Plasma phospho-tau 181 and p-tau 217 can distinguish people with Alzheimer’s disease from controls most, but not all, of the time. To increase diagnostic accuracy, researchers are creating algorithms combining p-tau with other fluid biomarker and diagnostic data. In the May 24 Nature Medicine, researchers led by Oskar Hansson and Sebastian Palmqvist, both at Lund University, Sweden, reported just such an algorithm to predict AD. They report that among people with a subjective memory complaint, plasma p-tau, APOE genotype, executive function, and memory scores together predicted AD onset within two to six years with 90 percent accuracy. In the same cohort, clinicians were about 72 percent accurate. Alzforum first reported on the algorithm at this year’s International Conference on Alzheimer’s and Parkinson’s Diseases (Apr 2021 conference news). 

  • Algorithm predicts AD in people with memory complaint or MCI.
  • It includes plasma p-tau181 or 217, APOE genotype, cognitive tests.
  • AUCs are 0.90 and greater for dementia onset within two to six years.

To build and compare prediction algorithms, first author Palmqvist and colleagues analyzed data from 340 participants in the Biofinder cohort, of whom 164 had subjective cognitive decline and 176 had mild cognitive impairment. The researchers considered a number of variables, including demographics, ApoE4 genotype, cortical thickness, cognitive measures, and levels of plasma and cerebrospinal fluid markers. First, they prioritized algorithms that found the best compromise between fit and complexity; then they sequentially omitted variables to search for simpler ones that retained the same performance power.

Algorithm Design. Clinicians were least accurate in predicting Alzheimer’s disease dementia (black box). Plasma p-tau217 by itself did slightly better, while incorporating additional measures increased accuracy. The best model had an AUC of 0.92, but the most parsimonious one did as well with fewer variables. [Courtesy of Palmqvist et al., Nature Medicine, 2021.]

Which combination of variables worked best? Plasma p-tau217 alone predicted dementia within two, four, and six years with AUCs of 0.78, 0.83, and 0.85, respectively—AUC, aka Area Under the Curve, reflects the specificity and sensitivity of a measure, with a value of 1.0 being perfection and rarely achieved. A model that incorporated APOE genotype, executive function and memory scores did better, predicting AD within four years with an AUC of 0.90. Even better was adding plasma NfL and cortical thickness measured by MRI; these six variables combined brought the AUC to 0.92 (see image above and plot below).

Improving Prediction. Plasma p-tau217 (pink line) predicted Alzheimer’s disease over four years with an AUC of 0.81. Sequentially adding cognition (blue), and APOE genotype (green) improved accuracy, but the best fit (AUC 0.92) came from including plasma NFL and cortical thickness as well (red). [Courtesy of Palmqvist et al., Nature Medicine, 2021.]

Other models did better over shorter or longer time frames. The same six factors plus sex, years of education, and verbal ability gave the best two-year predictions, with an AUC of 0.91. The six plus plasma Aβ42/Aβ40 best predicted AD over six years, with an AUC of 0.94.

How did this compare to clinical diagnosis? In a subset of 285 Biofinder participants, memory clinic clinicians predicted AD within four years with an AUC of 0.72.

Next, the authors tested their algorithms using data from ADNI. Of the 543 volunteers in this cohort, 106 had enrolled with subjective cognitive decline and 437 with MCI. Palmqvist and colleagues tested similar algorithms here, though they had to exclude plasma Aβ42/Aβ40 because it was not available for all participants, and they had to swap plasma p-tau181 for p-tau217 because the latter was unavailable.

Again, the four-variable model with p-tau181, APOE4, executive function, and memory posted the same power as in Biofinder. It predicted AD over a four-year period with an AUC of 0.90, with the proviso that each algorithm used a different plasma p-tau isoform.

