Two-stage blood algorithm reduces amyloid PET use in validated Alzheimer’s triage study

A two-stage blood-biomarker algorithm called GRAD classified most participants without an amyloid positron emission tomography scan and narrowed the group whose results were too uncertain for a firm classification. In external validation involving 1,644 cognitively unimpaired participants from the Anti-Amyloid Treatment in Asymptomatic Alzheimer’s disease study, GRAD increased correct classifications within the p-Tau217 gray zone from 41.4% to 56.2% while reducing the proportion referred for PET from 35.9% to 24.3%. The findings support the algorithm as a triage tool rather than a replacement for confirmatory imaging.

Phosphorylated tau 217, commonly called p-Tau217, is a blood marker used to identify cerebral amyloid-beta pathology associated with Alzheimer’s disease. The study notes that PET has cost and accessibility limitations, while p-Tau217 can be more accessible and less costly. However, p-Tau217 values overlap between people with and without brain amyloid, particularly in early stages. Researchers therefore define a gray zone in which the marker alone cannot confidently distinguish the two groups. In this study, p-Tau217 left 33.8% of 145 cognitively impaired development-cohort participants and 38.3% of the external validation cohort in this uncertain range.

GRAD, short for the Gatekeeper-Reflex model for Alzheimer’s disease, concentrates additional testing on uncertain results rather than sending every person for imaging. Its first stage, the Gatekeeper, is a logistic-regression model based solely on harmonized p-Tau217. A predicted probability below 0.25 indicates high-confidence amyloid negativity, while a probability above 0.75 indicates high-confidence positivity. Results from 0.25 through 0.75 advance to the second stage. The Gatekeeper resolved 96 of 145 development participants and 1,014 of 1,644 validation participants. Accuracy among these resolved cases was 93.8% in the development cohort and 86.2% in validation.

The Reflex stage is an L2-regularized logistic regression combining p-Tau217 with glial fibrillary acidic protein, the amyloid-beta 42-to-40 ratio, a tau-amyloid divergence measure, age, APOE epsilon-4 carrier status and a selected interaction between GFAP and p-Tau217. Missing biomarker values were imputed with training-set medians rather than information drawn from the test cohort. The investigators established regularization and decision boundaries before external validation and used leave-one-out cross-validation for internal estimates. They selected interaction terms from biological reasoning rather than automated feature discovery to limit dependence on the development sample. Stage 2 returns an estimated probability of amyloid positivity; values outside 0.40–0.60 are treated as resolved, while remaining uncertainty triggers referral for Aβ-PET.

Among the 49 development-cohort participants who reached the gray zone, the Reflex model achieved an area under the receiver operating characteristic curve of 0.781, with 75.5% accuracy, 76.9% sensitivity and 73.9% specificity. This compared with an AUC of 0.739 for p-Tau217 alone in the same subgroup, a 0.042 difference that was not statistically significant in the small sample. Stage 2 resolved 33 of the 49 cases and referred 16 for PET. In the external gray zone of 630 participants, Stage 2 reached an AUC of 0.793, 70.8% accuracy, 66.7% sensitivity and 80.0% specificity. Its improvement over p-Tau217 alone was 0.073 AUC points, from 0.719, with p less than 0.001. It resolved 477 participants, or 75.7% of the gray-zone group, and referred 153 for imaging.

The model developed in the Alzheimer’s Disease Neuroimaging Initiative was applied without refitting or threshold adjustment to the A4 and LEARN cohorts, despite differences in cognition, amyloid prevalence and p-Tau217 assay. Across the full external pipeline, GRAD achieved an AUC of 0.879, 76.7% sensitivity and 88.6% specificity, and 90.7% of participants received a classification without imaging. Its probabilities also tracked continuous amyloid burden: the correlation with PET Centiloid values was 0.748, with p less than 0.001. Calibration was less satisfactory, however. Predicted risk averaged 57.5% in a cohort with 69.6% observed amyloid positivity, and the calibration intercept was 0.93. The findings therefore indicate that discrimination and routing transferred better than the numerical probabilities, which may require recalibration in populations with different prevalence or assay characteristics.

