Predicting continuous amyloid PET levels with CSF and plasma brain‐derived p‐tau217

Abstract

BACKGROUND

Staging amyloid beta (Aβ) burden along a continuous spectrum provides clinically relevant information. Conventional phosphorylated tau (p‐tau)217 reliably identifies amyloid positron emission tomography (PET) positivity, but its ability to capture the full Aβ continuum is limited. Recently developed brain‐derived (BD) p‐tau217 assays enrich for central nervous system tau species and may improve specificity.

METHODS

We tested how closely BD p‐tau217 reflected continuous Aβ and how accurately it could classify participants as Aβ positive or negative across different Centiloid (CL) thresholds in 924 participants from the Alzheimer's Disease Neuroimaging Initiative. Conventional and BD cerebrospinal fluid (CSF) and plasma p‐tau217 biomarkers were quantified using the Alamar NULISAseq panel.

RESULTS

Across the full cohort, plasma BD p‐tau217 demonstrated the strongest continuous association with PET CL (R² = 0.63), and the best prediction across the CL spectrum (root mean square error = = 24.5). Prediction was precise in cognitively unimpaired participants at low CL and reasonably accurate in cognitively impaired participants across intermediate‐to‐high CL.

CONCLUSIONS

Plasma BD p‐tau217 best captured the amyloid PET continuum, with BD processing improving discrimination in plasma but not CSF.

Highlights

• Plasma brain‐derived (BD) phosphorylated tau (p‐tau)217 best predicts continuous amyloid positron emission tomography burden.

• BD p‐tau217 outperforms conventional p‐tau217 in plasma and cerebrospinal fluid.

• BD p‐tau217 improves discrimination at low Centiloid thresholds.

• BD p‐tau217 may aid trial enrichment and amyloid beta staging.

1. BACKGROUND

Amyloid beta (Aβ) pathology biomarkers are often classified in a binary manner, typically based on a threshold at which abnormality can be reliably detected. However, quantifying the severity of Aβ accumulation in a continuous manner provides additional clinically relevant information. ¹ , ² , ³ For instance, amyloid positron emission tomography (PET) staging is increasingly used to guide inclusion criteria in clinical trials. ⁴ , ⁵ Continuous values also enable more precise monitoring of disease progression over time, which is particularly valuable in both observational studies and interventional trials. However, PET imaging is costly and not widely available, which limits its use in large‐scale or repeated assessments. Conventional phosphorylated tau (p‐tau) 217, measured in both cerebrospinal fluid (CSF) and plasma, reliably identifies individuals with abnormal Aβ, particularly among cognitively impaired (CI) participants, confirming its value as a biomarker for binary classification. ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ However, limited evidence shows that it captures the full continuum of Aβ accumulation measured by PET Centiloids (CL), ¹² , ¹³ highlighting the need for more precise characterization of p‐tau217 performance across the Aβ continuum, and for biomarkers that may further improve staging across its full spectrum. Moreover, comparing CSF and plasma performance is important to determine whether minimally invasive plasma measures can approximate CSF‐derived signals while maintaining accuracy across disease stages.

Recently, brain‐derived (BD) p‐tau assays have been developed to improve specificity for AD‐related pathology ¹⁴ by enriching for tau species originating from the central nervous system. These assays reduce potential confounding from peripheral sources of tau and, when combined with ultrasensitive proteomic platforms, allow precise measurement of low‐abundance central nervous system (CNS)‐relevant proteins in both plasma and CSF. ¹⁵ , ¹⁶ BD p‐tau217 has shown promise in predicting amyloid PET positivity and may offer a minimally invasive approach to estimate Aβ accumulation. In addition, comparing BD p‐tau217 to conventional p‐tau217 enables assessment of whether CNS‐enriched signal improves prediction performance.

In this study, we evaluated the ability of CSF and plasma BD p‐tau217 to reflect Aβ accumulation across the full continuum of PET CL values.

2. METHODS

2.1. Study population

In this cross‐sectional study, we included 924 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. ADNI is a multicenter study launched in 2004, aimed at developing clinical, imaging, genetic, and fluid biomarkers for the detection and tracking of Alzheimer's disease (AD). Exclusion criteria included a recent or current cancer diagnosis, severe mental illness, current or recent alcohol or substance abuse, recent traumatic brain injury with loss of consciousness, and any contraindications to PET or magnetic resonance imaging. Detailed inclusion and exclusion criteria are available at https://adni.loni.usc.edu/data‐samples/. Our study population included cognitively unimpaired (CU) individuals, as well as participants with mild cognitive impairment (MCI) or dementia, all with available CSF and plasma p‐tau217 measurements. Diagnostic classifications were assigned by ADNI at each visit using standardized clinical criteria, including structured interviews, functional assessments, and neuropsychological testing. All participants provided written informed consent, and ADNI protocols were approved by the institutional review boards at each participating site.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature (PubMed and Google Scholar) using traditional sources. Amyloid beta (Aβ) biomarkers are usually interpreted using binary thresholds. However, continuous staging of Aβ provides clinically meaningful information for trial enrichment and disease monitoring. Conventional phosphorylated tau (p‐tau)217 in cerebrospinal fluid and plasma reliably predicts binary amyloid positron emission tomography (PET) positivity, particularly in cognitively impaired individuals, but captures only part of the full Aβ spectrum. Brain‐derived (BD) p‐tau217 assays enrich for central nervous system tau species and may improve specificity across the Aβ spectrum. These papers are cited in the article.