To run a direct comparison, the researchers converted continuous p-tau values to binary ones, i.e., positive or negative, based on cutoffs. They set cutoffs based on plasma p-tau levels in 215 and 547 Aβ-negative, cognitively normal participants in Biofinder and ADNI, respectively. Their rationale was that anyone Aβ-negative should have normal levels of p-tau since the latter has been found to rise only in Aβ-positive people (Aug 2019 news). The authors set the cutoffs for p-tau positivity at two standard deviations above the mean value in the Aβ-negative groups. This turned out to be 0.387 pg/mL for plasma p-tau217, and 38.2 pg/mL for p-tau181.

Even though binary values are by definition less informative than a continuous variable, in this case the cutoffs only marginally reduced the four-factor algorithm's accuracy, dropping the four-year AUC by 0.01 in Biofinder and by 0.04 in ADNI. In essence, plasma p-tau181 and p-tau217 gave almost identical results. “This implies that normalized, well-functioning assays measuring any of the p-tau epitopes could be used in this algorithm, making it widely applicable,” Hansson said. The scientists built an online tool for this modified algorithm

How about using this in the clinic? That is a hot topic in the field right now. Co-author Kaj Blennow, University of Gothenburg, Sweden, thinks algorithms like this are best suited to gauge AD risk. “At present, I’m not ready to call these algorithms diagnostic tools, but they give very good risk scores for future AD,” he told Alzforum.

Hansson believes their algorithms can soon be implemented in memory clinics but need more work for use in primary care. “While the Biofinder and ADNI cohorts enrolled representative, heterogenous memory clinic populations, we need to validate our algorithm in multiple primary care cohorts,” Hansson told Alzforum. Palmqvist agreed, adding, “We also need to see how the algorithm works in a more ethnically diverse population.”

To this end, Palmqvist, Hansson, and colleagues last spring began recruiting older people in 15 Swedish primary care clinics. So far, they have included 140 of their 600-person goal.

Before such algorithms find use in routine diagnosis, assays for various p-tau isoforms need to be standardized, just as the field is currently doing for plasma Aβ assays.

In the meantime, p-tau plasma testing is already rolling out in some leading clinics to aid in diagnosis. “In our lab, we added plasma NfL as a clinical diagnostic last year and we plan to add p-tau, either 181 or 231, this year,” said Blennow.—Chelsea Weidman Burke


  1. This paper highlights the utility of algorithms as part of a next step in taking blood-based biomarkers to the clinic and maximizing prediction of Alzheimer’s disease. Incorporating information from multiple measures, including cognition and plasma p-tau, provides better accuracy for predicting progression to Alzheimer’s disease dementia than either alone.

    As the authors mention, additional validation is needed in larger sample sizes, especially from studies that use the same exact measures, because the variables in the current analysis slightly differed between studies. It will be important for this validation to be conducted among community-based populations seen in primary care, in more diverse samples, and also in older populations. The mean age of volunteers in both ADNI and BIOFINDER was 72-73 years. 

    Additional work is also needed to determine what is feasible for primary care use as compared to memory clinics. Primary care physicians likely will not have the time to conduct cognitive testing beyond a general screening measure such as the MMSE, MoCA, or AD8. Therefore, future work should incorporate these screening measures into the algorithm and then compare the accuracy to the more extensive cognitive testing. It is likely that different algorithms will be developed depending on the population, for example, one for screening purposes in primary care with only a cognitive screen, and another for more specialized memory clinics with more intensive testing. 

  2. This is an impressive study that builds on the usefulness of plasma p-tau in AD. We know from different studies that plasma p-tau provides high diagnostic value for AD. Palmqvist et al. now show that plasma p-tau can be also useful for predicting progression to AD dementia, and it highlights the value of combining plasma p-tau with other clinical measures, instead of using measures separately. These additional measures are often obtained during the clinical evaluation but the information is not used for predicting the disease course.

