A Blood Test for Alzheimer's: How Close Are We?

Two groups of Alzheimer's researchers have spent the last few years arguing, politely but persistently, about what a positive blood test actually means.

The first camp, associated with the Alzheimer's Association workgroup that revised the diagnostic criteria in 2024, defines Alzheimer's disease biologically. If you have the core pathology, you have the disease, whether or not you have symptoms. Under that framework, an abnormal plasma p-tau217 result counts as a core biomarker, and a cognitively normal 68-year-old with a high value has Alzheimer's disease at an early stage.

The second camp, represented most clearly by the International Working Group, argues that this is a category error. Many amyloid-positive people never develop dementia in their lifetimes. Calling them "patients" carries real costs, psychological, financial and legal, and the predictive value isn't there yet to justify the label. For this group, a positive biomarker in someone without symptoms marks risk, not disease.

I find myself closer to the second camp for clinical use and closer to the first for research. That split position turns out to be a useful way to read the whole p-tau217 literature, because the assay works very differently depending on who you test and why. You can follow both sides of the argument in the Alzheimer's Association revised criteria and the International Working Group's response.

Why p-tau217, and not the markers we had before

Blood biomarkers for Alzheimer's have had several false dawns. Plasma Aβ42/40 does track brain amyloid, but the difference between amyloid-positive and amyloid-negative people is small, often in the range of ten percent or less. A marker with that little dynamic range is fragile. Small shifts in tube type, time to centrifugation or assay calibration can push people across the threshold. Neurofilament light is robust but nonspecific, since it rises with axonal injury from almost any cause. GFAP is interesting biologically and less useful diagnostically on its own.

Phosphorylated tau at threonine 217 behaves differently. In amyloid-positive individuals, plasma concentrations are several-fold higher than in amyloid-negative individuals. That fold change is what makes the marker clinically workable. It survives noise that would swamp Aβ42/40.

The mechanism, as far as we understand it

The working model goes like this. As amyloid plaques accumulate, neurons respond by increasing phosphorylation and secretion of soluble tau species, with T217 among the most responsive sites. These soluble fragments reach the CSF and, at much lower concentrations, the bloodstream. Importantly, plasma p-tau217 seems to reflect the neuronal response to amyloid more than the tangle burden itself, at least early on. That's why it correlates so tightly with amyloid PET in preclinical and prodromal stages and increasingly with tau PET as disease advances.

This distinction matters more than it might seem. A test that tracks the response to amyloid is a superb amyloid proxy. It's a weaker staging tool, and it doesn't directly tell you how much tau pathology a person carries. Much of the current work is trying to close that gap.

The other mechanistic point concerns clearance. Plasma levels depend on production and on peripheral handling, including renal function. That's the main source of biological confounding, which I come back to below.

The strongest evidence: large cohorts, head-to-head comparisons and real-world accuracy

A few things are now established well enough that I'd stop hedging about them.

Plasma p-tau217 identifies Alzheimer's pathology with accuracy comparable to CSF biomarkers. Across multiple independent cohorts, including BioFINDER, ADNI and memory clinic samples in Europe and North America, high-performing p-tau217 assays discriminate amyloid-PET-positive from amyloid-PET-negative individuals with areas under the curve typically in the low-to-mid 0.9s. The better assays perform on par with FDA-approved CSF tests. Head-to-head comparisons consistently put p-tau217 ahead of p-tau181 and p-tau231 in plasma.

It distinguishes Alzheimer's from other neurodegenerative diseases. Patients with frontotemporal dementia, Parkinsonian disorders and vascular cognitive impairment generally have low values unless they also carry Alzheimer's co-pathology. That last clause matters a great deal in older patients, where mixed pathology is the norm.

It outperforms clinical judgment in symptomatic patients. The most convincing work here came from a prospective Swedish study published in 2024, in which a combined plasma measure (percentage p-tau217 together with Aβ42/40) reached roughly 90 percent accuracy for identifying Alzheimer's pathology in both primary and secondary care. Clinicians' diagnostic accuracy in the same settings, before seeing biomarker results, was substantially lower, with the gap largest in primary care. That's the result that moved this field from "promising" to "deployable." You can find it and its follow-ups through a PubMed search on p-tau217 accuracy in primary and secondary care.

Regulatory and guideline infrastructure now exists. In May 2025 the FDA cleared the first blood test intended to aid Alzheimer's diagnosis, a plasma p-tau217/Aβ42 ratio, for adults 55 and older with cognitive symptoms. Later in 2025, the Alzheimer's Association released its first clinical practice guideline on blood-based biomarkers. It recommends that tests meeting roughly 90 percent sensitivity and 75 percent specificity can serve as triage, and that tests reaching about 90 percent for both can substitute for PET or CSF confirmation. Both thresholds apply to patients with objective cognitive impairment in specialist care. Assays vary widely, and the guideline is deliberately assay-agnostic. Plenty of commercially available tests don't meet those bars.

The two-cutoff approach is the real clinical innovation

If I had to pick the single most practically important idea in this literature, it would be the two-threshold strategy rather than any one assay. Instead of one cutoff, you set a lower threshold below which Alzheimer's pathology is very unlikely and an upper threshold above which it's very likely. Values in between go to CSF or PET.

In most published implementations, the intermediate zone captures something like 15 to 30 percent of people, depending on the assay and population. That isn't a failure. It means the large majority of patients avoid a lumbar puncture or PET scan, while the uncertain minority still get confirmatory testing. I'd argue any clinical rollout that uses a single binary cutoff is throwing away the main advantage of the marker.

Observational evidence: prognosis, monitoring and the populations we haven't studied enough

This is where the confidence drops a notch, and where most of the open debate lives.

