Published on 16 September 2026
Amyloid-positive mild cognitive impairment: does the result tell you when dementia will follow?
In brief
In the French BALTAZAR cohort, 173 patients with mild cognitive impairment were followed for three years, with paired measurement in cerebrospinal fluid and plasma of more than one hundred and twenty central nervous system biomarkers. Sixty-three of them, or 36%, converted to dementia. The result separates two questions that practice often runs together. To say whether amyloid pathology is present, plasma phosphorylated tau markers are excellent, with an area under the curve above 0.90. To say which of these patients will progress to dementia within three years, p-tau217 falls to 0.73 in cerebrospinal fluid and 0.72 in plasma. The most predictive combination in cerebrospinal fluid brings together an inflammatory marker, IL-16, a neurodegeneration marker, total tau, and a synaptic marker, NPTX2, with an area under the curve of 0.86 and a 95% confidence interval of 0.80 to 0.91. The hazard ratio attached to this combination, 39.8, compares the third tertile of the score with the first and carries a confidence interval of 9.6 to 165.2: it should be read as a signal of direction, not as a usable estimate. The assay platform used is a research tool, and the publication reports no routine use of this signature.
The context
A seventy-eight-year-old patient comes in because of forgetfulness. The work-up comes back positive for amyloid. The question he asks next is not the one the test answers. He does not ask whether the disease is there, he asks when it will start to get in his way. The amyloid cascade has shaped diagnosis for twenty years, and plasma phosphorylated tau markers have made that diagnosis possible without a lumbar puncture. Prognosis, for its part, has been left without a tool.
This study is of interest because it does not replace one marker with another. It shows, in a single sample and with a single assay technique, that a marker can perform very well on one question and poorly on another, and that diagnostic performance does not automatically carry over to prognostic performance.
The mechanism
Two families of markers, two distinct pieces of information
Amyloid and phosphorylated tau markers signal that the pathological process is present. Neurodegeneration markers, such as total tau and neurofilament light chain, and synaptic markers, such as NPTX2, signal the functional damage under way. The hypothesis the authors put to the test is that the second family says more about the pace of the disease than the first.
One point of reading is worth setting out from the start. In the model retained, the coefficients for IL-16 and NPTX2 are negative and the coefficient for tau is positive: it is low concentrations of NPTX2 and IL-16, together with high tau, that point towards conversion. Taken on its own, IL-16 has no predictive value in this cohort, with an area under the curve of 0.54. Its contribution exists only inside the model.
What the study measures: protein concentrations in cerebrospinal fluid and in plasma, and their statistical association with conversion to dementia. What it does not measure: the causal mechanism linking synaptic loss to cognitive decline, or the effect of any intervention on these markers. The authors themselves point out the lack of neuropathological confirmation.
The study at a glance
| Population | |
| 173 patients with mild cognitive impairment according to Petersen’s criteria, drawn from the French prospective multicentre BALTAZAR cohort (NCT01315639). Mean age 77.8 years (standard deviation 5.5), 38.7% men, 38.2% carrying one or two APOE4 alleles. Amyloid-positive status in 55.5% of participants, defined by the Aβ42/40 ratio in cerebrospinal fluid. Conversion to dementia at three years in 63 patients, or 36%, including 50% of amyloid-positive and 19% of amyloid-negative participants. The number of participating centres is not stated in the publication. | |
| Exposure measured | |
| Simultaneous measurement of more than one hundred and twenty central nervous system biomarkers in cerebrospinal fluid and in plasma, sampled on the same day, on the NULISA multiplex platform. The panel comprises 52 inflammatory markers, 24 neurodegeneration markers, 12 vascular and metabolic markers, 21 markers related to synuclein and synaptic function, and 22 related to amyloid and tau. | |
| Comparator | |
| Markers of amyloid pathology and of phosphorylated tau taken individually, p-tau217 in particular, together with amyloid status established on cerebrospinal fluid. | |
| Outcome | |
| Conversion from mild cognitive impairment to dementia at three years, defined by cognitive decline and loss of autonomy (MMSE, IADL, ADL, CDR greater than or equal to 1) and validated by an adjudication committee applying the NIA-AA criteria, blinded to biomarker results. Performance assessed by area under the curve and by hazard ratios comparing the third with the first tertile of the score. | |
| Design | |
| Prospective biomarker analysis carried out on a previously assembled cohort, three-year follow-up with reassessment every six months. Variable selection by elastic net penalised regression (alpha 0.5), five-fold cross-validation, 100 bootstrap iterations, then 300 resamples to estimate the areas under the curve and their intervals. Combinations limited to three markers to respect the ten events per variable rule. Kaplan-Meier survival analysis with log-rank test. No independent validation cohort. |
The quality check
| Criterion | Judgement |
|---|---|
| Prospective data collection and adjudication | Sound |
