Highly accurate blood test for Alzheimer's disease is similar or superior to clinical cerebrospinal fluid tests

Alzheimer's disease is the leading cause of dementia worldwide, with prevalence expected to rise substantially and associated costs already exceeding $1 trillion annually. Accurate detection of AD pathology is crucial for diagnosis and for identifying candidates for recently approved anti-amyloid disease-modifying therapies such as lecanemab. Current gold-standard biomarkers—amyloid and tau PET imaging and CSF assays—are accurate but limited by cost, invasiveness, and accessibility, restricting widespread clinical use. Blood-based biomarkers (BBMs), particularly plasma phosphorylated tau species, have emerged as promising, scalable alternatives. Among these, p-tau217 has shown the highest accuracy for identifying AD pathology and predicting cognitive decline. However, plasma p-tau levels can be influenced by comorbidities like kidney disease; using the ratio of p-tau217 to non-phosphorylated tau (%p-tau217) mitigates some confounding. This study tests whether plasma %p-tau217 is clinically equivalent or superior to FDA-approved CSF assays for detecting Aβ and tau pathology, using PET imaging as the reference, focusing on cognitively impaired individuals who are the target population for anti-amyloid therapies.

Prior work has established that plasma p-tau species (p-tau181, p-tau217, p-tau231) correlate strongly with AD pathology measured by PET and CSF, with p-tau217 often demonstrating the best diagnostic performance and prognostic value for cognitive decline. Blood biomarker tests have begun entering clinical practice but have not been widely endorsed as standalone diagnostics due to limited head-to-head equivalence studies against established CSF/PET modalities. Comorbidities, especially renal dysfunction, can elevate plasma p-tau, which can be mitigated using ratios like %p-tau217. FDA-approved CSF assays (Roche Elecsys p-tau181/Aβ42; Fujirebio Lumipulse Aβ42/40 and p-tau181/Aβ42) show high concordance with Aβ PET (AUCs ~0.96–0.97) and are used clinically, sometimes employing two-cutoff algorithms to optimize PPV/NPV. Visual reads of Aβ PET have high concordance (~95%) with quantitative measures. Collectively, the literature supports BBMs as promising tools but underscores the need for large, rigorous comparisons demonstrating clinical equivalence or superiority to CSF and PET.

Design: Comparative diagnostic accuracy study across two independent cohorts: BioFINDER-2 (Sweden; NCT03174938; enrollment April 2017–June 2022) and Knight ADRC (USA; samples 2013–2020). Participants: Cognitively unimpaired (controls or subjective cognitive decline) and cognitively impaired (MCI or dementia). In BioFINDER-2, MCI defined by performance ≤−1.5 SD in any domain; AD dementia required Aβ positivity by CSF/PET and DSM-5 criteria. Knight ADRC participants had CDR=0 (CU) or CDR>0 with AD-type clinical syndrome (MCI/dementia). Inclusion required both Aβ and tau PET within 2 years of CSF and available plasma. Cohorts: BioFINDER-2 n=1,422; Knight ADRC n=337. Primary outcome: Aβ PET positivity (Centiloids ≥37). Secondary outcomes: Tau PET positivity (SUVR >1.32 in Braak I–IV meta-ROI) and agreement with clinical AD diagnosis. Biomarkers: Plasma %p-tau217 measured via LC-MS/HRMS as p-tau217 divided by non-phosphorylated mid-region tau. Matched CSF assays: BioFINDER-2 used Roche Elecsys automated immunoassays for Aβ42, Aβ40, and p-tau181 (p-tau181/Aβ42 and Aβ42/40 ratios). Knight ADRC used Fujirebio Lumipulse G1200 for Aβ42, Aβ40, p-tau181 (Aβ42/40 and p-tau181/Aβ42 ratios). Imaging: Aβ PET tracers were [18F]flutemetamol (BioFINDER-2) and [18F]florbetapir or [11C]PiB (Knight ADRC), harmonized to Centiloids. Visual reads (FDA-approved) also performed for [18F]flutemetamol in BioFINDER-2. Tau PET tracers were [18F]RO948 (BioFINDER-2) and [18F]flortaucipir (Knight ADRC). Tau SUVR computed in Braak I–IV meta-ROI; positivity >1.32. Statistical analysis: ROC analyses to compute AUCs overall and within cognitively impaired (CI) and cognitively unimpaired (CU) subgroups; DeLong’s test assessed AUC differences. In CI participants, two categorization approaches were used: (1) single cutoff optimizing sensitivity while fixing specificity at 90%, reporting accuracy, PPV, NPV, sensitivity; (2) two cutoffs yielding three categories (negative, intermediate, positive): lower cutoff fixed sensitivity at 95% maximizing specificity; upper cutoff fixed specificity at 95% maximizing sensitivity, reporting accuracy, PPV, NPV among non-intermediate and percentage intermediate. Bootstrapping (n=1,000 resamples) provided mean estimates and 95% CIs; plasma %p-tau217 served as reference for pairwise differences. Sensitivity analyses included out-of-bag estimates, external cutoff derivation (plasma cutoffs from Knight ADRC; CSF cutoffs from UCSF cohort), alternative modeling using p-tau217 with non-phosphorylated tau as covariate, and longitudinal consistency of %p-tau217 (n=40; ~3 years).

