importance of considering both reference standards when establishing diagnostic thresholds.

Furthermore, the use of a fully automated chemiluminescent enzyme immunoassay (CLEIA) platform offers further advantage, with recent cross-platform comparisons demonstrating that CLEIA-based assays perform on par with mass spectrometry methods for detecting amyloid pathology.⁵⁰ In addition to strong clinical performance, CLEIA testing provides practical benefits, including widespread instrument availability, fast turnaround time, affordability, and scalability, facilitating broader access to accurate biomarker testing across healthcare settings.

A few limitations in this study should also be acknowledged. Our cohort composition was racially unbalanced, which may affect the generalizability of our findings across diverse populations. However, Cousins et al.⁵¹ recently showed that the plasma pTau217/Aβ_1–42 ratio performs comparably in Black/African American and White participants. Discriminative accuracy for detecting amyloid PET positivity was similarly high in Black/African American (AUC = 0.88) and White (AUC = 0.91) individuals, with no significant interaction between race and amyloid PET status, and a consistently low proportion of indeterminant classifications. This study's results suggest that race does not substantially influence the diagnostic accuracy of the ratio in these racial groups. Nevertheless, additional studies in more diverse populations, including Hispanic and other underrepresented groups, are needed to establish generalizability and ensure robust applicability across demographics and settings. Additionally, while in our study age, sex, and APOE ε4 carriership did not impact biomarker performance, other potential confounders, such as comorbidities, medication use, and body mass index, were not comprehensively evaluated. Although these factors were not the focus of the current analysis, they may influence biomarker levels⁵² and warrant further investigation.

As previously mentioned, the intended use population for the Lumipulse G pTau 217/Aβ 1–42 Plasma Ratio (and the population evaluated in this study) comprises symptomatic individuals across the spectrum from SCD to AD. However, due to the small sample sizes in the AD (n = 12) and SCD (n = 7) groups (Supplementary Table 9), further research is warranted to better assess the ratio's performance in these individual diagnostic categories.

Additionally, while we recognize the value of Centiloid conversion, it was not possible at the time of analysis.

Finally, our study did not include prospective plasma analyses over longer time periods, which would better mimic the use of the test in clinical practice. Future clinical performance studies should validate the established plasma ratio thresholds in larger, more diverse cohorts and clinical settings.

In conclusion, this study demonstrates that the Lumipulse G pTau 217/Aβ 1–42 Plasma Ratio outperforms Lumipulse G plasma pTau 217 alone in detecting amyloid pathology based on amyloid PET/CSF results, providing greater diagnostic accuracy and a reduced rate of indeterminate results. The use of a two-threshold approach (which provides a negative, positive, or indeterminate result) for amyloid pathology classification, and the implementation of the PLR 14/NLR 20 target-derived thresholds for the pTau217/Aβ_1–42 plasma ratio, achieved the highest diagnostic accuracy and reduced patient misclassification, providing a framework for clinical implementation that reduces diagnostic uncertainty while minimizing the need for additional testing. These findings represent an essential step towards developing accessible and reliable blood-based biomarkers for AD, with significant implications for clinical practice, research participant selection, and access to treatment.

Data availability

The data that support the main findings of this study are available from the corresponding author, on reasonable request.

Funding

The BioFINDER study was funded by the National Institute on Aging (#R01AG083740), the Alzheimer’s Association (#SG-23-1061717, ALZSI-26-1523522), the Swedish Brain Foundation (#FO2024-0284, FO2025-0055), the Kamprad Foundation (#20243058), the Family Rönström (#FRS-0011 and FRS-0004, FRS-0013), the Swedish Alzheimer Foundation (#AF-1011949, AF-1011799), Lilly Research Award Program, Michael J. Fox Foundation (MJFF-025507), Swedish federal government under the ALF agreement (2022-Projekt0107), Bundy Academy and MultiPark at Lund University. ¹⁸F-Flutemetamol was provided by GE Healthcare. Fujirebio provided funding and covered the costs associated with the collection, handling, and analysis of the material. The BIOCARD study is supported by two grants from the National Institute on Aging (NIA) (#P30AG066507 and #U19AG033655).

Competing interests

N.B. is an employee of Fujirebio Diagnostics Inc. L.B. is an employee of Fujirebio Diagnostics Inc. F.I.D.S. is an employee of Fujirebio Diagnostics Inc. R.R.R. is an employee of Fujirebio Diagnostics Inc. M.C.M. is an employee of Fujirebio Diagnostics Inc. K.M. is an employee of Fujirebio Diagnostics Inc. D.D. is an employee of Fujirebio Diagnostics Inc. S.H. has no conflict of interest to declare. A.M. has acquired research support (for the institution) from Fujirebio Diagnostics Inc. M.A has no conflict of interest to declare. N.M.-C. has received consultancy and/or speaker fees from BioArctic, Biogen, Eli Lilly, Merck, Novo Nordisk, and Owkin. S.P. has acquired research support (for the institution) from Avid and ki elements through ADDF. In the past 2 years, he has received consultancy/speaker fees from Bioartic, Biogen, Eisai, Eli Lilly, Novo Nordisk, and Roche. R.O. has received research funding/support from Avid Radiopharmaceuticals, Janssen Research & Development, Roche, Quanterix and Optina Diagnostics, has given lectures in symposia sponsored by GE Healthcare, received speaker fees from Springer, is an advisory board/steering committee member for Asceneuron, Biogen, Johnson & Johnson and Bristol Myers Squibb. All the aforementioned have been paid to his institutions. M.F.J. has no conflict of interest to declare. O.H. is an employee of Lund University and Eli Lilly. E.S. has acquired research support (for the institution) from C2N Diagnostics, Fujirebio Diagnostics Inc., GE Healthcare and Roche. P.S. is an employee of Eisai Co., Ltd. H.N. is an employee of Eisai Co., Ltd. D.V. is an employee of Eisai Co., Ltd. D.M.H. serves as a consultant for Fujirebio Diagnostics, Inc.

Supplementary material

Supplementary material is available at Brain online.