| FindingFrench prospective multicentre cohort, reassessment every six months over three years, conversion validated by an adjudication committee blinded to the assays. The biomarker question is asked of a pre-existing cohort, which makes this a secondary analysis and not a study built for this hypothesis. | |
| Exploration without a prior hypothesis | Sound |
| FindingMore than one hundred and twenty markers measured simultaneously, penalised statistical selection, bootstrap resampling and a limit of three markers per model, which reduces overfitting without eliminating it. | |
| Consistency of the main message | Sound |
| FindingThe dissociation between diagnostic and prognostic performance can be read directly from the areas under the curve compared within the same sample, it is not inferred. | |
| Precision of the hazard ratio | Unstable |
| FindingThe confidence interval of the hazard ratio for the best combination runs from 9.6 to 165.2. The authors state that these values cannot be carried directly into clinical practice, since they compare the first and third tertiles, leaving out the middle tertile and without defining a cut-off. | |
| Two conversion figures in the text | To note |
| FindingThe abstract and results give a conversion rate of 36%, the discussion mentions 40%. The 36% figure is the one that matches the numbers in Table 1, 63 converters out of 173. | |
| Status of the technique | Research |
| FindingThe multiplex assay platform used is a research tool, and the publication reports no routine use of this signature. An important nuance: the authors show that tau measured with a routine ELISA, substituted for the multiplex measurement, gives the same gain when combined with NPTX2 and IL-16. The substitution concerns tau only: for these last two markers, the publication reports no routine assay. | |
| Funding and competing interests | Public, none declared |
| FindingFunded by the French Ministry of Health (hospital clinical research programme, PHRC 2009/01-04 and PHRC-13-0404), the Fondation Plan Alzheimer, the Fondation pour la Recherche Médicale, France Alzheimer and the Gérontopôle d’Île-de-France, with no role for the funders in design, analysis or the decision to publish. The authors declare no conflict of interest related to this manuscript. | |
The findings
| Outcome | Value |
|---|---|
| Plasma phosphorylated tau markers, detection of amyloid pathology | Area under the curve above 0.90 |
| ReadingHigh diagnostic performance, consistent with the recent literature on these markers. This value appears in the abstract of the publication, with no marker-by-marker figures in the tables, which all concern conversion. | |
| p-tau217, prediction of conversion | 0.73 (0.64 to 0.81) in cerebrospinal fluid, 0.72 (0.63 to 0.80) in plasma |
| ReadingThe same marker loses a good deal of performance when the question becomes prognostic. The authors add that sensitivity and specificity remain below 75%, which makes it difficult to use for an individual decision. This is the central result of the study. | |
| Best combination in cerebrospinal fluid | IL-16, tau and NPTX2: area under the curve 0.86 (0.80 to 0.91) |
| ReadingAn inflammatory marker, a neurodegeneration marker and a synaptic marker. Taken individually, the same markers reach 0.54 for IL-16, 0.63 for NPTX2 and 0.73 for tau: the gain comes from the combination, not from its components. | |
| Hazard ratio for this combination | 39.8 (9.6 to 165.2), third tertile versus first |
| ReadingThe interval spans more than an order of magnitude, and the comparison leaves out the middle tertile. To be read as the direction of a signal, not as an estimate of the size of the risk. | |
| Two-marker combination | Tau and NPTX2: 0.83 (0.76 to 0.88), hazard ratio 15.8 (5.6 to 44.8) |
| ReadingTwo markers are enough to come close to the performance of the trio, which the authors note as an argument for feasibility. | |
| Amyloid status alone, on cerebrospinal fluid | Hazard ratio 3.2 (1.9 to 5.0) |
| ReadingA useful reference point: amyloid status as it is already used in routine practice carries real prognostic information, but on an entirely different scale from that of the models explored here. | |
| Plasma markers | Modest improvement |
| ReadingAdding YWHAG, the gamma isoform of the 14-3-3 protein, takes p-tau217 from 0.72 (0.63 to 0.80) to 0.79 (0.73 to 0.86) and p-tau181 from 0.72 (0.64 to 0.79) to 0.78 (0.70 to 0.85). The best plasma combination in the table, p-tau217, VCAM1 and YWHAG, reaches 0.81 (0.75 to 0.87). The authors attribute this limit to the absence of a robust synaptic marker in blood. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Selection of participants | Low risk |
| FindingProspective multicentre recruitment of patients with mild cognitive impairment according to Petersen’s criteria. The population is older, mean age 77.8 years, and mostly female, which brings it closer to the patients seen in memory clinics, French ones in this case, than to the cohorts of therapeutic trials. | |
| Measurement of the exposure | Low risk |
| FindingSimultaneous assays on a single platform, from paired samples taken on the same day, which removes methodological differences between the markers compared. Two cerebrospinal fluid samples were excluded because of detection failure, and ten markers detected in fewer than 40% of samples were left out of the analysis. | |
| Measurement of the outcome | Low risk |
| FindingConversion remains a clinical diagnosis, but it was reviewed by an adjudication committee ruling on clinical and neuropsychological data alone, blinded to biomarker results. | |
| Overfitting and validation | Moderate risk |