2. Interpretation: Plasma BD p‐tau217 best captured continuous amyloid PET burden and showed superior discrimination for amyloid PET positivity, especially at low Centiloid thresholds and among cognitively impaired participants. Prediction was most precise in cognitively unimpaired individuals. BD p‐tau217 modestly improves staging over conventional p‐tau217, with greatest gains in ranges relevant for early intervention and trial enrollment.

3. Future directions: Longitudinal studies should evaluate BD p‐tau217 for tracking within‐person changes and its integration with accessible biomarkers for scalable Aβ staging.

2.2. Fluid biomarkers

Plasma and CSF samples were analyzed using the Alamar NULISAseq CNS Disease Panel at the Michael T. Zuendel Family Laboratory, Banner Sun Health Research Institute. ¹⁷ Before analysis, samples were thawed and centrifuged at 2200 × g for 10 minutes to remove particulate matter. A total of 25 µL of supernatant from each sample was transferred to 96‐well plates for assay. The NULISAseq CNS Disease Panel (Alamar Biosciences), ¹⁵ a panel encompassing 131 analytes, was used to quantify protein levels. Protein‐specific immunocomplexes were formed using paired oligonucleotide‐conjugated antibodies. Subsequent steps, including oligo‐dT bead capture, release, secondary capture with streptavidin beads, and ligation, were performed on the automated Alamar ARGO HT system following the manufacturer's instructions. Reporter DNA libraries containing target‐specific molecular identifiers (TMIs) and sample‐specific molecular identifiers (SMIs) were pooled, amplified by polymerase chain reaction, purified, and sequenced on the Illumina Novaseq X platform. Assay quality was monitored using pooled plasma samples, internal control reporters, inter‐plate controls, and negative controls. Raw target counts were first normalized to internal controls and then to inter‐plate control medians. Values were subsequently rescaled, log₂‐transformed, and reported as NULISA Protein Quantification (NPQ) units, reflecting relative protein abundance and log₂ fold changes between samples. All biomarker concentrations in this study are presented as NPQ values.

2.3. Imaging methods

Cerebral Aβ burden was measured using PET scans acquired through the ADNI protocol, ¹⁸ with either [^{1 8}F]florbetapir or [^{1 8}F]florbetaben tracers. Image acquisition followed standardized ADNI procedures to ensure consistency across study sites and tracer types. All PET images were centrally preprocessed by ADNI using validated pipelines, which included realignment of frames, averaging, spatial normalization to a common template, and intensity normalization. Global cortical Aβ burden was quantified using standardized uptake value ratios (SUVRs) averaged across predefined cortical regions typically affected by Aβ pathology, including frontal, temporal, parietal, and cingulate cortices. SUVR values normalized to the whole cerebellum obtained with 8 mm smoothing were converted to 6 mm equivalents using ADNI‐derived conversion equations, which have been shown to strongly correlate with native 6 mm SUVRs (R ² ≥ 0.96). ¹⁹ To allow comparison across tracers, SUVRs were transformed into CL units using established conversion procedures. ²⁰

2.4. Statistical analyses

All statistical analyses were conducted in R v.2024.04.2. First, to ensure robustness of the NPQ measurements, we performed COMBAT harmonization and found that corrected and uncorrected values were highly concordant (r = 0.962–0.982) and yielded comparable associations with amyloid PET (Figure S1 in supporting information), supporting the use of the original data in primary analyses.

Analyses focused on evaluating the ability of conventional and BD p‐tau217 in both CSF and plasma to predict continuous Aβ burden as measured by PET CL, as well as binary amyloid PET positivity across multiple thresholds. All analyses were completed in (1) the entire sample, (2) in CU individuals and (3) in CI individuals. All models were adjusted for age and sex, given their established associations with AD pathology. ²¹ , ²²

