Clinical performance of scalable automated p‐tau 217 multi‐analyte algorithmic blood test with reduced intermediate zone using multiplexed digital immunoassay

Abstract

INTRODUCTION

To address an urgent need for a scalable, accurate blood test for brain amyloid pathology that provides a conclusive result for the greatest number of patients, we developed a multi‐analyte algorithmic test combining phosphorylated tau (p‐tau) 217 with four other biomarkers.

METHODS

Multiplexed digital immunoassays measured p‐tau 217, amyloid beta 42/40, glial fibrillary acidic protein, and neurofilament light chain in 730 symptomatic individuals (training set) to establish an algorithm with cutoffs, and 1082 symptomatic individuals (validation set) from three independent cohorts to identify brain amyloid pathology.

RESULTS

The algorithmic in validation gave an area under the curve = 0.92, yielding 90% agreement with amyloid positron emission tomography and cerebrospinal fluid. Positive predictive value was 92% at 55% prevalence. The multi‐marker algorithm reduced the intermediate zone ≈ 3‐fold from 34.4% to 11.9% versus p‐tau 217 alone. Diagnostic performance was similar across subgroups.

DISCUSSION

The LucentAD Complete multi‐analyte blood test demonstrated high clinical validity for brain amyloid pathology detection while substantially reducing inconclusive intermediate results.

Highlights

• We developed a multi‐analyte blood test for assessing brain amyloid status that significantly minimizes the ambiguous “intermediate zone,” a key limitation of plasma phosphorylated tau (p‐tau) 217 alone.

• Our test combines plasma levels of p‐tau 217, amyloid beta 42/40 ratio, glial fibrillary acidic protein, and neurofilament light chain for a more comprehensive evaluation of amyloid status.

• We rigorously validated the test's clinical performance in > 1000 samples from symptomatic individuals across three independent cohorts, using cerebrospinal fluid biomarkers and amyloid positron emission tomography as comparators.

1. BACKGROUND

The advent of new Alzheimer's disease (AD) therapies underscores the urgent need for accessible and scalable early diagnostic tools. While clinical diagnosis has historically been the standard, biomarkers, particularly amyloid and phosphorylated tau (p‐tau) markers, are increasingly crucial for an etiological diagnosis and eligibility for amyloid targeting therapies ¹ , ² . Traditionally, these biomarkers required costly and invasive positron emission tomography (PET) or cerebrospinal fluid (CSF) assessments. However, recent rapid advancements in blood‐based biomarker tests offer non‐invasive alternatives for amyloid pathology assessment, promising expanded access to diagnosis through simpler, more efficient procedures. Among blood‐based biomarkers, tau phosphorylated at threonine 217 (p‐tau 217) has emerged as the most accurate single biomarker for detecting amyloid pathology in AD, with the Alzheimer's Association (AA) recommending it as the only blood‐based biomarker comparable to US Food and Drug Administration (FDA)–cleared CSF tests for diagnostic use when analyzed with high‐performance technologies. ¹ Recognizing overlap between amyloid‐positive and ‐negative assay signals, the AA and others advocate for a two‐cutoff test design, creating an intermediate risk zone ³ . To maximize clinical utility, the AA and Global CEO Initiative (CEOi) recommend an amyloid classification accuracy of ≥ 90% and an intermediate zone of ≤ 20% ⁴ . Accurately classifying borderline cases within this intermediate zone remains a key challenge.

Beyond p‐tau 217, several other established plasma biomarkers are relevant to detecting the presence of amyloid pathology. These biomarkers, including the amyloid beta 1‐42/1‐40 ratio (Aβ42/Aβ40), glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL), reflect amyloid pathology directly or detect the presence of AD‐associated disease pathways. Aβ42/Aβ40 directly reflects amyloid plaque development, ⁵ GFAP indicates astrocytic activation linked to amyloid pathogenesis, ⁶ and NfL signals neuroaxonal damage in neurodegenerative diseases like AD ⁷ . While these biomarkers can individually predict AD and their predictive power is enhanced in combination ⁸ , they are less accurate than p‐tau 217 alone ⁹ , and combining them with p‐tau 217 has shown some, but minimal, overall accuracy improvement ¹⁰ , ¹¹ , ¹² . We hypothesized that these additional markers reflecting ongoing AD processes could augment amyloid classification in p‐tau 217 borderline cases, thereby reducing the intermediate zone.

In this multi‐cohort clinical validation study, we therefore aimed to establish an algorithm augmenting the accuracy of p‐tau 217 for cases with low and high amyloid burden with additional amyloid classification ability for borderline cases through the addition of Aβ42/Aβ40, GFAP, and NfL. This approach combines the LucentAD p‐tau 217 immunoassay with a single multiplexed Simoa digital immunoassay (N4PE). This report describes the test and its clinical performance based on a multi‐cohort clinical validation study using pre‐established diagnostic cutoffs. These validation data support the availability of the test under Clinical Laboratory Improvement Amendments (CLIA; as “LucentAD Complete”) for clinical use to provide accurate results to the greatest number of patients with a simple, scalable immunoassay format.