| FindingPenalised selection, cross-validation and resampling reduce overfitting, but there is no validation in an independent sample. The authors say so explicitly: the effect sizes may partly reflect model instability or overfitting in the context of a single cohort. The reported performance is therefore optimistic. | |
| Partial circularity | Moderate risk |
| FindingAmyloid status was defined from cerebrospinal fluid biomarkers, and other biomarkers from the same fluid were then analysed according to that status. The authors acknowledge this circular reasoning, which does not affect the conversion outcome, determined independently. | |
| Length of follow-up | Moderate risk |
| FindingThree years. Later conversions are not captured, which classifies as non-converters patients who will go on to convert. Note also that fewer than 10% of conversions occurred without biological evidence supporting Alzheimer’s disease, which limits the contribution of non-specific profiles. | |
| Independence | No interests declared |
| FindingFrench public and charitable funding, with no role in the conduct of the study. The authors declare no conflict of interest related to this manuscript. The assay platform is a commercial product, which remains a piece of context worth knowing even in the absence of any declared interest. | |
Level of evidence
Oxford level of evidence 2b: a prospective cohort with complete three-year follow-up, but an exploratory analysis deriving prognostic models, without validation in an independent sample. Confidence is high in the central observation, which is an internal comparison within the same sample and with the same assay methods: phosphorylated tau discriminates amyloid pathology better than it discriminates the risk of conversion. It is low for the precise prognostic value of the combination identified, because of the width of the confidence interval, the absence of a defined cut-off and the absence of external validation. It is nil for the idea that acting on these markers would change the course of the disease, a question the study does not address.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ This changes nothing about what I prescribe tomorrow: in this paper, the two markers that make the difference come from a research platform, not from a routine test. But it changes what I tell the patient. The amyloid work-up says the disease is there. It does not say when it will show. ”
What this means in practice: when giving the diagnosis, explicitly separate the presence of the pathology from its pace. The two pieces of information do not come from the same markers, and the publication reports no routine test that provides the second.
What you can do with this
- Do not turn an amyloid-positive result into a prognosis. Saying so to the patient in those terms avoids a disclosure heavier than the test allows.
- Keep in mind the order of magnitude of the only reference point that can be transferred today: in this cohort, a positive amyloid status on cerebrospinal fluid is associated with a hazard ratio for conversion of 3.2 (1.9 to 5.0) over three years. That is real information, it is not a date.
- When a report includes neurodegeneration markers, total tau in particular, read them as the information closest to the pace of progression, without giving them a decision threshold that does not exist.
- Wait for validation in an independent cohort before granting prognostic value to any combination of markers, including the one identified here.
- Be ready to answer the question “how much time do I have left” with what is established: none of the tests examined in this publication, including those already used in routine practice, answers that question properly, and this study documents precisely why.
Frequently asked questions
Should clinicians stop ordering p-tau217 in mild cognitive impairment?
No. It remains highly accurate for what it is used for, detecting amyloid pathology. The study does not call that use into question, it marks out a different one: with an area under the curve of 0.73 in cerebrospinal fluid and 0.72 in plasma for conversion, it is not enough to support an individual prognosis.
Can this biomarker signature be used in routine practice?
The multiplex platform used here is a research tool, and the publication does not report routine use of the signature identified. One point is nonetheless encouraging for what comes next: the authors obtain the same gain by replacing the multiplex tau measurement with a routine ELISA. That substitution concerns tau only: for NPTX2 and IL-16, the publication does not document a routine assay.
Does a hazard ratio of 39.8 mean a forty-fold higher risk of dementia?
It should not be read that way. It compares patients in the third tertile of the score with those in the first, leaving out the middle tertile and with no defined cut-off, and its confidence interval runs from 9.6 to 165.2. The authors themselves state that these values cannot be directly extrapolated to clinical use. What can be retained is the direction of the association, not its size.
Do these findings apply to patients seen in a memory clinic?
The cohort is French and multicentre, which is uncommon and makes transfer to French practice easier in particular. The population is older, mean age 77.8 years, which is a good match for memory clinic patients. The limit lies in the technique and in the absence of independent validation, not in the population.
Annotated bibliography
Source study. Delaby C, Schraen-Maschke S, Paquet C, Blanc F, Vidal JS, Hirtz C, Assou S, Allinquant B, Bombois S, Gabelle A, Hanon O, Lehmann S. Beyond amyloid and tau: synaptic and neurodegenerative biomarkers shape MCI progression. Molecular Psychiatry, 2026, published online on 24 June 2026. DOI 10.1038/s41380-026-03708-y. Open access article. BALTAZAR cohort, ClinicalTrials.gov registration NCT01315639.