    This paper is important for clinical practice because historically, predicting individual progression in AD has been challenging. This is confirmed by Palmqvist et al. in that plasma p-tau alone performed better than clinical prediction. It will be important to know how to integrate plasma p-tau and other measures with CSF in clinical practice and how the model predicts progression in those with prodromal AD confirmed by CSF or PET. In addition, it will be important to see if the model also works in less selected individuals who often have comorbidities that can also influence disease progression.  

  3. Algorithms that predict progression to AD dementia using blood-based measures and readily available individual-level factors (e.g., age and APOE genotype) are likely to be very useful in research, clinical trials, and the clinic. Adding cognitive test scores may also be feasible, although the normative ranges may vary in different groups. Adding complex imaging measures (e.g., AD signature cortical thickness by MRI) is much less useful, as this metric is rarely available outside of highly specialized research centers. Fortunately, Palmqvist et al. have found that a model with plasma p-tau217 or p-tau181 and relatively easily obtained measures (age, years of education, number of APOE ε4 alleles, and performance on a few cognitive tasks) is highly predictive of progression to AD dementia. Their model was established in the Swedish BioFINDER study and tested in ADNI. Despite significant differences between these cohorts, the algorithm performed well.

    Although plasma p-tau181 alone was less accurate than CSF p-tau181 alone in predicting progression, it is remarkable that Palmqvist et al. did not see a major difference between models using plasma biomarkers compared to CSF biomarkers. This suggests that through the use of algorithms, plasma biomarkers may achieve comparable accuracy to CSF biomarkers.

    It seems likely that companies performing blood-based biomarker tests will also provide an interpretation that considers individual-level variables; this will improve the accuracy and usefulness of the test. For example, C2N’s PrecivityAD test includes an “amyloid probability score” that factors in plasma Aβ42/Aβ40, ApoE proteotypes, and age. Notably, all current clinical AD tests are considered to be for brain amyloidosis, not symptomatic AD. Algorithms including variables that mediate the association between brain amyloidosis and symptomatic AD may finally enable a test for brain amyloidosis to become a test for symptomatic AD. 

  4. This paper contributes to the growing evidence that blood-based biomarkers can be used to detect and predict AD.

    The highlight is creating an algorithm for predicting AD dementia within four years that only includes plasma p-tau217 and cognitive assessments. Importantly, these tests are noninvasive, cheap, and readily available at most clinical-care centers. Currently, the use and generalization of such prediction algorithms is limited by the absence of standardized data acquisition methods and by testing in cohorts that do not reflect the general population. In this paper however, the algorithm was established and validated in two independent cohorts, thus reducing some of this bias and increasing its reproducibility.

    Although we should not rely completely on these tools, this is very encouraging because it offers the perspective of a cheap and readily accessible option for clinicians to help evaluate patient risk and optimize diagnosis, both of which will result in earlier and adapted clinical care.

  5. The potential of plasma p-tau to revolutionize the work-up of dementia cases has been well-documented in a number of high-profile publications in the last two years. However, it has not been definitively shown to what extent p-tau can be combined with other standard measures in a battery of simplistic tests that can be used in clinical routine to improve individualized prognosis of cognitive decline.

    Now, Palmqvist and co-workers, using the BioFINDER and ADNI cohorts, have demonstrated the superior combination of p-tau, memory and executive function, and APOE status for the progression to AD dementia. This study is crucial in bringing an AD blood test closer to reality by signifying that p-tau measures do improve the clinical work-up of suspected AD.

    There are other important lessons from this study. Plasma p-tau alone is not the perfect biomarker for the prodromal phase of disease and has less accuracy than separating AD dementia from non-AD dementias (Janelidze et al., 2020Karikari et al., 2020; Palmqvist et al., 2020; Lantero-Rodriguez et al., 2020; Ashton et al., 2021). Here, it is shown that the interplay between clinical symptomology, genetic risk, and a biological marker is the best prediction of prognosis at this stage and that plasma p-tau plays a central role in such a model. Secondly, this article goes some way in deducing the true importance of assay platform and p-tau epitope choices, which is an important question currently unanswered in the field. While p-tau217 (Meso-Scale discovery) alone was marginally better than p-tau181 (SIMOA) alone (AUC 0.83 versus 0.79) in a combinational model of this kind, where values are standardized or binarized, the inclusion of p-tau181 or p-tau217 were identical. Remarkably, replacing plasma measures with CSF measure did not significantly improve the prediction models. Lund University, Eli Lily, and Gothenburg University should be commended for their collaborative efforts on this seminal study.


