Blood tests for Alzheimer's: what they measure and how accurate they are

Alzheimer's is defined by two proteins that build up in the brain: amyloid, which forms plaques, and tau, which forms tangles. Until recently the only ways to see them in a living person were a PET scan or a sample of spinal fluid taken by lumbar puncture[FDA]. The 2024 revision of the diagnostic criteria defines the disease by that biology rather than by symptoms, and it admits blood tests into the diagnostic categories for the first time[11]. The first accurate blood tests appeared in 2017 and 2018, using amyloid and tau PET as the reference standard[12].

A blood test measures traces of these brain proteins, or of the brain's reaction to them, that leak into the bloodstream. Five markers matter in practice. Two are forms of tau carrying a phosphate at a particular position, p-tau217 and p-tau181; both rise when amyloid plaques are present. One is the ratio of two amyloid fragments, Aβ42 to Aβ40, which falls as plaques form. The other two are not specific to Alzheimer's: GFAP, released by the brain's star-shaped support cells when they react to injury, and NfL, released from damaged nerve fibres whatever the cause[56][45].

The case for blood is access. The FDA's own wording, in clearing the first test, was that PET scans are costly, time-consuming and involve radiation exposure; CSF testing requires a lumbar puncture; a blood draw is less invasive and easier for patients to access[FDA]. What that is worth in money has so far only been modelled, not measured. A US model that assumed a test with 88% sensitivity, 89% specificity put its value-based price at $290 as a triage test, $1,150 as a confirmatory test in primary care, where PET or CSF use would fall by 47% (triage) and 86% (confirmatory); in specialty care the figures were $450 as a triage test, $1,950 as a confirmatory test[13]. A second model found that adding p-tau217 to a standard evaluation saved $342 per patient in diagnostic costs, at $17,571 per quality-adjusted life year gained[14]. A 2026 systematic review of economic evaluations found nine, all model-based: six of CSF tests and three of blood tests; CSF tests were mostly cost-effective; blood tests showed mixed results[15]. In 348 memory-clinic patients with MCI (BioFINDER-1 and 2), a workflow that ran the blood model first and sent only uncertain cases for CSF testing reached an accuracy of 88.2%, 90.5% and 92.0% with lenient, moderate and stringent thresholds while avoiding 85.9%, 72.7% and 61.2% of the CSF tests[16]; the automated-platform study put the avoidable CSF tests at up to 60%[6]. No record in the library states the list price of a commercial blood test, so this page does not give one.

What it reflects

Tau carrying a phosphate at position 217 rises in blood when amyloid plaques are present, and rises further as tangles spread. In the 2020 study that first showed its value, plasma p-tau217 separated autopsy-confirmed Alzheimer's from other brain diseases with an AUC of 0.89 (95% CI 0.81 to 0.97), against 0.50 to 0.72 for p-tau181 and NfL in the same people; it correlated with tangles at 0.64 in people with plaques, 0.15 in people without[17]. In carriers of a gene that causes early Alzheimer's it was raised from about age 25, 20 years before expected MCI[17]. The best-performing version is a ratio: the share of tau at position 217 that carries the phosphate, written %p-tau217, measured by mass spectrometry[3][18].

The best reported discrimination

Pooled over 113 studies, 29,625 participants, plasma p-tau217 reached 88.1% sensitivity, 88.7% specificity, AUROC 91.1%, the best of the blood tau markers[1]. The 2024 review of 30 studies, 9,787 people, 10 countries put plasma p-tau217 at 82% sensitivity, 86% specificity for amyloid, against 79% sensitivity, 91% specificity for the spinal-fluid version[2]. In two cohorts of 1,422 (BioFINDER-2, Sweden) and 337 (Knight ADRC, US), %p-tau217 had an AUC of 0.95 to 0.97, clinically equivalent to FDA-approved CSF tests, and for tau PET 0.95 to 0.98, generally superior to CSF tests; in cognitively impaired people its accuracy was 89 to 90%, rising to 95% with two cutoffs[3]. In 1,213 patients with cognitive symptoms, 2020 to 2024, accuracy was 91% for the APS2 blood test, 73% for dementia specialists, 61% for primary care doctors[4], and %p-tau217 alone did as well as the combined score: 90% (95% CI 88 to 91%) vs 90% (88 to 92%), not different[5]. On a fully automated platform in 1,767 participants in 5 cohorts, including 548 in primary care, accuracy was 89 to 94% in memory clinics, 85 to 92% in primary care; AUC 0.93 to 0.96[6]. In an ordinary outpatient memory clinic of 509 patients, p-tau217 had 95% sensitivity, 82% specificity; 86% specificity with the p-tau217/Aβ42 ratio, and a two-cutpoint approach gave a definitive answer in 86% of patients[21]. In 912 people with cognitive concerns, mean age 71, the AUC was 0.94 (95% CI 0.90 to 0.97), validated in an independent 118[20].