2. METHODS

2.1. Cohorts and reference methods

2.1.1. Amsterdam Dementia Cohort

The Alzheimer Center Amsterdam Dementia Cohort (ADC) comprised patients referred for cognitive evaluation by general practitioners or specialists. All patients underwent a standardized, multidisciplinary work‐up, including neurological history and examination, vital function assessment, informant history, dementia nurse consultation, neuropsychological testing, brain magnetic resonance imaging, electroencephalogram, standard laboratory tests, and either lumbar puncture for CSF biomarker analysis or amyloid PET imaging ¹³ , ¹⁴ . Diagnoses were determined via multidisciplinary consensus, according to established diagnostic guidelines ¹⁵ , ¹⁶ . Here, AD dementia required abnormal CSF biomarkers or positive amyloid PET. Amyloid PET scans used [18F]Florbetaben or [18F]Florbetapir, with positivity defined by visual assessment of neocortical fibrillary amyloid by a nuclear medicine physician. CSF amyloid positivity was determined using Roche Elecsys p‐tau 181/Aβ42 assays (cut‐off 0.02) ¹⁷ or Fujirebio Innotest p‐tau 181/Aβ42 enzyme‐linked immunosorbent assay (ELISA; cut‐off 0.06) ¹⁸ . As the intended use population is objectively impaired patients, mild cognitive impairment (MCI; n = 223) and AD dementia (n = 272) cases were selected to comprise part of the training set. Another 178 MCI cases and 96 AD dementia cases were included in the validation set.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed the growing body of literature on blood‐based biomarkers for Alzheimer's disease (AD)–related brain amyloid pathology. This review highlighted the emerging utility of plasma phosphorylated tau (p‐tau) 217 as a key indicator, but also revealed the potential of multi‐analyte approaches to reduce inconclusive indeterminate results while maintaining high overall diagnostic accuracy.q

2. Interpretation: We demonstrated that a novel multi‐analyte algorithmic blood test incorporating plasma p‐tau 217, the amyloid beta 42/40 ratio, glial fibrillary acidic protein, and neurofilament light chain, significantly reduces the inconclusive “intermediate zone” compared to p‐tau 217 alone. This algorithmic approach leverages the complementary information from these biomarkers to provide a more robust assessment of brain amyloid status in symptomatic individuals, enabling accurate classification of two thirds of intermediate cases from p‐tau 217 alone.

3. Future directions: Further research in larger, more diverse cohorts, including longitudinal studies and comparisons across different clinical settings, will help fully establish the clinical utility and optimal implementation of this multi‐analyte blood test in the diagnostic pathway for AD. Studies exploring its performance in pre‐symptomatic individuals and its potential for monitoring disease progression are also warranted.

2.1.2. Bio‐Hermes cohort

From April 2021 to November 2022, 17 clinical trial sites recruited community‐based participants for the Bio‐Hermes cohort, aiming to enrich ethnic/racial diversity. Participants, meeting established inclusion criteria, were categorized as cognitively unimpaired, MCI, or clinical mild AD dementia ¹⁹ . MCI participants had a documented MCI diagnosis (National Institute on Aging–AA criteria ²⁰ ) or met screening criteria: Mini‐Mental State Examination (MMSE) 24 to 30, Rey Auditory Verbal Learning Test (RAVLT) delayed recall ≥ 1 standard deviation (SD) below age‐adjusted mean, and minimal functional impairment (Functional Activities Questionnaire [FAQ]). Mild AD dementia participants had a probable AD diagnosis ²⁰ or met screening criteria: MMSE 20 to 24, RAVLT delayed recall ≥ 1 SD below age‐adjusted mean, and evidence of functional decline (FAQ). All participants underwent [18F]Florbetapir amyloid PET scans, interpreted centrally by IXICO Technologies Inc. Underrepresented groups (Hispanic and non‐Hispanic Black) comprised 27.8% of the symptomatic sub‐cohort (MCI and mild AD). Consistent with the intended use population, only MCI (n = 136) and mild AD (n = 99; clinical diagnosis, not PET confirmed) cases were included in the analyses. Another 149 MCI cases and 122 mild AD cases were randomly selected to comprise part of the validation set.

2.1.3. Alzheimer's Disease Neuroimaging Initiative cohort

The Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, derived from the ADNI database (adni.loni.usc.edu) ²¹ , comprised participants from a Foundation for the National Institutes of Health Biomarker Consortium study examining longitudinal trajectories of blood‐based biomarkers in relation to amyloid PET ²² . ADNI, initiated in 2003, aims to assess the progression of MCI and early AD using multimodal biomarkers. For this study, participants with plasma samples collected within 6 months of amyloid PET at three distinct time points were selected. From 406 subjects with 1231 samples (2–5 time points), 1010 had amyloid status and all five biomarkers assayed. Exclusion of samples outside the intended use population yielded 537 samples from 236 individuals, spanning up to 180 months.

2.2. Plasma sample analysis

2.2.1. Instrumentation

All assay testing was performed on the Simoa HD‐X instrument, a fully automated digital immunoassay analyzer using Simoa technology for rapid isolation and counting of single molecules. The HD‐X is a sequential cuvette processing robot with a clock speed of 66 samples/hour. Details of the instrument and its principles are given elsewhere ²³ .

2.2.2. Assay principle and protocol

Simoa technology, a digitized bead‐based ELISA, achieves attomolar sensitivity through single‐molecule detection within 40 femtoliter microwells ²⁴ . By confining fluorescent reporter molecule diffusion, single enzyme labels generate detectable signals within 30 seconds. Arrays of 216,000 wells enable rapid, simultaneous counting, resulting in assay processing time for a single sample in 45 to 60 minutes, and a batch run time of 2 to 4 hours.