We analyzed the ability of each biomarker to reflect amyloid PET burden using two different perspectives. First, gradient boosting machine (GBM) models were implemented using the R package gbm with a Gaussian loss function. Models included up to 1000 trees with an interaction depth of 3, shrinkage (learning rate) of 0.01, and a minimum of 10 observations per terminal node. These hyperparameters were chosen based on values used in prior biomarker PET modeling studies and were held constant across analyses to ensure comparability. ¹² Five‐fold cross‐validation was used within the GBM framework to monitor model performance and reduce overfitting during model fitting. No additional hyperparameter grid search was performed. GBM predictions were used to generate centered individual conditional expectation (ICE) curves for each participant, which show how the predicted PET CL changes across the observed biomarker values for that individual while adjusting for age and sex. For each participant, the centered prediction profile was obtained by subtracting the predicted CL value at the lowest biomarker level from all predicted values. Second, predicted generalized additive model (GAM) curves, adjusting for age and sex, were plotted with shaded 95% confidence intervals to represent uncertainty in the fitted trajectories. Model calibration was evaluated by fitting linear regressions of predicted versus observed CL, reporting the slope, intercept, and R ², and the 5th to 95th percentile range of predicted CL values was derived to characterize the typical prediction span for individuals. These modeling approaches were selected based on their use in previous studies of biomarker PET relationships, ¹² , ¹³ facilitating comparability across studies, while also providing complementary advantages in flexibility (GBM) and interpretability (GAM). To assess local prediction accuracy across the CL continuum, we computed sliding window root mean square error (RMSE, window whole sample = 85 participants; window CU = 35, window CI = 50), within overlapping windows for participants sorted by observed CL values. This approach allowed identification of CL ranges in which biomarker predictions were most and least accurate.

For binary amyloid PET classification, CL thresholds of 20, 40, 60, and 80 were applied to define amyloid PET positivity. Raw biomarker values were used to construct receiver operating characteristic (ROC) curves for each biomarker and threshold, and the area under the curve (AUC) was computed to assess how well a biomarker distinguished participants above versus below that threshold. Comparisons between conventional and BD p‐tau217 biomarkers were performed using a DeLong test to assess whether the BD assays improved discrimination. P values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure across the comparisons. In sensitivity analyses, we repeated the models correcting for apolipoprotein E (APOE) ε4 allelic count, and repeated the analyses using predicted CL values from each biomarker's GAM.

The code used to generate the results reported in this study is available from the corresponding author upon request.

3. RESULTS

3.1. Participant characteristics

A total of 924 participants were included in this study. The mean age was 72.2 ± 7.17 years, and 51.4% were female. Overall, 58.6% of participants were CI (MCI or dementia). PET CL values ranged from −39 to 201, with a mean of 37.9 ± 47.2. Demographic characteristics of the sample are listed in Table 1.

TABLE 1.

Demographic, clinical, and biomarker characteristics of the study sample.

	CU (N = 383)	MCI (N = 433)	Dementia (N = 108)	Overall (N = 924)
Age, mean (SD), years	72.4 (6.35)	71.9 (7.51)	74.1 (8.25)	72.4 (7.18)
Female, no. (%)	241 (62.9)	190 (43.9)	43 (39.8)	474 (51.3)
Education, mean (SD), years	16.8 (2.40)	16.2 (2.65)	15.9 (2.44)	16.4 (2.55)
Race, no. (%)
Asian	2 (0.5)	1 (0.2)	0 (0)	3 (0.3)
Black	16 (4.2)	8 (1.8)	3 (2.8)	27 (2.9)
White	349 (91.1)	386 (89.1)	95 (88.0)	830 (89.8)
More than one race	5 (1.3)	2 (0.5)	0 (0)	7 (0.8)
Unknown	11 (2.9)	36 (8.3)	10 (9.3)	57 (6.2)
CDR‐SB	0.0445 (0.176)	1.49 (0.920)	4.42 (1.60)	1.23 (1.58)
Centiloids, mean (SD)	20.7 (35.4)	41.5 (48.3)	84.9 (44.7)	37.9 (47.3)
APOE ε4 carriers, no. (%)	128 (33.4)	199 (46.0)	74 (68.5)	401 (43.4)
CL 20, no. (%)
≤20	260 (67.9)	199 (46.0)	13 (12.0)	472 (51.1)
> 20	123 (32.1)	234 (54.0)	95 (88.0)	452 (48.9)
CL 40, no. (%)
≤40	302 (78.9)	241 (55.7)	18 (16.7)	561 (60.7)
> 40	81 (21.1)	192 (44.3)	90 (83.3)	363 (39.3)
CL 60, no. (%)
≤60	325 (84.9)	279 (64.4)	25 (23.1)	629 (68.1)
> 60	58 (15.1)	154 (35.6)	83 (76.9)	295 (31.9)
CL 80, no. (%)
≤80	348 (90.9)	328 (75.8)	41 (38.0)	717 (77.6)
> 80	35 (9.1)	105 (24.2)	67 (62.0)	207 (22.4)

Abbreviations: APOE, apolipoprotein E; CDR‐SB, Clinical Dementia Rating Sum of Boxes; CL, Centiloid; CU, cognitively unimpaired; MCI, mild cognitive impairment; SD, standard deviation.