Not every p-tau217 test is the same test. Compared head to head in 998 people (BioFINDER-2), replicated in 219 (Washington University), mass-spectrometry %p-tau217 reached an accuracy of 0.93 for amyloid PET status against 0.83 to 0.88 for the immunoassays; the authors' reading was that %p-tau217 may be considered a stand-alone confirmatory test; some immunoassays might be better suited as triage tests whose positive results are confirmed with a second test[18]. In 344 people, median ages 62 and 76, four immunoassays agreed with each other at correlations of 0.40 to 0.86, reached an AUC of 0.84 to 0.90 for amyloid, and only 0.52 to 0.61 for tau pathology[19].

Where it is weakest

Kidney disease. p-tau217 is cleared by the kidneys, so poor kidney function raises it without any Alzheimer's. In 202 people, mean age 78, 43% with CKD stage 3 to 4, p-tau217 was 31% to 55% higher at an eGFR of 45 than at 60 in amyloid-negative people, and the %p-tau217 ratio rose by only 19%[22]. A Swedish study found the same: the pT217/T217 ratio was less associated with kidney disease than p-tau217 alone[23]. In the outpatient clinic above, reduced kidney function was associated with higher p-tau217 in patients without Alzheimer's[21]. In a Korean study of 2,571 people (UGOT assay), 1,578 (Roche), 304 (C2N %p-tau217), accuracy in people with kidney disease rose from 0.65 (95% CI 0.57 to 0.72) with the standard cutoff to 0.83 (0.76 to 0.89) with a subgroup-specific one, and in anaemia from 0.80 to 0.86; a double cutoff cut false classifications but left 12% to 39% of results intermediate[24].

Body weight, age and other conditions. In a US study with a training set of 289 (28% Black) and a test set of 846 (12% Black), false negatives were more likely in people with normal cognition (OR 7.79), women, and people with a high body mass index[25]. On the automated platform, accuracy fell to 83% at age 80 and over[6]. In 617 people, 44% Black, drawn from electronic health records rather than a research clinic, p-tau217/Aβ42 was also associated with kidney function, diabetes and hearing loss, and medical conditions were most frequent in Intermediate classifications[31].

Race and ethnicity. The evidence here is thinner than for the test itself. In the US study just cited, the p-tau217/Aβ42 ratio had an accuracy of 0.88 vs 0.91, with no classification difference by race (OR 1.45, 95% CI 0.66 to 2.95), but 73% of Black participants were cognitively normal, and samples came from specialised memory centres[25]. In a Chicago cohort of 431 Black and 583 White participants, followed up to 15 years, pTau217 predicted dementia within three years with an AUC 0.81 (Black), 0.77 (White); biomarker levels by race were similar; associations with kidney dysfunction and BMI persisted after adjustment[26]. The outpatient clinic study records that most patients were non-Hispanic White[21]. A 2025 review of the FDA clearance notes that the cleared test is not well validated across diverse racial and ethnic groups; not for screening people without symptoms or for distinguishing Alzheimer's from other brain diseases[7].

What it reflects

Tau phosphorylated at position 181 was the first blood tau marker to work, and it also rises with amyloid plaques. In 1,113 people, 89% non-Hispanic White, change in plasma p-tau181 tracked cognitive decline and shrinkage in Alzheimer's-typical brain regions and, unlike NfL, was Alzheimer's-specific[33]. In a single-centre clinical cohort, p-tau181 distinguished frontotemporal dementia from Alzheimer's; NfL separated degenerative dementias from healthy controls[34]. Measured in midlife in 1,525 people, mean age 58 at the first sample, 26% Black, p-tau181 and the amyloid ratio were the only markers already linked to dementia decades later: HR 1.15 per SD higher p-tau181 (95% CI 1.06 to 1.25); 1.11 per SD lower Aβ42/40 (1.02 to 1.21)[35].