The Simoa p‐tau 217 assay, a three‐step sandwich immunoassay, involves capture, sandwich formation with biotinylated detector antibodies, and labeling with a streptavidin‐β‐galactosidase conjugate ²⁵ . After magnetic bead collection and washing at each step, beads are resuspended in resorufin β‐D‐galactopyranoside substrate. Digital processing occurs upon bead transfer to the Simoa array disc, where captured p‐tau 217 leads to substrate hydrolysis and fluorescence. p‐tau 217 concentration is determined via four‐parameter logistic curve interpolation. p‐tau 217 concentration is determined via four‐parameter logistic curve interpolation for results read out.

The commercially available Simoa N4PE simultaneously measures Aβ40, Aβ42, GFAP, and NfL using a two‐step digital immunoassay. Detector antibodies are combined with analyte‐specific beads pre‐coated with distinct fluorescent dyes. After substrate resuspension and transfer to the array, bandpass filters identify analyte‐specific beads, enabling multiplexing. Single‐sample processing time for all four analytes is ≈ 1 hour.

2.2.3. Sample collection and testing

K2 ethylenediaminetetraacetic acid (EDTA) plasma was collected via venipuncture for all three cohorts. For the ADC, samples were centrifuged within 2 hours (1800 × g, 10 minutes, room temperature), aliquoted (≤ 0.5 mL), and stored at −80°C until use at the Neurochemistry Laboratory Amsterdam or dry‐ice shipment to Quanterix. For the Bio‐Hermes cohort, 2.0 mL whole blood was transferred to conical tubes and centrifuged (1500 × g, ≥15 minutes, room temperature), with plasma aliquoted and frozen at −80°C within 4 hours. For the ADNI cohort, plasma was centrifuged within 1 hour (1500 × g, 15 minutes, room temperature), aliquoted, and frozen at −80°C. Amsterdam‐based samples were thawed, vortexed, and centrifuged per manufacturer's instructions (10 000 × g, 10 minutes) before being analyzed. Samples were run in singlicate for Aβ40, Aβ42, GFAP, and NfL using the Simoa N4PE kit, with calibration and quality controls (QCs) run in duplicate. Inter‐assay coefficients of variation (CVs) were < 15% ¹⁸ . For testing at Quanterix, samples were thawed (60 minutes, room temperature), centrifuged (10000 × g, 10 minutes), and analyzed, with calibrators, controls, and QCs according to Quanterix standard operating procedures.

QC samples, included in all p‐tau 217 and N4PE runs, demonstrated acceptable precision, consistent with previous reports ¹⁸ , ²⁵ . The combined precision of all five biomarkers, estimated by evaluating inter‐assay imprecision from a multi‐day precision study for each assay, was 7.33% (95% confidence interval [CI]: 6.87%–7.80%; Figures S1, S2 in supporting information).

2.2.4. Training and validation cohorts

The training set for the multi‐analyte algorithm was composed of cases from two cohorts: the ADC and the Bio‐Hermes study. The training set comprised 495 ADC cases (45% MCI, 55% AD dementia) and 235 Bio‐Hermes cases (58% MCI, 42% mild AD). The validation set, which was unseen during training, included a mix from the same two cohorts plus an additional cohort: 274 ADC cases (65% MCI, 35% AD dementia), 271 Bio‐Hermes cases (55% MCI, 45% mild AD), and 545 ADNI cases (83% MCI, 17% AD dementia). While training and validation sets from ADC and Bio‐Hermes had similar demographics due to blinded randomization (Table S1 in supporting information), the two cohorts themselves had significant demographic differences in age and amyloid prevalence (P < 0.0001).

The ADC training samples were analyzed at the Neurochemistry Laboratory Amsterdam, while the Bio‐Hermes training samples were analyzed at the Quanterix CLIA laboratory with a different reagent lot. To ensure consistency in biomarker values across the different laboratories and reagent lots, we used a bridging process, which involved testing a crossover set of 100 samples at both sites to normalize values. The bridging of each analyte used the full equations from the Passing–Bablok regression. The overall mean percent differences between the laboratories for NfL, p‐tau 217, Aβ42/40, and GFAP were 15%, 16%, 17%, and 22%, respectively. Importantly, these differences were consistent across the full range of each assay.

The multi‐analyte algorithm thresholds were established using the normalized p‐tau 217 and N4PE data from the combined ADC and Bio‐Hermes training sets. The p‐tau 217 clinical thresholds used in the algorithm were established previously at the Quanterix CLIA laboratory with a single reagent lot, as described ²⁵ . The validation of the multi‐analyte thresholds was conducted at the Quanterix CLIA laboratory using the combined ADC and Bio‐Hermes validation sets and the independent ADNI cohort. Two different reagent lots were used for the validation testing. These lots read samples consistently and did not require bridging.

2.3. QC monitoring

Depending on the analyte, two or three QC controls in EDTA plasma were included with each instrument run. Inter‐assay CVs across the runs ranged from 3.0% to 8.9% at the following concentrations (pg/mL): p‐tau 217 at 0.0860, 0.358, and 1.53; NfL at 23.5 and 516; GFAP at 205.7 and 4243.2; Aβ40 at 9.4 and 361; and Aβ42 at 7.50 and 36.1. These precision results align well with the separate multi‐day precision study conducted to model the amyloid risk score imprecision, as described in Figures S1, S2.