APOE epsilon-4 carrier status, p-Tau217 and the amyloid-beta 42-to-40 ratio contributed most strongly to resolving the gray zone. The estimated odds ratio per standard deviation was 3.14 for APOE epsilon-4, with a 95% confidence interval of 1.53–6.44; 2.34 for p-Tau217, with a 95% confidence interval of 1.20–4.56; and 0.32 for Aβ42/40, with a 95% confidence interval of 0.13–0.77. GFAP and its interaction with p-Tau217 added little discrimination, and their removal did not significantly degrade performance in external ablation analyses. The researchers also compared GRAD with a published p-Tau181 integration strategy in 256 validation participants with matched measurements. GRAD achieved an AUC of 0.799 in this subset, compared with 0.597 for p-Tau181 and 0.743 for p-Tau217 alone; the difference from p-Tau181 was significant, while the difference from p-Tau217 was not.

A simulated economic analysis modeled 10,000 patients using 2024 US Medicare reimbursement assumptions: $350 for a p-Tau217 test, $250 for the multianalyte reflex panel and $3,000 for a PET scan. Universal PET was projected to cost $30 million. Screening with p-Tau217 followed by PET for every gray-zone case was projected to cost $15 million, while GRAD was projected to cost $7.25 million under external-cohort routing rates. Development-cohort routing produced a GRAD estimate of $7.655 million. These scenarios indicated savings of 74%–76% against universal PET and 44%–52% against p-Tau217 screening with PET confirmation, while reducing projected imaging by 89%–91%. Cost per correct diagnosis was $863 for GRAD under external routing, compared with $1,639 for p-Tau217 screening and $3,000 for universal PET. The simulation did not use real-world financial data and excluded the downstream consequences and costs of misclassification.

Several limitations constrain the findings. The development cohort was small, and only 49 participants entered the gray zone; the significant incremental performance came from the larger external cohort. Both cohorts were predominantly White and highly educated, with White participants representing 89% of ADNI and 94% of A4/LEARN, so performance in racially and ethnically diverse populations is unknown. The study did not account for renal function, comorbid proteinopathies or mixed pathologies. Aβ42/40 was available for 919 of 1,644 validation participants, and missing values required imputation, although a complete-panel sensitivity analysis without imputation produced consistent results. The cohorts also differed in cognitive status, sex distribution, amyloid prevalence, assay platform and PET positivity definition.

Because the study was cross-sectional, it did not test prediction of future amyloid accumulation, cognitive decline, treatment monitoring or repeated measurements. The authors also note that prospective validation would be needed in clinically representative populations using locally applicable assays and thresholds before fixed probability cutoffs could guide individual decisions. The development cohort produced operating points consistent with cited triage benchmarks: a rule-out threshold had 91.8% sensitivity and 81.7% specificity, while a rule-in threshold had 90.0% specificity and 81.2% sensitivity. Those thresholds were derived in the development cohort, and the external Reflex stage in isolation had 66.7% sensitivity and 80.0% specificity, below the cited benchmark for replacing Aβ-PET.

The most defensible conclusion is that selective reflex testing can make blood-based amyloid assessment more decisive while preserving PET for residual uncertainty. GRAD does not classify every patient from blood alone: across the full pipeline, 89.0% of the development cohort and 90.7% of the external cohort received a classification without imaging, while the remaining uncertain cases were directed toward confirmatory PET. The study identifies a potential strategy for reducing unnecessary imaging and supporting trial enrichment, but not a standalone substitute for confirmatory testing. Prospective studies in broader and more clinically heterogeneous populations, together with assessment of setting-specific calibration, access to the multianalyte panel and real-world economic outcomes, are needed to establish whether the projected reductions in PET use and testing costs occur in practice.

Subject of Research: External validation of a two-stage plasma biomarker algorithm for amyloid-PET triage in Alzheimer’s disease

Article Title: A Two‐Stage Plasma Biomarker Algorithm Reduces Amyloid PET Use and the p‐Tau217 Gray Zone

Article References: Parankusham, H., Vanderlip, C., Birkenbihl, C., Krishna, E., Ugboaja, C., Budson, A. E., Frank, B., & for the Alzheimer's Disease Neuroimaging Initiative (2026). A Two‐Stage Plasma Biomarker Algorithm Reduces Amyloid PET Use and the p‐Tau217 Gray Zone. Annals of Clinical and Translational Neurology, Article acn3.70560. https://doi.org/10.1002/acn3.70560

Image Credits: AI Generated

DOI: 10.1002/acn3.70560

Keywords: Alzheimer’s disease, amyloid PET, blood biomarkers, p-Tau217, APOE ε4, Aβ42/40, triage algorithm, external validation, diagnostic economics, plasma testing, Stage, Plasma