3.2. Continuous associations between fluid biomarkers and PET CL

To examine the continuous relationship between fluid p‐tau217 and amyloid PET CL, we first generated ICE plots for each biomarker using predictions from GBM models, adjusting for age and sex (Figure 1). ICE plots revealed that plasma BD p‐tau217 closely tracked PET CL values, with tight, consistent increases across participants, whereas conventional plasma p‐tau217 showed more heterogeneous trajectories, despite an overall positive trend. In CSF, both BD and conventional p‐tau217 exhibited greater inter‐individual variability and less consistent alignment with PET CL. Similar results were observed in CI individuals. In CU participants, plasma p‐tau217 and BD p‐tau217 showed a tight, consistent increase in predicted values up to ≈ 80 CL, after which the trajectories dropped and plateaued. Individual trajectories were heterogeneous beyond this point. In contrast, CSF biomarkers exhibited highly variable and inconsistent trajectories across the CL range. Sensitivity analyses correcting for APOE ε4 allelic count yielded similar results (Figure S2 in supporting information).

FIGURE 1.

ICE profiles for CSF and plasma biomarkers. The relationship between biomarkers and predicted PET CL levels is illustrated using ICE profiles for both conventional p‐tau217 and BD p‐tau217 in CSF and plasma. Each gray line represents the model‐predicted trajectory for a single participant across the biomarker range, adjusted for age and sex. The bold colored line indicates the average predicted relationship across all participants for each biomarker. This visualization highlights both individual variability and the overall trend in how biomarker levels relate to predicted amyloid beta burden. BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; ICE, individual conditional expectation; PET, positron emission tomography; p‐tau, phosphorylated tau.

Predicted CL values were positively associated with observed values for all biomarkers, as indicated by linear calibration analyses including slope, intercept, and R ² (Table 2 and Figure 2). These metrics provide a global summary of agreement between predicted and observed CL values across the full PET spectrum, whereas Figure 2 illustrates local non‐linear patterns across the CL continuum. The slope of the fitted models ranged from 0.37 for CSF and CSF BD p‐tau217 to 0.63 for plasma BD p‐tau217, with intercepts spanning 14.1 to 23.8. Corresponding R ² values indicated moderate‐to‐strong continuous associations, with the highest for plasma BD p‐tau217 (R² = 0.63) and the lowest for CSF biomarkers (R² = 0.37). The predicted CL ranges captured most of the observed PET values, with plasma biomarkers extending slightly beyond the extremes of the measured CL spectrum. The summary of sliding‐window RMSE provided an additional global measure of prediction accuracy, with lower average RMSE observed for plasma BD p‐tau217 (24.5) compared to conventional plasma (27.0) and CSF biomarkers (31.8), reflecting tighter, more consistent continuous prediction across the cohort. In other words, an RMSE of 24.5 means a participant predicted at 75 CL based on the biomarker may be anywhere between 50 and 100 CL. Sensitivity analyses correcting for APOE ε4 allelic count yielded similar results (Table S1 in supporting information).

TABLE 2.

Calibration metrics for GAM‐predicted PET CL values across biomarkers.

All	Slope	Intercept	R 2	Range	Summary RMSE
CSF p‐tau217	0.37	23.8	0.37	2 to 82	31.8
CSF BD p‐tau217	0.37	23.8	0.37	2 to 83	31.8
Plasma p‐tau217	0.56	16.6	0.56	−5 to 94	27.0
Plasma BD p‐tau217	0.63	14.1	0.63	−1 to 100	24.5
Cognitively unimpaired
CSF p‐tau217	0.24	15.6	0.25	−2 to 50	23.4
CSF BD p‐tau217	0.24	15.7	0.24	−2 to 50	23.3
Plasma p‐tau217	0.45	11.4	0.45	−7 to 68	20.8
Plasma BD p‐tau217	0.54	9.4	0.55	−3 to 79	18.9
Cognitively impaired
CSF p‐tau217	0.39	30.7	0.39	6 to 90	34.0
CSF BD p‐tau217	0.39	30.4	0.40	6 to 91	34.0
Plasma p‐tau217	0.57	21.5	0.57	−4 to 100	29.1
Plasma BD p‐tau217	0.62	19	0.62	−5 to 102	27.1

Notes: The slope and intercept of linear regressions of predicted versus observed CL are shown, reflecting the agreement between model predictions and measured values. Slopes closer to 1 indicate stronger proportional agreement, while intercepts indicate systematic over‐ or underestimation. The 5th to 95th percentile range of predicted CL values is also included to illustrate the typical prediction span for each biomarker. Models were adjusted for age and sex. Summary RMSE summarizes local prediction accuracy across the CL continuum.

Abbreviations: BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; GAM, generalized additive model; PET, positron emission tomography; p‐tau, phosphorylated tau; RMSE, root mean square error.

FIGURE 2.