2.4. Statistical methods

All data analysis was performed using JMP Pro 18 (JMP Statistical Discovery LLC). Discovery analyses, using multiple modeling techniques, graphical data interrogations, and performance metrics, resulted in the LucentAD Complete test configuration, consisting of a locked logistic regression formula and thresholds. This test configuration was subsequently validated in three independent cohorts, assessed individually and with the validation cohorts combined. Performance metrics, including 95% CIs, are reported, with specific mention of instances in which indeterminate zone results were excluded. Details on the calculation of the accuracy estimates are provided in the supplemental materials, Section C, Table S2 in supporting information.

The locked test algorithm builds upon established thresholds for the p‐tau 217 assay ²⁵ . Initially, these thresholds were used to identify intermediate zone cases. A logistic regression model was constructed using all five biomarkers across all cases in the training dataset. A separate logistic regression transformed p‐tau 217 concentrations into a model score. These two models were then combined using a decision tree based on the initial p‐tau 217 thresholds. Thresholds for the resulting multi‐analyte model score were determined by fitting best‐fit distributions to the negative and positive training samples. Specifically, the upper threshold was set at the 93rd percentile of a sinh‐arcsinh distribution fit to the negative samples (corresponding to a 7% false positive rate), and the lower threshold was set at the seventh percentile of a beta distribution fit to the positive samples (corresponding to a 7% false negative rate). These 7% false negative and false positive rates were selected to establish robust thresholds, aiming for 10% false negative and false positive rates in independent validation cohorts.

3. RESULTS

3.1. Demographic and clinical characteristics

Demographic and clinical characteristics of the three validation cohorts and the combined validation set stratified by amyloid status are detailed in Tables S3 and S4 in supporting information, respectively. The combined validation cohort (n = 1082) had a mean age of 69.9 years (SD 7.8) and was 48.0% female. The mean age varied across cohorts: ADC (64.9 years), Bio‐Hermes (73.4 years), and ADNI (70.6 years). The validation set was predominantly White (93.3%), with Bio‐Hermes contributing the majority of underrepresented minorities. All participants were symptomatic, with diagnoses of MCI (71.3%), mild AD dementia (11.3%), or AD dementia (17.5%); 43.6% were apolipoprotein E (APOE) ε4 carriers. Amyloid prevalence varied significantly across cohorts and clinical subgroups, ranging from 34.7% in Bio‐Hermes MCI cases to > 99% in ADC AD dementia cases. The combined validation set had an overall amyloid prevalence of 55.4%.

3.2. Analyte measurement in plasma samples

K2 EDTA plasma samples were analyzed for p‐tau 217, Aβ42/Aβ40, GFAP, and NfL. Figure 1 exhibits individual biomarker results stratified by amyloid status across all samples (n = 1812) to highlight overall data trends and analytical capability of the assays to quantify all target analytes. All results were above the assay limits of detection, and 99.5% of p‐tau 217 samples exceeded the lower limit of quantification (LLoQ), with all other assays yielding 100% LLoQ‐exceeding results. Median p‐tau 217 concentration was 3.14‐fold higher in amyloid‐positive participants (P < 0.0001), with an overall area under the curve (AUC) of 0.91 (95% CI: 0.89–0.92). Amyloid status differences were statistically significant for Aβ42/40, GFAP, and NfL (P < 0.05). Their respective AUCs were 0.72 (95% CI: 0.70–0.75), 0.75 (95% CI: 0.72–0.77), and 0.55 (95% CI: 0.53–0.58).

FIGURE 1.

Biomarker levels (n = 1812) across training (o) and validation (+) cohorts, stratified by amyloid status. All biomarkers showed statistically significant differences between amyloid‐positive and ‐negative groups. Dashed lines indicate the established intermediate p‐tau 217 range (0.04–0.09 pg/mL). Continuous variables were used for all biomarkers in the algorithm, eliminating the need for discrete diagnostic thresholds for Aβ42/40, GFAP, and NfL. Aβ, amyloid beta; GFAP, glial fibrillary acidic protein; NfL, neurofilament light chain; p‐tau, phosphorylated tau.

3.3. Development of multi‐analyte algorithm

With the training set (n = 730), we explored the potential benefit of combining p‐tau 217 with other biomarker assays. We observed a stepwise increase in classification accuracy for cases within the p‐tau 217 intermediate range when more biomarkers were included. After testing various model types, a nominal logistic regression model was selected for its simplicity and consistent performance. Non‐tau pathology markers (Aβ42/40 ratio, NfL, and GFAP) exhibited statistically significant, non‐zero parameters in the models, with improved model metrics upon their inclusion. Notably, despite NfL's limited standalone amyloid discrimination, it demonstrated the strongest statistical contribution to the all‐biomarker multivariate model (P < 0.0001), surpassing the Aβ42/40 ratio (P < 0.0048). Applying these three‐ and five‐biomarker models to the training cohort, we observed slight but non‐significant increases in AUC (0.91 to 0.92) across models (Figure S3 in supporting information).

To better characterize the potential of the additional biomarkers to classify amyloid status in the p‐tau 217 intermediate zone, we derived thresholds from the best‐fit distributions of the multi‐analyte score. Compared to p‐tau 217 alone, the intermediate zone decreased from 31.2% to 14.9% of samples with the three‐biomarker model, and further to 10.5% with the five‐biomarker model (Table 1). This model's stability and marker necessity were confirmed via forward stepwise selection, a self‐validating ensemble model that uses internal cross‐validation, and multinomial regression, with all biomarkers demonstrating significant contributions (P < 0.0083 for GFAP, P < 0.0001 for others). Therefore, the five‐biomarker model was selected for subsequent validation.

TABLE 1.

Comparison intermediate zones of p‐tau 217 and two multi‐analyte models.