The best reported discrimination

In the Lancet Neurology review it pooled to 80.5% sensitivity, 76.4% specificity, AUROC 81.5%, clearly below p-tau217[1]. In an international study of 1,298 plasma samples, six centres, p-tau181 was the best single marker for separating Alzheimer's from controls, frontotemporal dementia and amyloid-negative Lewy body dementia, but pre-defined cutoffs did not reproduce that performance[32]. It still has a place in panels: in 450 patients, three academic clinics, showing clinicians a panel of p-tau181, GFAP and NfL after the standard work-up moved their median confidence from 80% to 90%, and changed the suspected cause in about 5%[36].

Where it is weakest

Its specificity is the problem: at 76.4% specificity pooled, roughly one person in four without the pathology would test positive[1]. Kidney disease raises it as it raises p-tau217[23]. On race and ethnicity the picture is more encouraging than for most markers, with caveats. In 2,086 people of African American, Caribbean Hispanic and Peruvian ancestry, p-tau181 was higher in Alzheimer's in every cohort and levels were consistent across cohorts, but its predictive value varied; Aβ42/40 showed minimal differences[29]. In 617 people, p-tau181, Aβ40 and NfL were lower in African American participants by self-reported race but not by genetic ancestry[28]. In 379 people, 57% Hispanic, p-tau181 tracked amnestic MCI, dementia and amyloid, with no difference by ethnicity[30]. In a biracial community sample of 267 African American and 268 White adults, there were no biomarker differences by race; kidney disease, sex and vascular risk did matter[27].

What it reflects

Amyloid comes in fragments of different lengths. As the 42-length fragment is drawn into plaques, its ratio to the 40-length fragment in blood falls. The fall is small, and blood amyloid is partly made outside the brain: in 2,279 dementia-free adults, median age 72, followed 8.3 years on average, Aβ42/40 was the one marker kidney function did not disturb, which the authors explained by noting that blood amyloid is lower than in CSF and partly peripheral in origin, so it only partially reflects brain pathology[40]. Older assays could not detect the ratio at all: a 2016 meta-analysis found plasma Aβ42 and Aβ40 were not associated with Alzheimer's[39].

The best reported discrimination

High-precision mass spectrometry changed that. In 158 mostly cognitively normal people, the ratio had an AUC of 0.88 (95% CI 0.82 to 0.93); 0.94 with age and APOE ε4[37]. In 465 people from the US, Australia and Sweden, with blood collected and processed differently at each site, the AUC was 0.84 (95% CI 0.80 to 0.87); 0.88 with APOE ε4[38]. Its real strength is as a partner. PrecivityAD2 combines %p-tau217 with Aβ42/40: in 583 people, 53% amyloid-PET positive, its AUC against amyloid PET was 0.94 with agreement of about 88%, though several authors are employees or consultants of the maker[41]; in an independent cohort it reached accuracy 91% (95% CI 86 to 94%), sensitivity 90%, specificity 92%, AUC 0.95[42]. The FDA-cleared test is a ratio of p-tau217 to Aβ42: in its study of 499 plasma samples from cognitively impaired adults, 91.7% of positive results matched amyloid on PET or CSF and 97.3% of negative results matched a negative PET or CSF result, with less than 20% indeterminate[FDA]. The same ratio on the same platform in 391 clinic and 121 community participants had an AUC 0.963 to 0.966 and fewer intermediate results than p-tau217 alone, 10.6% vs 13.0% (clinic), 16.5% vs 31.4% (community)[43]; a separate cohort put it at AUC 0.927 (95% CI 0.900 to 0.954), the highest of the markers tested[44].

Where it is weakest

On its own it is the weakest of the three Alzheimer's-specific markers, and it is the most sensitive to how blood is collected and stored[38]. In the diverse-ancestry study above, Aβ42/40 showed minimal differences[29]. In people with type 2 diabetes and overweight or obesity, neither Aβ42/40 nor p-tau181 tracked cognitive decline[59]. Both PrecivityAD2 validations were written by people with a financial interest in the test[41][42].

What it reflects

Glial fibrillary acidic protein comes from astrocytes, the brain's support cells, when they react to injury. It rises early: in people with amyloid but no symptoms, plasma GFAP was 285 vs 185 pg/mL in one cohort and 170 vs 122 pg/mL in another, and it was associated with tau pathology only in people who also had amyloid[45]. In 139 nursing-home dementia patients, with autopsy on average 139 days after the last blood sample, serum GFAP was associated with tau pathology independent of amyloid, and separately with hippocampal sclerosis and argyrophilic grain disease[47]. It is not specific to Alzheimer's.