Predicted versus observed PET CL levels from GAM models for CSF and plasma biomarkers. Observed PET CL values are plotted against predicted values generated by GAMs for each biomarker, including conventional p‐tau217 and BD p‐tau217 in CSF and plasma. Models were adjusted for age and sex. Each point represents an individual subject. Smooth colored lines depict the GAM‐predicted relationship, with shading showing the 95% confidence interval. BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; GAM, generalized additive model; PET, positron emission tomography; p‐tau, phosphorylated tau.

We repeated the analyses in clinical subgroups. In CU participants, slopes were lower across all biomarkers (0.24–0.54), intercepts were generally lower than in the full cohort (9.4–15.7), and R ² values were moderate (0.24–0.55), with plasma BD p‐tau217 showing the strongest association and lowest RMSE (18.9). Predicted CL ranges for CU participants were narrower, particularly for CSF biomarkers (–2 to 50 CL), whereas plasma biomarkers covered a wider span (–7 to 79 CL). In CI participants, slopes and intercepts were higher, reflecting larger predicted CL values, with R ² values similar to the overall cohort (0.39–0.62). Plasma BD p‐tau217 again exhibited the strongest association (R² = 0.62) and lowest RMSE (27.1), whereas CSF biomarkers showed higher RMSE (34.0), indicating more variable predictions. Overall, plasma BD p‐tau217 provided the most robust and consistent predictions across both CU and CI subgroups. Sensitivity analyses correcting for APOE ε4 allelic count yielded similar results (Table S1).

3.3. Discrimination and prediction performance across PET CL thresholds

To assess biomarker performance in classifying amyloid PET positivity at multiple PET CL thresholds (20, 40, 60, 80 CL), we computed ROC curves and corresponding AUC values for each biomarker (Figures 3 and 4). Across all participants, plasma BD p‐tau217 consistently exhibited the highest discrimination (AUC range across thresholds: 90.0–93.2), followed by conventional plasma p‐tau217 (AUC range across thresholds: 88.0–91.4) and CSF biomarkers (CSF BD p‐tau217: 80.2–83.3; conventional CSF p‐tau217: 80.1–83.4). Comparisons using DeLong tests (Table 3) indicated that plasma BD significantly outperformed conventional plasma at all thresholds (P < 0.001), whereas CSF BD and conventional CSF showed no significant differences (P > 0.05). Sensitivity analyses using the GAM‐derived values yielded similar results (Table S2 in supporting information). Sensitivity analyses correcting for APOE ε4 allelic count also yielded similar results (Table S3 in supporting information).

FIGURE 3.

Diagnostic performance of plasma and CSF p‐tau217 measures across CL thresholds. Heatmap displaying AUC for the conventional and BD biomarkers in both CSF and plasma across PET CL thresholds. Color intensity reflects AUC magnitude, with darker shades indicating higher discriminative performance. AUC, area under the curve; BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; PET, positron emission tomography; p‐tau, phosphorylated tau.

FIGURE 4.

Biomarker performance across CL thresholds. ROC curves for predicting amyloid PET positivity at different PET CL thresholds. Curves show the ability of raw biomarkers values to identify amyloid‐positive participants at CL thresholds of 20, 40, 60, and 80. BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; PET, positron emission tomography; p‐tau, phosphorylated tau; ROC, receiver operating characteristic.

TABLE 3.

Comparison of conventional and brain‐derived p‐tau biomarkers.

Conventional	AUC conventional	AUC brain‐derived	Absolute difference in AUC	CL threshold	p value (FDR adjusted)
All
CSF	83.2	82.9	0.3	20	0.120
Plasma	89.9	92.1	2.2	20	0.000221
CSF	83.3	83.3	0	40	0.803
Plasma	91.4	93.2	1.8	40	0.000221
CSF	81.9	82.0	0.1	60	0.776
Plasma	90	91.6	1.7	60	0.000515
CSF	80.1	80.2	0.2	80	0.465
Plasma	88	90.0	2	80	9.27e‐06
Cognitively unimpaired
CSF	74.7	74.3	0.4	20	0.274
Plasma	81.7	84.7	3	20	0.0876
CSF	77.4	77	0.4	40	0.274
Plasma	87.6	90.3	2.7	40	0.045
CSF	74.9	74.7	0.3	60	0.515
Plasma	87.4	89.8	2.4	60	0.103
CSF	80.1	79.5	0.7	80	0.260
Plasma	88.2	90.2	2	80	0.180
Cognitively impaired
CSF	86.4	86.3	0.1	20	0.547
Plasma	93.1	95.0	1.9	20	0.00284
CSF	84.7	84.9	0.2	40	0.373
Plasma	91.4	93.1	1.7	40	0.00284
CSF	82.4	82.6	0.2	60	0.461
Plasma	88.7	90.4	1.7	60	0.00284
CSF	77	77.5	0.4	80	0.248
Plasma	85.3	87.5	2.2	80	1.33e‐04

Notes: Pairwise comparison of diagnostic performance (AUC) between brain‐derived and conventional p‐tau biomarkers in CSF and plasma using the DeLong test. Bold values indicate statistically significant differences after FDR correction (q < 0.05). Analyses were conducted in the total sample, in cognitively unimpaired individuals and in cognitively impaired individuals.