(A)
	p‐tau217
	Low risk		Intermediate		High risk		All
Amyloid status	N	Row %	N	Row %	N	Row %	N
PET/CSF
Positive	34	6.6%	149	29.0%	331	64.4%	514
Negative	130	60.2%	79	36.6%	7	3.2%	216
All	164	22.5%	228	31.2%	338	46.3%	730

(B)
	p‐tau217 + Aβ42/Aβ40
	Low risk		Intermediate		High risk		All
Amyloid status	N	Row %	N	Row %	N	Row %	N
PET/CSF
Positive	42	8.2%	69	13.4%	403	78.4%	514
Negative	158	73.1%	40	18.5%	18	8.3%	216
All	200	27.4%	109	14.9%	421	57.7%	730

(C)
	p‐tau217 + Aβ42/Aβ40 + GFAP + NfL
	Low risk		Intermediate		High risk		All
Amyloid status	N	Row %	N	Row %	N	Row %	N
PET/CSF
Positive	46	8.9%	41	8.0%	427	83.1%	514
Negative	162	75.0%	36	16.7%	18	8.3%	216
All	208	28.5%	77	10.5%	445	61.0%	730

Abbreviations: Aβ, amyloid beta; CSF, cerebrospinal fluid; GFAP, glial fibrillary acidic protein; NfL, neurofilament light chain; PET, positron emission tomography; p‐tau, phosphorylated tau.

The highlighted boxes signifies the crucial novel finding from this study: that the addition of multiple biomarkers reduced the inconclusive intermediate zone in stepwise fashion with increasing biomarkers: 217> 217+amyloid ratio> 217+amyloid ratio+GFAP+NfL.

To perform the test, samples are processed concurrently on the p‐tau 217 and N4PE 4‐plex assays. A multi‐marker model score is calculated using a decision tree, initially stratifying samples based on p‐tau 217 concentration thresholds. Figure 2 illustrates the test's workflow: identifying p‐tau 217 intermediate zone cases (Figure 2A) and analyzing these cases with the multi‐analyte algorithm (Figure 2B). The algorithm was focused only on the samples that could not be classified with p‐tau 217 to avoid the introduction of unnecessary noise. The test's output for all samples is a risk score between 0 and 1, which is multiplied by 100 and rounded to the nearest whole number for user convenience.

FIGURE 2.

Multi‐analyte interrogation of plasma samples for amyloid status. A, p‐tau 217 classifies samples with low and high amyloid burden, identifying intermediate risk cases. B, Multivariate analysis of all five biomarkers resolves approximately two thirds of intermediate cases. All results are converted to an amyloid risk score for final interpretation. p‐tau, phosphorylated tau.

3.4. Clinical performance validation

Clinical performance of the multi‐analyte algorithmic test was assessed with the pre‐defined p‐tau 217 and multi‐analyte logistic thresholds in the validation datasets, which included the ADC cohort (n = 274), Bio‐Hermes validation sub‐cohort (n = 271), and the ADNI cohort (n = 571), as detailed in Table S3. All 1082 validation samples were tested in the Quanterix CLIA laboratory. In addition to the full sub‐cohort of ADNI samples, the individuals from this set having two or three longitudinal intervals from baseline in common (24 and 48 months) were analyzed separately at each time point. Figure 3 depicts the results across the three validation cohorts. Overall, 88.1% of the results were amyloid classifiable after applying the algorithm, with 91.1% true positives and 90.7% true negatives identified. Of 372 p‐tau 217 intermediate zone cases (34.4% of the ≥ validation set), 55% were amyloid‐positive by PET/CSF. The multi‐analyte algorithm successfully reclassified 243 (67%) of these cases, reducing the indeterminate group to 129.

FIGURE 3.

Diagnostic performance of LucentAD Complete amyloid risk score across the full validation data set. Density of test results are exhibited in (A), while individual results from each validation cohort, separated based on reference method amyloid status, are depicted in (B). The gray box shows the lower and upper cutoffs (45, 70) and the intermediate zone pre‐established with separate training cohorts (not depicted here). Vertical dashed lines in (A) represent the intermediate zone of p‐tau 217 alone. ADC, Amsterdam Dementia Cohort; ADNI, Alzheimer's Disease Neuroimaging Initiative BH, Bio‐Hermes; CSF, cerebrospinal fluid; PET, positron emission tomography; p‐tau, phosphorylated tau.

Table 2 summarizes the clinical performance characteristics of the multi‐analyte test in the validation cohort. Overall, the combined validation statistics met the target accuracy of ≥ 90% irrespective of the calculation method (Section C in supporting information) and exhibited an intermediate range of 11.9%, well below the proposed maximum of 20% ⁴ . Importantly, the accuracy was statistically indistinguishable from the accuracy of p‐tau 217 alone as reported previously with a similar validation cohort ²⁵ . The results demonstrate robust and reproducible performance across the diversity of the three cohorts and in relation to amyloid prevalence rates. This diversity included variations in race/ethnicity, age, geography, clinical settings, and reference methods. The test maintained positive predictive values (PPVs) > 90% across amyloid prevalence rates ranging from 41% to 70%. Model‐derived PPVs and negative predictive values (NPVs) across a broader range of amyloid prevalence rates are detailed in Table S5 in supporting information.

TABLE 2.

Clinical performance characteristics of LucentAD Complete.