The best reported discrimination

In 160 patients followed 4.7 years on average, GFAP detected abnormal spinal-fluid amyloid with an AUC 0.79 (95% CI 0.72 to 0.86); 0.86 with APOE genotype, and predicted conversion to Alzheimer's dementia with an AUC 0.84 (95% CI 0.77 to 0.91)[46]. Its stronger suit is prognosis. In 949 cognitively unimpaired older adults from the A4 trial and its LEARN cohort, followed 240 weeks, higher GFAP predicted faster decline, CDR progression, cortical atrophy and amyloid conversion, with stronger effects in women[48]. In 1,712 dementia-free adults followed 19 years, the hazard ratios for dementia in the top third against the bottom third were HR 1.49 (NfL), 1.38 (GFAP); both in the top third, 2.06, and for death from dementia HR 2.87 (NfL), 2.76 (GFAP); both, 9.22 (95% CI 4.48 to 18.9)[49]. Among 300 people who were cognitively normal when first seen at a memory clinic, the figures were high GFAP: HR 3.6 for incident dementia (95% CI 2.2 to 5.7); NfL 1.4 (0.8 to 2.5), not significant[50].

Where it is weakest

There is no diagnostic meta-analysis of blood GFAP for Alzheimer's; the astrocyte-marker meta-analysis covered 33 articles, 3,204 people; CSF GFAP SMD 0.94 (95% CI 0.71 to 1.18); the authors call the field early days[51]. Sex matters: the A4 effects were stronger in women[48]. Vascular disease moves it: midlife diabetes was associated with a 0.15 SD faster GFAP rise[35], while in a biracial community sample diabetes and GFAP were inversely associated[27]. A 2026 meta-analysis, still a preprint and not yet peer reviewed, pooled 168 publications, 139 cohorts and found that GFAP separation in Alzheimer's was higher in predominantly Asian cohorts, p-tau181 separation in MCI lower; no between-population differences in amyloid-defined comparisons[52].

What it reflects

Neurofilament light chain is a structural protein of nerve fibres, released when they are damaged by anything. A 2024 review calls it a validated marker of nerve-fibre damage that rises in multiple sclerosis, ALS, stroke, brain injury and frontotemporal dementia as well as Alzheimer's; ageing and other confounders must be considered[56]. In spinal fluid from 10,059 people, it was raised in a majority of the neurological conditions studied[55]. The blood level tracks the spinal-fluid level: in 193 controls, 197 people with MCI and 180 with dementia, plasma correlated with CSF NfL (ρ 0.59)[53].

The best reported discrimination

As a marker of progression rather than diagnosis. In 405 people from families with inherited Alzheimer's, 243 mutation carriers and 162 non-carriers, the rate of change in serum NfL separated carriers from non-carriers 16.2 years before expected symptom onset, against 6.8 years for a single measurement[54]. In the ARIC cohort, late-life NfL had the largest association with later dementia of any marker, HR 1.92 per SD (95% CI 1.72 to 2.14)[35]. In the Cardiovascular Health Study its top-third hazard ratios were 1.49 for dementia and 2.87 for death from dementia[49].

Where it is weakest

As an Alzheimer's test. In the 2020 autopsy cohort it reached only 0.50 to 0.72 against p-tau217's 0.89[17]; in 1,275 people from normal cognition to dementia, NfL correlated weakly with core Alzheimer's pathology and tracked planning and attention more than memory[57]. Among cognitively normal memory-clinic attenders it did not predict dementia once other markers were accounted for[50]. Kidney function moves it: in 744 70-year-olds, 76 with chronic kidney disease, plasma NfL was 18.8 vs 14.1 pg/mL with and without kidney disease, and top-quartile NfL carried an HR of 2.39 for later kidney disease[58]; the association between NfL and dementia was HR 3.85 with impaired against 1.84 with preserved kidney function[40]. Midlife hypertension was associated with a 0.15 SD faster NfL rise per decade[35]. Yet in 758 people, mean age 61.5, the markers that tracked decline were rising NfL and GFAP, not Aβ42/40 or pTau-181[59]: in people with vascular and metabolic disease, the non-specific markers may carry information the Alzheimer's-specific ones miss.