Abbreviations: AUC, area under the curve; CL, Centiloid; CSF, cerebrospinal fluid; FDR, false discovery rate; p‐tau, phosphorylated tau.

Sliding‐window RMSE across the PET spectrum further quantified classification uncertainty for plasma BD p‐tau217: errors were higher in the intermediate range (20–40 CL), reflecting heterogeneity in this transitional zone, but predictions were tighter at very low CL values (< 20 CL), consistent with ROC‐derived discrimination (Figure 5 and Table S4 in supporting information). Sensitivity analyses correcting for APOE ε4 allelic count yielded similar results (Table S5 in supporting information).

FIGURE 5.

Sliding RMSE of predicted PET CL values across the observed CL continuum. For each biomarker, predictions from GAMs corrected for age and sex were compared to observed PET CL values. Sliding RMSE was computed over a window of 85 participants to estimate local prediction error along the CL range. The figure shows how prediction accuracy varies across low, intermediate, and high CL levels, with the table (or inset) reporting the mean RMSE across all windows. The gray shaded area indicates the early amyloid PET range, corresponding to CL values near the threshold for positivity, where amyloid deposition is beginning. BD, brain derived; CL, Centiloid; CSF, cerebrospinal fluid; GAM, generalized additive model; PET, positron emission tomography; p‐tau, phosphorylated tau; RMSE, root mean square error.

When stratifying by cognitive status, CI participants showed strong classification, with plasma BD p‐tau217 AUCs ranging from 87.5 to 95.0 and the best sliding‐window RMSE ≈ 25 CL at high CL thresholds (≥ 60 CL), indicating that individual predictions were typically within ± 25 CL of observed values. Conventional plasma p‐tau217 in CI participants showed AUCs of 85.3 to 93.1, while CSF biomarkers had lower AUCs (CSF BD p‐tau217: 77.5–86.3; CSF p‐tau217: 77.0–86.4). BD processing did not substantially improve CSF performance. Sensitivity analyses correcting for APOE ε4 allelic count yielded similar results.

In CU participants, plasma BD p‐tau217 performed best at low CLs (0–40 CL) with RMSE ≈ 15 to 20 CL. Across all thresholds, AUCs ranged from 84.7 to 90.3 for plasma BD p‐tau217, compared to 81.7 to 88.2 for conventional plasma p‐tau217, and lower AUCs for CSF biomarkers (CSF BD p‐tau217: 80.0–83.3; CSF p‐tau217: 80.1–83.4). Performance decreased at higher CLs (≥ 60 CL; RMSE ≈ 30–40 CL). BD processing had minimal impact on CSF discrimination.

4. DISCUSSION

In this study, we comprehensively evaluated BD and conventional p‐tau217 biomarkers in relation to amyloid PET across complementary dimensions of performance, including continuous association with Aβ burden, prediction accuracy, and discrimination of PET positivity. Overall, plasma BD p‐tau217 provided the best discrimination and lowest prediction error across the CL spectrum. BD p‐tau217 significantly improved performance over conventional p‐tau217 in plasma in discrimination of PET positivity, whereas differences in CSF between conventional and BD measures were small and mostly non‐significant. Across cognitive groups, prediction accuracy varied with Aβ burden: in CU participants, predictions were tightest at lower CL values (< 20), and remained fairly precise in early to intermediate Aβ accumulation (CL 20–40), while in CI participants, predictions were most precise at higher Aβ levels (≥ 60 CL). Taken together, these results suggest that plasma BD p‐tau217 offers slightly improved correspondence with the Aβ continuum relative to conventional plasma p‐tau217 and performs reasonably across cognitive groups, indicating potential utility for early detection and clinical applications.

The ability to accurately identify individuals at early stages of Aβ accumulation has become increasingly critical as disease‐modifying therapies shift toward earlier intervention. ⁴ , ²³ In this context, our findings suggest that plasma BD p‐tau217 may offer modestly improved performance for early‐stage enrichment strategies. Among CU participants, plasma BD p‐tau217 demonstrated slightly higher classification accuracy than its conventional counterpart at lower CL levels (≤ 40), at which distinguishing subtle Aβ elevation from true negativity is essential for preclinical trial enrollment. This advantage may reflect the increased specificity of BD assays for CNS‐derived tau species, which could enhance sensitivity to early pathological changes when signal is still low. However, once Aβ burden enters the elevated range (≥ 40 CL), plasma BD p‐tau217 offers little additional advantage over conventional plasma p‐tau217, likely because classification becomes straightforward at higher Aβ levels, with plasma showing higher discrimination than CSF.