(A)					(B)
	Cross sectional		Longitudinal	Combined validation	Longitudinal, single sample/individual
	Training	Validation	Validation		Baseline	24 mo.	48 mo.
Cohort	ADC + BH	ADC + BH	ADNI	ADC + BH + ADNI	ADNI	ADNI	ADNI
n	730	545	537	1,082	115	107	111
AUC	0.920 (0.894–0.939)	0.910 (0.880–0.934)	0.923 (0.897–0.943)	0.916 (0.896–0.932)	0.941 (0.876–0.973)	0.939 (0.874–0.971)	0.883 (0.802–0.934)
Aβ prevalence	70.4% (67.0%–73.6%)	58.2% (54.0%–62.2%)	52.5% (48.3%–56.7%)	55.4% (52.4%–58.3%)	40.9% (32.3%–50.0%)	46.7% (37.6%–56.1%)	50.5% (41.3%–59.6%)
False neg rate	8.9% (6.8%–11.7%)	10.1% (7.2%–13.9%)	9.3% (5.8%–12.4%)	9.4% (7.3%–11.9%)	8.5% (3.4%–19.9%)	10.0% (4.3%–21.4%)	14.3% (7.4%–25.7%)
False pos rate	8.3% (5.3%–12.8%)	8.8% (5.8%–13.2%)	9.0% (6.1%–13.2%)	8.9% (6.7%–11.8%)	4.4% (1.5%–12.2%)	5.3% (1.8%–14.4%)	12.7% (6.3%–24.0%)
% Intermediate	10.5% (8.5%–13.0%)	11.4% (9.0%–14.3%)	12.5% (9.9%–15.5%)	11.9% (10.1%–14.0%)	10.4% (6.1%–17.4%)	8.4% (4.5%–15.2%)	13.5% (8.4%–21.1%)
Accuracy	90.2% (87.7%–92.2%)	89.2% (86.2%–91.7%)	90.0% (87.0%–92.4%)	89.6% (87.5%–91.4%)	93.2% (86.6%–96.7%)	91.8% (84.7%–95.8%)	84.4% (75.8%–90.3%)
Sensitivity	90.3% (87.3%–92.6%)	88.4% (84.1%–91.7%)	90.6% (86.3%–93.6%)	89.5% (86.6%–91.8%)	90.0% (76.9%–96.0%)	89.1% (77.0%–95.3%)	84.0%(71.5%–91.7%)
Specificity	90.0% (84.7%–93.6%)	90.3% (85.5%–93.6%)	89.4% (84.5%–92.8%)	89.8% (86.5%–92.3%)	95.2% (86.9%–98.4%)	94.2% (84.4%–98.0%)	84.8% (71.8%–92.4%)
PPV	96.0% (93.7%–97.4%)	92.5% (88.6%–95.1%)	90.9% (86.7%–93.9%)	91.7% (89.0%–93.8%)	92.3%(79.7%–97.3%)	93.2% (81.8%–97.7%)	85.7% (73.3%–92.9%)
NPV	77.9% (71.8%–83.0%)	85.3% (80.0%–89.4%)	88.9% (84.1%–92.5%)	87.1% (83.6%–89.9%)	93.8% (85.0%–97.5%)	90.7% (80.1%–96.0%)	83.0% (69.9%–91.1%)

(C)
	PET or CSF
LucentAD Complete	Ab positive	Ab negative	Total	Frequency	Predictive value	Likelihood ratio
High risk	475	43	518	47.9% (44.9%–50.9%)	91.7% (89.0%–93.8%)	8.91 (6.58, 12.05)
Intermediate	68	61	129	11.9% (10.1%–14.0%)	52.7% (44.1%–61.1%)	0.90 (0.44, 1.83)
Low risk	56	379	435	40.2% (37.3%–43.2%)	12.9% (10.0%–16.3%)	0.12 (0.09, 0.16)
Total	599	483	1082	Prevalence = 55.4%

Note: Clinical performance characteristics of LucentAD Complete LDT. Performance parameters exclude samples in the intermediate zone (calculation method 2, supplemental section C in supporting information; using an alternative calculation method 3, sensitivity, specificity, and accuracy were 90.7%, 91.1%, and 90.9% respectively). (A) Performance characteristics broken out by training and validation cohorts (validation cohorts separately and combined). (B) Performance characteristics for a subset of ADNI participants with aligned timepoints illustrating performance reproducibility in smaller longitudinal samplings. (C) Likelihood ratios and predictive values.

Abbreviations: Aβ, amyloid beta; ADC, Amsterdam Dementia Cohort; ADNI, Alzheimer's Disease Neuroimaging Initiative; AUC, area under the curve; BH, Bio‐Hermes; CSF, cerebrospinal fluid; NPV, negative predictive value; PET, positron emission tomography; PPV, positive predictive value.

3.5. Performance in non‐AD dementias and subgroups

Demographic characteristics and a detailed analysis of the test's performance in patients with dementia due to other etiologies (vascular dementia [VaD], frontotemporal dementia [FTD], dementia with Lewy bodies [DLB]) and co‐pathology are provided in Section G, Tables S6–S12, and Figure S4 in supporting information. These groups showed amyloid positivity rates of 36.0% (VaD), 19.8% (FTD), and 47.6% (DLB) by CSF, consistent with expected prevalence ranges for these dementia syndromes ²⁶ . Amyloid detection accuracies were 90.6% for VaD, 86.7% for FTD, and 76.3% for DLB. While detection accuracy for AD pathology in VaD and FTD was statistically consistent to the combined validation cohorts, the accuracy for DLB was significantly lower. The percentage of cases in the intermediate zone in DLB patients (25%) was statistically higher than for the other two dementia types (VaD: 11.7%; FTD: 8.7%) and compared to the overall validation cohort (11.9%). Inclusion of these co‐pathology cases into the validation sample set at expected proportions for a memory clinic population ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² did not significantly alter the test's amyloid classification performance. Except for DLB, accuracy of amyloid detection across demographic subgroups ranged from 87% to 95% with the percent of intermediate results ranging from 3.5% to 14% (Figure 4). Although the accuracy estimates for non‐White participants (95%) was numerically higher than for White participants, this difference was not statistically significant. Similarly, age, sex, and APOE ε4 carrier status did not significantly impact test accuracy (Figure 4).