Predicting decline in cognitively unimpaired people

Longitudinal cohort data show that higher baseline p-tau217 predicts faster cognitive decline and conversion to MCI or dementia in people who are cognitively normal at baseline. The signal is consistent across cohorts. The predictive value for an individual is a different matter. In low-prevalence populations, even a test with good specificity produces many false positives, and the time horizon matters too. Plenty of people with elevated values remain cognitively stable for years.

That is precisely the argument the clinical-biological camp is making. The cohort data justify using p-tau217 to enrich prevention trials. In my view they don't yet justify telling an asymptomatic person they have a disease.

Confounders that change the numbers

Chronic kidney disease raises plasma p-tau217 and other plasma biomarkers, likely through reduced clearance. Higher BMI tends to lower concentrations, probably through a larger blood volume. Age, cardiovascular comorbidity and other conditions add further noise. Ratio-based measures, particularly the percentage of phosphorylated to non-phosphorylated tau217 used in some mass spectrometry platforms, appear to reduce some of this confounding. The literature on p-tau217 and chronic kidney disease is worth reading before interpreting any borderline value in an older patient.

The bigger limitation is who has been studied. The flagship cohorts are disproportionately white, highly educated, research-engaged and recruited from specialist settings. Performance in community samples, in populations with higher burdens of vascular and metabolic disease, and across ethnic groups is less well characterized. Some studies suggest the marker generalizes reasonably well. Others show shifts in optimal cutoffs. I wouldn't assume a cutoff validated in Lund transfers to a primary care clinic in rural Alabama.

Monitoring treatment response

Plasma p-tau217 falls in people treated with anti-amyloid antibodies, roughly in parallel with amyloid PET reductions. That makes it an attractive pharmacodynamic marker. Whether the decline predicts clinical benefit for an individual, as opposed to tracking target engagement at the group level, remains unresolved. I'd treat it as a target-engagement marker for now, not a surrogate endpoint.

Early-stage work: staging, tau-specific markers and mechanism

The most interesting frontier is the attempt to recover from blood the information that p-tau217 doesn't fully capture.

• Tangle-specific tau fragments. Microtubule-binding region tau species, notably MTBR-tau243 in CSF, track tau PET more closely than phosphorylated N-terminal fragments do. Translating these into plasma assays is technically harder and still early.

• Brain-derived tau. Assays designed to exclude peripherally expressed tau isoforms aim to provide a neurodegeneration marker more specific than total tau. Promising, with limited independent replication so far.

• Biological staging from plasma. Several groups are combining p-tau217 with other phosphorylation sites and fragments to approximate the sequence of amyloid, early tau and widespread tau pathology. If it works, it would let trials stratify participants by stage without serial PET.

• Mechanistic models. Cell and mouse work suggests tau release is activity-dependent and that site-specific phosphorylation responds to amyloid exposure. Mouse models only partly recapitulate human tau pathology, though, so I'd treat these as hypothesis-generating rather than explanatory.

None of these approaches is ready for clinical decisions. Several will probably matter for trial design within a few years.

What this means for trials

For trialists, the effect is already concrete. Screening has historically been one of the most expensive and slow parts of Alzheimer's trials, because large numbers of candidates underwent amyloid PET or lumbar puncture only to screen out. Prescreening with plasma p-tau217 sharply reduces that waste. Recent anti-amyloid programs, including the TRAILBLAZER-ALZ series, have built plasma p-tau217 into enrollment, and the preclinical prevention trials depend on it heavily. The protocols are visible on ClinicalTrials.gov.

The trial benefit is the least controversial use of the marker. Enriching a sample for people with pathology raises the event rate and statistical power without anyone being given a diagnostic label. It also makes decentralized recruitment possible, which could finally broaden the demographics of who gets into these studies.

Where the evidence sits, at a glance

Where I think the field gets it wrong

The common framing, "a blood test for Alzheimer's," is misleading in a specific way. What we have is a blood test for Alzheimer's pathology, which is a different thing in a 75-year-old with mixed vascular disease and a decade of slow decline. A positive result tells you amyloid is present and probably contributing. It doesn't tell you it's the whole story.

I also think the assay-to-assay variability gets underplayed in coverage. "P-tau217" isn't one test. Antibodies, platforms, calibrators and reporting units differ, and so do cutoffs. When someone tells you a p-tau217 test is 95 percent accurate, the right first question is which test, in which population, against which reference standard.

What to watch next

• Standardization. Certified reference materials and harmonized cutoffs across platforms would do more for clinical adoption than another accuracy study.

• Diverse, community-based validation. Prospective studies in primary care populations with realistic comorbidity burdens, outside the flagship cohorts.

• Plasma tau staging. Whether combinations of tau species can approximate tau PET well enough to guide treatment eligibility.

• Disclosure research. How people respond to learning a biomarker result when they have no symptoms, and whether that changes behavior or harms wellbeing.

• Health-economic data. Whether blood-first pathways actually shorten time to diagnosis and reduce costs once implemented at scale.

This literature is moving quickly enough that a review from eighteen months ago is already missing important papers. If you're tracking it, a living search is worth more than a static reading list. I keep mine in PACR, an AI-native research workspace where saved searches, reading notes and drafts sit in one place. When a new assay comparison appears, it lands next to my annotations on the earlier ones instead of in a separate folder I'll forget about. If you want to set up something similar for the biomarker literature, you can start exploring the research there or create a workspace and build your own living search.

My suggestion for anyone entering this area: pick one assay, learn its validation history in detail, and read every new accuracy claim against it. The field's progress is real. Its precision is something you have to check for yourself.