In contrast, among CI participants, prediction performance remained high overall, but variability differed across the Aβ spectrum. In the 20 to 40 CL range, RMSE was highest, reflecting greater biological heterogeneity during transitional stages of Aβ accumulation. This observation aligns with previous studies, which frequently describe this range as a gray zone or intermediate phase. ²⁴ , ²⁵ In this range, plasma BD p‐tau217 slightly outperformed conventional plasma p‐tau217, but RMSE remained relatively high, likely reflecting greater noise in the intermediate range due to most CI participants being either clearly amyloid negative or strongly amyloid positive, with fewer individuals in the intermediate zone; furthermore, some of these intermediate cases may have AD as a co‐pathology contributing to their impairment. ²⁶ In addition, the lower number of CI individuals in the 20 to 40 CL range may also partially explain the lower AUCs observed in our study. In the higher range (≥ 60 CL), RMSE decreased, indicating more consistent continuous predictions, but AUC tended to decrease because plasma values cluster at high CL levels. This apparent divergence between RMSE and AUC at high CL levels likely reflects the different properties captured by these metrics: RMSE evaluates continuous prediction error within a restricted high Aβ range, whereas AUC depends on discrimination around the classification threshold, which becomes more challenging as values begin to plateau. ²⁷ In this connection, the plateau of amyloid PET in later disease and the poorer relationship of Aβ pathology with clinical severity suggest that staging using tau PET may provide greater utility at this stage of the disease. ²⁸ , ²⁹ Finally, across the CL spectrum, plasma biomarkers generally showed higher discrimination than CSF, yet even BD‐specific improvements offered little additional benefit at high CLs.

More broadly, our findings are consistent with the phenomenon of biomarker saturation observed in prior studies of amyloid‐related fluid biomarkers. ¹² , ¹³ , ³⁰ At higher levels of Aβ burden, plasma and CSF p‐tau217 values appeared to plateau, resulting in reduced dynamic range despite persistently elevated biomarker levels. This likely contributed to the reduced discrimination observed at higher CL ranges, particularly for threshold‐based metrics such as AUC, for which separation between groups becomes more challenging once biomarker values cluster near an upper limit. Importantly, this apparent plateau does not necessarily indicate loss of biological relevance, but rather reflects compression of biomarker variability at advanced stages of Aβ accumulation. Similar saturation effects have previously been reported for plasma p‐tau biomarkers and may reflect the non‐linear relationship between soluble tau measures and fibrillar Aβ burden in later disease stages. Furthermore, because fluid biomarkers capture the concentration of specific soluble proteins at a specific point in time, and amyloid PET reflects the aggregation of insoluble protein aggregates over time, a saturation effect is also expected. ³¹ In this context, fluid biomarkers may be most informative during early and intermediate phases of Aβ accumulation, while PET‐based measures and tau PET staging may provide greater resolution at more advanced stages of disease.

Individuals enrolled in anti‐amyloid trials typically had mean CL values ranging from ≈ 76 to 100. ²³ , ³² , ³³ A biomarker that accurately identifies individuals in this CL range is particularly valuable, enabling clinicians to achieve the highest accuracy in identifying candidates for disease‐modifying anti‐amyloid therapies within this interval. In our study, among CI participants, both conventional and brain‐derived plasma p‐tau217 demonstrated strong classification performance at higher CL ranges. DeLong testing nonetheless showed that BD plasma p‐tau217 provided statistically significant improvements in discrimination versus the conventional assay across thresholds, indicating a modest though consistent gain in accuracy. In this context, BD processing appears to refine plasma performance, likely by improving the signal‐to‐noise ratio through enrichment of CNS‐derived tau and reduction of peripheral contributions, rather than fundamentally changing classification, offering a modest improvement in identifying individuals who fall within the high Aβ range typical of anti‐amyloid trials. In contrast, CSF already reflects a CNS‐enriched compartment, which may limit the additional benefit of BD processing in this biofluid, consistent with the nearly identical performance observed between conventional and BD p‐tau217 measures across all CSF analyses.

Our findings align with a recent study that compared NULISAseq BD‐pTau217 and Simoa ALZpath pTau217 in two independent Chinese cohorts and similarly reported modestly higher AUCs for BD p‐tau217 (0.96–0.98 vs. 0.94–0.97) in classifying Aβ pathology with low (< 10), intermediate (10–30), and high (> 30) CL. ¹⁶ Despite these methodological differences, as well as a difference in the assay used to quantify conventional p‐tau217, the direction and magnitude of improvement were broadly similar. In the present study, AUCs for plasma BD p‐tau217 versus conventional p‐tau217 were 92.5% versus 90.0% at CL ≥ 20 and 93.4% versus 91.8% at CL ≥ 40, confirming a consistent but incremental BD advantage across populations, tracers, and conventional assay platforms.  Notably, the BD advantage was substantially larger for amyloid/tau–positive classification in that study (AUC 0.89 vs. 0.78), ¹⁶ a comparison not assessed here but one suggesting BD processing may offer greater added value for disease staging.