FIGURE 4.

LucentAD Complete accuracy (A) and intermediate zone percentages (B) across datasets and subgroups. Point estimates and 95% confidence intervals are shown for training, validation, and subgroup analyses. Vertical lines represent the full validation set (n = 1082). Sample sizes are in parentheses. Non‐White subgroup: Black, Asian, and Native American participants. Cohort/subgroup Validation 1, first validation cohort; Validation 2, second validation cohort. 24 M, 24 months; 48 M, 48 months; ADC, Amsterdam Dementia Cohort; ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; BH, Bio‐Hermes; BL, baseline; CI, confidence interval; DLB, dementia with Lewy bodies; FTD, frontotemporal dementia; VaD, vascular dementia.

4. DISCUSSION

This multi‐cohort study demonstrates the clinical validity of the LucentAD Complete test, a novel blood‐based assay for amyloid pathology detection. Our findings build upon the established value of p‐tau 217 as a highly accurate single biomarker but also address the persistent challenge of classifying individuals within the p‐tau 217 intermediate zone. The multi‐analyte approach demonstrated an AUC of 0.92% and ≥ 90% accuracy, while reducing the intermediate zone ≈ 3‐fold to 11.9%. This improvement enhances diagnostic confidence and reduces the need for more invasive and costly procedures like CSF analysis or PET scans. While p‐tau 217 alone offers excellent discrimination for many individuals, the presence of an indeterminate zone necessitates further investigation, a challenge addressed by this novel assay.

Our key finding is that the addition of four readily measurable plasma biomarkers (Aβ42/Aβ40, GFAP, and NfL) significantly enhances amyloid classification, specifically within this critical intermediate zone. This improvement can be rationalized by the multifactorial nature of AD, which extends beyond amyloid plaque deposition and p‐tau production ³³ . Variable levels of inflammation and neuronal damage across individuals with AD suggest that a broader assessment of ongoing pathophysiology using biomarkers representing these diverse pathological processes improves the identification of patients exhibiting multiple disease‐related biomarkers ³⁴ . Importantly, this enhancement is achieved without compromising the high accuracy of p‐tau 217 in clearly classified amyloid‐positive or ‐negative samples. This targeted approach, using a multi‐analyte algorithm solely within the p‐tau 217 intermediate range, optimizes the LucentAD Complete test for diagnostic power and with an automated algorithm interpretation; associated complexity of a multi‐analyte test is minimized. By running all biomarkers up front, the test ensures a rapid, comprehensive result with the quickest turnaround time, which we believe outweighs the logistical complexity and potential delays of a reflex‐based strategy. The observed nearly 3‐fold reduction in intermediate cases (34.4% to 11.9%) translates to a greater proportion of patients receiving clear and actionable results, thereby improving clinical workflow and potentially accelerating access to appropriate interventions.

The multi‐analyte algorithm with pre‐established cut‐offs demonstrated robust performance across three independent cohorts, comprising diverse demographics, clinical characteristics, and amyloid positivity prevalence rates. This heterogeneity, including variations in age, geography, race/ethnicity, and clinical presentation, strengthens the generalizability of our findings and suggests that LucentAD Complete can be reliably applied across a broad spectrum of clinical settings. Notably, the test maintained a PPV > 90% across amyloid prevalence rates ranging from 41% to 70%, a critical attribute for clinical utility. The longitudinal analysis within the ADNI cohort provides preliminary evidence for the test's consistency over time, an important consideration for potentially monitoring disease progression.

It is worth noting that our validation cohort's 34% intermediate zone for p‐tau 217 alone is consistent with other technologies, such as the Lumipulse G p‐tau 217 plasma test (analog chemiluminescent assay), which reported a similar intermediate zone of 34% across comparable cohorts ³⁵ . However, the addition of Aβ42 to the Lumipulse p‐tau 217 test, as a p‐tau 217/Aβ42 ratio, reduced the intermediate zone to 20% within the same sample set. This aligns with our own findings, in which combining the Aβ42/40 ratio with p‐tau 217 in a logistic regression model reduced the intermediate zone from 31.2% to 14.9%. But beyond adding biomarkers, the size of the intermediate zone is also heavily influenced by the cohorts being tested and specific cutoffs used. Palmqvist et al. highlighted this variability, showing that the intermediate zones for the FDA‐cleared Lumipulse p‐tau 217/Aβ42 ratio test varied from 19% to 33% (mean 27%) across four different cohorts. They also demonstrated that somewhat less stringent cutoff placement could dramatically reduce the intermediate cases to 6% to 10%, though at the cost of overall accuracy ³⁶ . This underscores the critical balance between reducing indeterminate results and maintaining diagnostic precision.