A positive blood test is not a diagnosis. The FDA's clearance says the test is not intended as a screening or stand-alone diagnostic test; results must be interpreted in conjunction with other patient clinical information[FDA]. It names the risks plainly: false positives, which could lead to inappropriate diagnosis and unnecessary treatment, and false negatives, which could delay effective treatment[FDA]. It is cleared for adults aged 55 and older with signs and symptoms of the disease, in a specialized care setting, through 510(k) clearance with Breakthrough Device designation, as substantially equivalent to the same maker's CSF β-amyloid ratio test[FDA]. It is not cleared for people without symptoms, and a 2025 review adds that it is not well validated across diverse racial and ethnic groups; not for screening people without symptoms or for distinguishing Alzheimer's from other brain diseases[7].

What the test says is whether amyloid is probably present. Amyloid is necessary for Alzheimer's under the 2024 criteria, which define Alzheimer's biologically and incorporate blood-based biomarkers into the diagnostic categories for the first time[11]. But many older people carry amyloid without symptoms, and a positive result in someone with mild cognitive impairment raises questions the evidence cannot yet answer: a 2026 perspective finds limited data on when to test, how to counsel about the risk of progression, or which treatments to recommend[60]. Nor do blood tests see tangles well in early disease: four immunoassays reached an AUC of only 0.52 to 0.61 for tau pathology[19].

The number on the report depends on which test was run. The Alzheimer's Association's 2025 guideline applies to individuals with objective cognitive impairment, including MCI or dementia, evaluated in specialized care for memory disorders where Alzheimer's is the suspected cause[AA], in other words specialised care settings and symptomatic patients only; evidence rated moderate to very low certainty[8]. It sets two bars: a blood test with 90% or higher sensitivity and 75% or higher specificity can be used to triage, and one with 90% or higher sensitivity and 90% or higher specificity can substitute for amyloid PET[AA]. Its panel warns of significant variability in diagnostic test accuracy; many commercially available BBM tests do not meet these thresholds, especially using a single cutoff[8][AA]. Its evidence review covered 49 observational studies, 31 different blood tests, with sensitivity 49.3 to 91.4%, specificity 61.5 to 96.7%, and advises that performance is assay- and platform-specific; interpret in the context of the specific test used and a comprehensive clinical assessment[9]. The Global CEO Initiative's workgroup had set similar targets a year earlier: for a triage test before PET or CSF, sensitivity of 90% or more with specificity of 85% or more in primary care, 75 to 85% in secondary care; for a confirmatory test used alone, about 90% sensitivity and specificity, equivalent to CSF tests[10]. It adds that predictive values vary with the pre-test probability of amyloid pathology and must be interpreted in the complete clinical context[10]: the same result means something different in a memory clinic, where most patients have the disease, and in a general practice, where most do not. A 2026 clinical chemistry review puts the concern in one sentence: variability in assay performance and uncertainty in interpreting results in the general population raise concerns about widespread implementation[61].

What a negative result means. In the FDA's study, 97.3% of negative results matched a negative PET or CSF result[FDA]. A negative p-tau217 result in a symptomatic person makes Alzheimer's unlikely as the cause and points the work-up elsewhere. It does not rule out other dementias, which is where p-tau181 against frontotemporal dementia[34] and the non-specific markers GFAP and NfL[32] may add something.

• Outcome trials. Every study above measures agreement with PET, CSF or autopsy, or association with later decline. None shows that testing changed what happened to the person tested. The 2026 perspective on positive results in MCI finds limited data on when to test, how to counsel about the risk of progression, or which treatments to recommend[60].
• Pre-registered thresholds. The Lancet Neurology review rated about 90% of pooled studies at high risk of bias for lack of predefined thresholds[1]; the international p-tau181 study could not reproduce its results with pre-set cutoffs[32]. Prospective studies that fix the cutoff first would settle the real-world numbers.
• Diverse and multimorbid populations. The cleared test is not well validated across diverse racial and ethnic groups[7]; the largest primary-care study was Swedish[4]; the Korean cutoff study found kidney disease and anaemia need their own thresholds[24]; the strongest ethnoracial pooling is still a preprint[52].
• People without symptoms. No test is cleared for screening[FDA]. Whether finding amyloid in a person with normal cognition helps them is a question for the prevention trials, not for the tests.
• Real costs. Every cost figure on this page is modelled[13][14][15]. Measured prices, coverage and turnaround times in practice would replace them.
• Tau in blood. The current tests flag amyloid well and tangles poorly in early disease[19]. A blood marker of tangle load would change how far a test can stage the disease.