It is important to note that although plasma BD p‐tau217 predictions were relatively precise, even the lowest RMSE values (≈ 25 CL in CI participants at high Aβ levels) indicate variability at the individual level. In practical terms, this means that for a given participant, the plasma‐based estimate could be off by roughly ± 25 CL compared to the measured PET value. While sufficient for classifying individuals into broad amyloid positive versus negative categories, this level of error limits the ability to accurately predict exact CL values for an individual, highlighting that plasma biomarkers are currently most useful for screening and risk stratification rather than precise quantification of Aβ burden.

Several limitations should be considered. First, our study population consisted of participants willing to enroll in a research study, which may limit the generalizability of our findings to broader, community‐based populations. Second, the BD p‐tau217 assay used in this study was generated on the NULISA research platform, which is not yet a fully automated, clinically routine platform, which impacts scalability and real‐world implementation. Third, we lack longitudinal data with BD p‐tau217, so its utility for tracking Aβ progression over time remains uncertain; biomarker stability and performance for longitudinal staging are unknown. Fourth, the incremental benefit of BD p‐tau217 in our study may be underestimated, as our sample is highly curated and free from many confounding factors. In real‐world clinical practice, conventional plasma p‐tau217 can occasionally yield false positives due to interference from comorbidities, and BD processing is hypothesized to mitigate these effects. Consequently, the practical advantage of BD p‐tau217 in improving diagnostic confidence may be even greater outside of a controlled research setting. Fifth, our analyses relied on a single assay for both CSF and plasma p‐tau217. Performance characteristics may vary across assay platforms, and different assays could potentially show different levels of agreement or divergence relative to BD p‐tau217. Sixth, our findings were derived from a single cohort and therefore require confirmation in independent populations. Replication across cohorts with different demographic and clinical characteristics will be important to establish the robustness and broader applicability of our findings. Finally, our analyses focused exclusively on p‐tau217, and it is unclear how these results would translate to other p‐tau isoforms or combinations of biomarkers.

In conclusion, these findings support the use of plasma BD p‐tau217 as an enrichment strategy in early preclinical trials, but highlight that PET remains essential for accurately quantifying Aβ burden along a continuous scale in later preclinical and symptomatic populations. These results underscore both the promise and the limitations of fluid biomarkers for trial recruitment, emphasizing their role as a complementary tool rather than a replacement for PET‐based quantification.

CONFLICT OF INTEREST STATEMENT

L. Trudel reports no disclosures relevant to the manuscript. J. Therriault has served as a consultant for the Neurotorium Educational Platform and Roche, and as a medical writer for Alzheon Inc., both outside of the scope of the present work. A. C. Macedo, E. Aumont, S. A. Hosseini, A. L. Benedet, and L. N. Hazrati report no disclosures relevant to the manuscript. S. Gauthier serves on scientific advisory boards for Alzeon, AmyriAD, Advantage, Eisai Canada, Enigma USA, Lilly Canada, Medesis, Lundbeck Foundation, Novo‐Nordisk Canada, Okutsa, and TauRx. H. Zetterberg has served on scientific advisory boards and/or as a consultant for Abbvie, Acumen, Alector, Alzinova, ALZpath, Amylyx, Annexon, Apellis, Artery Therapeutics, AZTherapies, Cognito Therapeutics, CogRx, Denali, Eisai, Enigma, LabCorp, Merck Sharp & Dohme, Merry Life, Nervgen, Novo Nordisk, Optoceutics, Passage Bio, Pinteon Therapeutics, Prothena, Quanterix, Red Abbey Labs, reMYND, Roche, Samumed, ScandiBio Therapeutics AB, Siemens Healthineers, Triplet Therapeutics, and Wave; has given lectures sponsored by Alzecure, BioArctic, Biogen, Cellectricon, Fujirebio, LabCorp, Lilly, Novo Nordisk, Oy Medix Biochemica AB, Roche, and WebMD; is a co‐founder of Brain Biomarker Solutions in Gothenburg AB (BBS), which is a part of the GU Ventures Incubator Program; and is a shareholder of CERimmune Therapeutics (outside submitted work). P. Vitali serves on the scientific advisory boards for NovoNordisc, Eisai, and Lilly, and received honoraria from IntelGenx Corp. N. J. Ashton reports no disclosures relevant to the manuscript. P. Rosa‐Neto has served on scientific advisory boards and/or as a consultant for Roche, Novo Nordisk, Eisai, and Cerveau Technologies, outside of the scope of the present work. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All human participants or their caregivers provided informed written consent to participate in the respective studies.

Supporting information

Supporting Information: alz71723‐sup‐0001‐SuppMat.docx

Supporting Information: alz71723‐sup‐0002‐Disclosureforms.pdf