Cohorts: BioFINDER-2 (n=1,422; mean age 69.3; 49.8% women; 49.3% CI); Knight ADRC (n=337; mean age 69.8; 51.9% women; 14.8% CI). Aβ PET-positive prevalence: BioFINDER-2 overall 25.8% (CU 15.3%; CI 49.7%); Knight ADRC overall 25.2% (CU 16.7%; CI 74.0%). Tau PET-positive: BioFINDER-2 overall 25.0% (CU 6.8%; CI 43.6%); Knight ADRC overall 10.4% (CU 2.4%; CI 56.0%). AUCs for Aβ PET status: Plasma %p-tau217 showed AUC 0.97 [0.95–0.98] (BioFINDER-2 all), clinically equivalent to CSF Elecsys p-tau181/Aβ42 (0.97 [0.96–0.98]) and CSF Aβ42/40 (0.96 [0.95–0.97]). Knight ADRC: plasma 0.97 [0.95–0.99], equivalent to CSF Lumipulse p-tau181/Aβ42 (0.97 [0.96–0.99]) and Aβ42/40 (0.96 [0.94–0.98]). Subgroup AUCs were similarly high (CU and CI). AUCs for tau PET status: BioFINDER-2: plasma 0.95 [0.93–0.97], superior to CSF p-tau181/Aβ42 (0.93 [0.92–0.95]) and CSF Aβ42/40 (0.88 [0.86–0.90]). Knight ADRC: plasma 0.98 [0.97–0.99], higher than CSF p-tau181/Aβ42 (0.96 [0.94–0.98]) and CSF Aβ42/40 (0.90 [0.87–0.94]). Diagnostic metrics in cognitively impaired (single-cutoff; Aβ PET Centiloids ≥37): BioFINDER-2 CI (n=304): plasma %p-tau217 accuracy 0.90 (95% CI: 0.86–0.93), PPV 0.91 (0.88–0.93), NPV 0.89 (0.81–0.96), clinically equivalent to CSF Elecsys p-tau181/Aβ42 (accuracy 0.91, PPV 0.91, NPV 0.91) and CSF Aβ42/40 (accuracy 0.87, PPV 0.90, NPV 0.85). Knight ADRC CI: plasma accuracy 0.94 (0.72–1.00), PPV 0.99 (0.97–1.00), NPV 0.89 (0.48–1.00), equivalent to CSF assays. Diagnostic metrics in cognitively impaired (single-cutoff; tau PET SUVR >1.32): BioFINDER-2 CI (n=663): plasma accuracy 0.88 (0.85–0.91), PPV 0.88 (0.86–0.90), NPV 0.88 (0.82–0.94), superior to CSF p-tau181/Aβ42 (accuracy 0.82) and CSF Aβ42/40 (accuracy 0.68). Knight ADRC CI showed similar performance between plasma and CSF. Two-cutoffs approach (cognitively impaired): Aβ PET (BioFINDER-2 CI): plasma accuracy 0.95 (0.94–0.97), PPV 0.95 (0.94–0.97), NPV 0.96 (0.94–0.98), with 16% (6–25%) intermediate; clinically equivalent to CSF Elecsys p-tau181/Aβ42 (accuracy 0.95; 17% intermediate) and CSF Aβ42/40 (accuracy 0.94; 25% intermediate). Tau PET (BioFINDER-2 CI): plasma accuracy 0.94 (0.94–0.95), PPV 0.93 (0.92–0.94), NPV 0.95 (0.94–0.96), superior to CSF p-tau181/Aβ42 (accuracy 0.93; 34% intermediate) and CSF Aβ42/40 (accuracy 0.91; 49% intermediate). Individuals with intermediate plasma %p-tau217 had imaging values near diagnostic thresholds, consistent with borderline pathology. Clinical AD diagnosis (BioFINDER-2 CI): plasma %p-tau217 AUC 0.94 (0.92–0.96), equivalent to CSF p-tau181/Aβ42 (0.95) and CSF Aβ42/40 (0.93); single-cutoff accuracy 0.86 (0.82–0.89), PPV 0.89 (0.87–0.91), NPV 0.84 (0.77–0.89); two-cutoffs increased metrics to 93–94% with 24% intermediate. Sensitivity analyses: Out-of-bag and external-cutoff analyses confirmed primary findings; modeling p-tau217 with non-phosphorylated tau as covariate yielded nearly identical results; longitudinal %p-tau217 status was stable (only 1/40 changed status over ~3 years).