    . Plasma P-tau181 in Alzheimer's disease: relationship to other biomarkers, differential diagnosis, neuropathology and longitudinal progression to Alzheimer's dementia. Nat Med. 2020 Mar;26(3):379-386. Epub 2020 Mar 2 PubMed.

    . Blood phosphorylated tau 181 as a biomarker for Alzheimer's disease: a diagnostic performance and prediction modelling study using data from four prospective cohorts. Lancet Neurol. 2020 May;19(5):422-433. PubMed.

    . Discriminative Accuracy of Plasma Phospho-tau217 for Alzheimer Disease vs Other Neurodegenerative Disorders. JAMA. 2020 Aug 25;324(8):772-781. PubMed.

    . Plasma p-tau181 accurately predicts Alzheimer's disease pathology at least 8 years prior to post-mortem and improves the clinical characterisation of cognitive decline. Acta Neuropathol. 2020 Sep;140(3):267-278. Epub 2020 Jul 27 PubMed.

    . Plasma p-tau231: a new biomarker for incipient Alzheimer's disease pathology. Acta Neuropathol. 2021 May;141(5):709-724. Epub 2021 Feb 14 PubMed.

  6. The plasma biomarkers of AD pathology that have arrived on the scene in recent years are truly impressive, and p-tau217 is among the best. Elevated p-tau217 or low Aβ42/40 ratio (by mass spec) places an individual on the AD spectrum. Biomarkers such as hippocampal volume, FDG PET, or cognitive assessment can provide additional information about staging. Two or more of these measures can be used to select individuals for trials, or indicate risk of cognitive/functional decline. And when disease-modifying treatments arrive (perhaps very soon), these markers will be useful in selecting individuals for therapy. Specific algorithms or models may not be necessary, though they may ease interpretation for clinicians.

    These new plasma assays are already providing powerful tools for trials, and as they reach clinical practice they will dramatically assist in diagnosis and planning.

  7. “Among people with a subjective memory complaint” (not among community-based populations seen in primary care) the algorithm predicted AD onset within two to six years with 90 percent accuracy.

    Some comments report that one factor limiting clinical relevance is doctor time, e.g., “Primary care physicians likely will not have the time to conduct cognitive testing.” However, there are longitudinally accurate cognitive tests (e.g., CANS-MCI) that are self-administered in doctors’ offices with less than three minutes of staff time, much less doctor time. Such tests accurately detect the onset of clinical changes, even in highly functional patients, and have been found to be predictive of CSF measures (Barber et al., 2018).

    Clinical relevance is not just the ability to evaluate patient risk and optimize diagnosis. Clinical relevance means accurate reporting of cognitive symptom onset as much as prognosis. Clinicians must be mindful of the consequences when a patient is told that an untreatable condition will emerge but might not be clinically evident for five years.


    . Poster Presentation: CSF markers of preclinical Alzheimer’s and deficits on a self-administered computerized test battery. Sunday, July 22, 2018: Alzheimer's Association International Conference

Make a Comment

To make a comment you must login or register.


News Citations

  1. Where to Now, Phospho-Tau?
  2. Move Over CSF, P-Tau in Blood Also Tells Us There’s Plaque in the Brain

External Citations

  1. Biofinder
  2. ADNI
  3. algorithm

Further Reading

Primary Papers

  1. . Prediction of future Alzheimer's disease dementia using plasma phospho-tau combined with other accessible measures. Nat Med. 2021 May 24; PubMed.