Regarding the use of Aβ42/40 ratio in the LucentAD Complete test versus Aβ42 alone or a p‐tau 217/Aβ42 ratio, it is well established that amyloid 1‐40 and 1‐42 peptides are sticky and unstable proteins leading to well‐documented pre‐analytical risk; however, the ratio of the two peptides normalizes this pre‐analytical variation, giving a much more stable biomarker ³⁷ , ³⁸ . The advantage of the Aβ42/40 ratio has been recently affirmed by Figdore et al., ³⁹ who concluded that the disparate stability profiles of Aβ42 and p‐tau217 can lead to altered p‐tau 217/Aβ42 ratio results if samples are not handled properly. This poses a risk of patient misclassifications in routine clinical practice ⁴⁰ , ⁴¹ . For these reasons, we did not explore the p‐tau 217/Aβ42 ratio as part of the LucentAD Complete algorithm.

A potential concern in developing a blood‐based AD biomarker algorithm is the confounding effect of age on certain markers, such as NfL and GFAP, which are known to rise with normal aging. However, our algorithm's high accuracy is primarily driven by the highly specific signal from p‐tau217, with NfL, GFAP, and amyloid ratio serving as secondary augmentative signals in borderline cases. During algorithm development, we found that using uncorrected values yielded equivalent clinical performance—in terms of AUC, sensitivity, and specificity—to a model using age‐corrected data. This suggests that the pathological signal from NfL and GFAP in the context of AD is sufficiently strong to overcome their age‐related “noise” and provide a useful secondary diagnostic signal when the p‐tau217 signal is more equivocal.

Several limitations should be acknowledged. First, the LucentAD Complete test exhibited reduced accuracy in the DLB patient group, with diminished sensitivity and a higher proportion of cases falling within the intermediate zone. This suggests that while a positive result strongly indicates amyloid co‐pathology in DLB, a negative result may not reliably rule it out. This finding should be considered when interpreting test results in patients with suspected DLB. Additionally, the use of different reference methods (CSF biomarkers in ADC, visual PET in Bio‐Hermes and ADNI) could contribute to some of the observed differences in test performance across cohorts. Quantitative PET demonstrates higher sensitivity and consistency than visual PET ⁴² , and CSF biomarkers exhibit superior clinical performance compared to amyloid PET, with a 22% discordance rate predominantly CSF+/PET– ⁴³ . Last, this study included only cognitively impaired individuals, and further work is needed to evaluate the test in asymptomatic populations, especially with emerging pre‐clinical AD treatments ⁴⁴ .

While this study focused on evaluating the diagnostic performance of LucentAD Complete for amyloid pathology, future research should investigate its potential utility in other clinical contexts. Investigating its potential in predicting conversion from MCI to AD or in identifying individuals most likely to benefit from specific therapies would be valuable. Future research should also investigate the potential of emerging biomarkers, such as brain‐derived tau, either alone or as part of multi‐analyte panels, to potentially reduce the diagnostic intermediate zone. The specificity of these brain‐derived biomarkers in differentiating AD from conditions with peripheral p‐tau changes, like amyotrophic lateral sclerosis, could lead to a reduction in uncertain results. For the LucentAD Complete test, the availability of results from all four biomarkers, particularly NfL, also opens the possibility for extending the test's application to help guide differential diagnosis ¹⁸ , ⁴⁵ . Finally, clinical utility and cost‐effectiveness analyses will be essential to fully understand the impact of LucentAD Complete on clinical practice and health‐care resource use. These studies are under way, including the multi‐site Accurate Diagnosis study ⁴⁶ .

In summary, LucentAD Complete marks a significant advancement in blood‐based amyloid pathology detection. By integrating p‐tau 217 with a panel of complementary biomarkers and using a targeted algorithmic approach, this test offers high accuracy and a substantially reduced intermediate zone, facilitating more efficient and accessible AD diagnosis. The consistent and robust performance observed across diverse cohorts underscores its potential clinical utility as a valuable tool for aiding in the evaluation of individuals with suspected AD.

CONFLICT OF INTEREST STATEMENT

Charlotte E. Teunissen performed contract research for Acumen, ADx Neurosciences, AC‐Immune, Alamar, Aribio, Axon Neurosciences, Beckman‐Coulter, BioConnect, Bioorchestra, Brainstorm Therapeutics, Celgene, Cognition Therapeutics, EIP Pharma, Eisai, Eli Lilly, Fujirebio, Grifols, Instant Nano Biosensors, Merck, Novo Nordisk, Olink, PeopleBio, Quanterix, Roche, Toyama, and Vivoryon. Charlotte E. Teunissen is editor in chief of Alzheimer Research and Therapy, and serves on editorial boards of Medidact Neurologie/Springer, and Neurology: Neuroimmunology & Neuroinflammation. Inge Verberk received a speaker honorarium from Quanterix, which was paid directly to her institution. David Wilson, Meenakshi Khare, Michele Wolfe, Patrick Sheehy, Ann‐Jeanette Vasko, and Mike Miller are employees of Quanterix. Karen Copeland and Lyndal Hesterberg are contractors of Quanterix. Author disclosures are available in the supporting information.

ETHICS STATEMENT

All studies were properly consented to and approved by applicable ethics committees. All human subjects provided informed consent. The studies were conducted in accordance with local legislation and institutional requirements.

Supporting information

Supporting Information

Supporting Information

Wilson DH, Copeland K, Miller M, et al. Clinical performance of scalable automated p‐tau 217 multi‐analyte algorithmic blood test with reduced intermediate zone using multiplexed digital immunoassay. Alzheimer's Dement. 2025;17:e70215. 10.1002/dad2.70215