Plasma %p-tau217 measured by mass spectrometry demonstrated very high accuracy for detecting Aβ and tau pathology, achieving clinical equivalence to FDA-cleared CSF assays for Aβ PET classification and superiority for tau PET classification. In cognitively impaired patients, who are the intended population for anti-amyloid therapies, plasma %p-tau217 achieved approximately 90% accuracy, PPV, and NPV with a single cutoff and around 95% with a two-cutoffs strategy, matching or exceeding CSF test performance. Given the accessibility of blood collection compared to lumbar puncture or PET imaging, these results indicate that high-performance blood tests can feasibly replace CSF and reduce the need for PET in clinical workflows to determine Aβ status, thereby improving diagnostic access and enabling more equitable deployment of disease-modifying therapies. The strong classification of tau PET status by plasma %p-tau217 suggests added value for determining whether cognitive impairment is due to AD. The two-cutoffs approach provides very high-confidence results in most patients while identifying an intermediate group near imaging thresholds who may benefit from repeat testing or confirmatory PET/CSF, mirroring FDA-approved workflows for CSF assays.

Across two large, well-characterized cohorts, plasma %p-tau217 provided clinically equivalent or superior performance to FDA-approved CSF assays for classifying Aβ and tau PET status, with especially strong results in cognitively impaired individuals relevant for anti-amyloid treatment decisions. Implementing plasma %p-tau217 in clinical practice could substantially decrease reliance on CSF and PET, expand access to accurate AD diagnosis, and streamline therapy eligibility assessments. Future research should evaluate performance in broader, more diverse primary care populations, directly compare mass spectrometry versus immunoassay platforms for p-tau217, and integrate longitudinal and multimodal strategies to manage intermediate results and monitor disease progression.

Key limitations include: relatively small number of cognitively impaired individuals in the Knight ADRC cohort; lack of a sufficiently large cohort with paired antemortem biomarker and postmortem neuropathology for definitive validation; PPV/NPV depend on disease prevalence, which may vary across settings; mass spectrometry assays, while highly accurate, may have higher cost and more limited availability and operational complexity compared to immunoassays; and underrepresentation of minoritized populations, limiting generalizability to diverse clinical settings.