Plasma p-tau217 performs well in broad Alzheimer’s classification, then loses ground in narrower clinical questions
Plasma p-tau217 continues to look strong when the question is broad and the cohort is mixed. It loses some of that strength when the task becomes more specific, especially when investigators ask whether the marker can separate tau-positive from tau-negative individuals inside an already amyloid-positive group, or when they try to map plasma signals onto MRI-defined hippocampal neurodegeneration.
This preprint tested those boundaries directly in the Alzheimer’s Disease Neuroimaging Initiative, using repeated nested cross-validation with fold-internal preprocessing and calibration. The design matters because blood-based biomarkers can look deceptively good if preprocessing or threshold tuning leaks information across folds. Here, the modeling framework stayed inside the training splits, which gives the performance estimates more credibility than a single pass through the data would have allowed.
The main result is straightforward: plasma-based prediction was strongest for global AT(N) tasks and weakest for conditional tau discrimination within A+ participants and for N as defined by MRI hippocampal neurodegeneration. In other words, a blood marker that can help flag Alzheimer’s biology does not automatically become a precise substitute for the downstream tissue-specific endpoints used in research staging.
ADNI modeling showed sharp separation between broad and conditional endpoints
The internal validation set included 1,097 participants for amyloid positivity, 929 for tau positivity, 361 for tau positivity within amyloid-positive participants, 1,303 for hippocampal neurodegeneration, and 852 for joint A+T+ classification. The observed AUROCs were 0.9425 for A, 0.9495 for T, 0.8852 for T within A+, 0.7865 for N, and 0.9614 for joint A+T+.
Those numbers are not interchangeable. A plasma classifier that approaches 0.95 for tau positivity in the full cohort is doing something materially different from one that falls to 0.8852 when the same biological question is restricted to amyloid-positive participants. Once the sample is conditioned on one axis, the remaining separation is harder because the biomarker has to distinguish a narrower biological contrast against a background of already abnormal pathology.
A single sentence captures the clinical tension here: the model remained strong, but the task got harder faster than the biomarker did.
At thresholds set to preserve 90% sensitivity, specificity reached 87.5% for tau positivity overall, but only 60.7% for tau positivity within A+. That drop is clinically meaningful because the latter task is the one clinicians and trialists increasingly care about when they want to identify which amyloid-positive patients are also tau-positive. In a setting like that, false positives would be expensive and biologically confusing, especially if the test were used to enrich studies or triage patients for more invasive confirmatory testing.
MRI-defined neurodegeneration was the weakest of the main ADNI endpoints, with AUROC 0.7865. Hippocampal atrophy is not a direct plasma readout in the same way amyloid or tau pathology may be, and the weaker performance fits that biology. Structural MRI reflects cumulative tissue loss, network injury, and the downstream effects of multiple processes, so expecting a single plasma analyte to map cleanly onto it may be too optimistic.
Separate axis-specific analyses showed moderate gains, not a universal fix
Post hoc analyses extended the story in two directions. In Bio-Hermes, investigators examined an A-axis staged random-forest model using Lilly p-tau217 and C2N A40/A42 in 908 participants. Under the primary PET classification, adding A40 and A42 to p-tau217 increased AUROC from 0.894 to 0.916, with a paired difference of 0.022 and a 95% conditional interval from 0.010 to 0.035. The Brier score also changed by -0.011, with an interval from -0.017 to -0.005.
Those are modest but real gains. They suggest that plasma amyloid measures may contribute incremental axis information when the model is already anchored by p-tau217, especially in an A-specific framework rather than a broad all-purpose classifier. The finding is consistent with a multimarker logic: one plasma measure may not fully capture amyloid biology, but a carefully chosen pair can improve calibration and discrimination enough to matter in a research pipeline.
The alternative PET classification in Bio-Hermes was concordant, which strengthens the case that the gain was not a one-off artifact of a single endpoint definition.
On the opposite side of the analysis, the A4 ordinal assessment tested plasma p-tau217 against an A4-specific medial-temporal tau-PET endpoint in amyloid-positive participants. Here the marker reached an AUROC of 0.7356, with average precision of 0.5657 and Spearman ρ of 0.4718. The performance is not trivial, but it is clearly lower than the broad ADNI tau classification. The difference is important because it shows how much performance depends on the precise endpoint, the cohort structure, and the assay implementation.
Biological specificity, assay choice, and endpoint design set the ceiling
The study’s most useful message is not that plasma biomarkers fail. It is that their ceiling depends on what exactly is being asked of them. Broad amyloid and tau classification is easier because the biological signal is larger and the contrast between groups is wider. Conditional tau prediction inside A+ participants is harder because the comparison is more subtle, with less room for a single circulating marker to separate the groups cleanly. MRI-based neurodegeneration is harder still because the endpoint sits further downstream from the blood compartment.
Assay implementation also matters. The Bio-Hermes analysis used a particular p-tau217 assay paired with amyloid measures, while the A4 assessment relied on an ordinal p-tau217 framework against a medial-temporal tau-PET endpoint. Even when the underlying biomarker is the same, platform differences, calibration strategy, and endpoint definition can shift performance enough to change how the test would behave in practice.
This is where the study becomes most relevant for Alzheimer’s trials. If a sponsor wants to enrich for amyloid positivity, one plasma model may be sufficient. If the goal is to identify tau burden within an amyloid-positive population, the evidence here suggests more caution. And if the goal is to infer MRI-defined neurodegeneration from plasma alone, the present data argue against overpromising.
What the preprint can and cannot claim
The analysis is retrospective, restricted to controlled-access datasets, and still awaiting peer review. It also mixes internal validation with post hoc external or semi-external axis-specific assessments, so the strongest claim is about performance boundaries rather than definitive clinical deployment. Even so, the pattern is coherent across datasets: plasma p-tau217 is useful, but its utility depends on the axis and the exact question being posed.
The practical consequence is a narrower one than many biomarker headlines would suggest. Plasma markers may support staging and enrichment, but they do not erase the need for endpoint-specific validation, especially when the target is conditional tau positivity or MRI-defined neurodegeneration. Future work will need to determine whether these ceilings can be moved by different assays, multi-marker combinations, or alternative modeling strategies, and whether the same limits hold outside ADNI and A4-style cohorts.