Accuracy of blood tests for specific tau and amyloid proteins in detecting Alzheimer’s disease using an automated platform

We included all consecutive individuals who underwent lumbar puncture for the analysis of AD CSF biomarkers assessed at the Sant Pau Memory Unit (SPIN cohort, Barcelona, Spain) as part of their diagnostic work-up [35] between January 2021 and December 2021 (shown in flowchart in Supplementary Material (Fig. 1). The study was approved by the Sant Pau Ethics Committee (Protocol code: EC/22/202/6880) following the standards for medical research in humans recommended by the Declaration of Helsinki. All participants or their legally authorized representative gave written informed consent to participate in biomarkers research studies.

At the time of CSF and plasma acquisition, participants had a diagnosis of dementia, mild cognitive impairment (MCI), or were cognitively unimpaired (CU). The clinical diagnosis was established after a thorough neurological and neuropsychological evaluation [35]. A more extensive description about how cognitively unimpaired participants were recruited and classified is provided in Supplementary Material (Text 1 & Fig. 1). To assess the impact of vascular risk factors and comorbidities, we collected information about the presence of high blood pressure, dyslipidemia, diabetes, obstructive sleep apnea and history of stroke. Participants were also classified according to the estimated glomerular filtrate rate (eGFR) in different stages (1–5) of chronic kidney disease (CKD) using CKD-EPI formula.

After a full evaluation that included analysis of AD CSF biomarkers, participants were classified according to their etiologic diagnosis, as Alzheimer disease (AD), other neurodegenerative diseases (OtherDem), not neurodegenerative diseases (OtherNotDeg) or CU. A proportion of participants' diagnosis was classified as “uncertain” as they had an unclear etiological diagnosis after a full initial evaluation and required clinical follow-up.

Cognitively unimpaired participants were patients with no evidence of cognitive impairment after a thorough neuropsychological evaluation and healthy volunteers interested in research.

Blood samples were collected in EDTA-K2 tubes and subsequently centrifuged (2000 rpm × 10 min, 4ºC) within 2 h after extraction. Plasma was aliquoted and stored at -80ºC until analysis. CSF samples were obtained through lumbar puncture, and were also centrifuged, aliquoted and stored at -80ºC until analysis. Blood and CSF samples were collected simultaneously. Full protocol for CSF and blood sample collection in our center has been previously reported [35, 36].

All plasma samples were measured in the Lumipulse fully-automated platform G600II using commercially available kits (Fujirebio Europe, Ghent, Belgium) for pTau₁₈₁, Aβ_1-42 and Aβ_1-40 between July and August 2022 with the same lot of reagents. Plasma pTau₂₁₇ was analyzed between August and September 2023 in another aliquot of the same samples using a novel assay recently developed by Fujirebio. On the day of the analysis, plasma samples were brought to room temperature, mixed thoroughly, centrifuged for 5 min at 2000 g, and subsequently transferred to specific cuvettes for analysis in the Lumipulse platform.

CSF markers Aβ_1–42, Aβ_1–40, pTau₁₈₁ and tTau were used in the diagnostic assessment of patients and measured in routine runs scheduled twice a month throughout 2021 following previously reported methods [37]. According to CSF markers, all participants were classified as amyloid positive (A + , CSF Aβ_1–42/Aβ_1–40 < 0.062) or negative (A-), and as tau positive (T + , pTau₁₈₁ > 63 pg/mL) or negative (T-). Validation of these cutoff values has been described elsewhere [37]. To validate that our main findings would not be overly influenced by analytical error in CSF measurements, we performed additional sensitivity analyses excluding participants ± 10% of the CSF cutoff points (Supplementary Material, Figs. 1–3).

DNA was extracted from full blood using standard procedures, and APOE was genotyped following previously reported methods [35]. Briefly, direct DNA sequencing of exon 4 was performed routinely for all participants in the SPIN cohort, followed by visual analysis of the resulting electropherogram to identify the two coding polymorphisms that encode the three possible APOE isoforms.

Data normality was assessed with the Shapiro–Wilk test. Non-normally distributed variables were log-transformed when necessary. Kruskal–Wallis and Wilcoxon rank sum test were used in continuous variables that were not normal distributed. Linear regression models and ANCOVA adjusted by age and sex were performed for group comparison. We used Chi Square test to assess differences in categorical variables and Fisher’s exact test in group comparisons with small number of observations, and Spearman test to assess the correlation between plasma and CSF markers.

Diagnostic accuracy of plasma biomarkers was assessed through receiver operating characteristic (ROC) analysis. We calculated the areas under the curve (AUC) of individual markers and that of logistic regression models that combined them with each other and with clinical variables. A basic model that included Age, Sex and APOEε4 status was used as a reference to assess the added diagnostic value of plasma markers. We compared the accuracy of individual markers and regression models using DeLong's test adjusted by multiple comparisons using Bonferroni method. We evaluated the sensitivity, specificity, and Youden’s J index of a range of cutoffs to discriminate A + from A- participants. We analyzed the concordance of pTau₂₁₇ with CSF amyloidosis. We followed a previously reported approach [38] to stratify our cohort in low, medium, and high risk of having CSF amyloidosis. Using predictive models, according to the risk of the participants of being A + , we selected conservative (97.5% sensitivity/specificity) and more lenient (95% sens/spec) cutoffs. We bootstrapped to assess the cutoff robustness. All tests were performed in R statistical software version 4.2.1. Alpha threshold was set at 0.05 for all analysis.

We included 290 participants who were syndromically classified as cognitively unimpaired (CU, n = 66), having mild cognitive impairment (MCI, n = 130) or a clinical diagnosis of dementia (n = 94). Table 1 shows the etiologic diagnoses in each category, main demographic characteristics, and biomarker measures in each group. CU participants were younger than those with MCI (p < 0.001) and those with dementia (p < 0.001). There were more female participants (62%). The proportion of A + , A + T + and APOEε4 positive increased according to the clinical stage. More extensive demographics details, including stratification by clinical diagnosis or A status, can be found in Supplementary Material (Tables 1 & 2).

All plasma measures for pTau₂₁₇, pTau₁₈₁, Aβ_1–42 and Aβ_1–40 were above their lower limit of quantification. The plasma concentration ranges in the study were 0.03 to 2.84 for pTau₂₁₇, 0.92 to 9.63 pg/mL for pTau₁₈₁, 14.06 to 50.84 pg/mL for Aβ_1–42 and 180.41 to 569.89 pg/mL for Aβ_1–40. Inter-assay coefficients of variation are shown in Supplementary Material (Table 3).

As per inclusion criteria, all participants had CSF biomarkers measures, and we explored the correlation between both matrices. The correlation between plasma and CSF was moderate for pTau₁₈₁ (Rho = 0.66, p < 0.001) and low for Aβ_1–42 (Rho = 0.26, p = 0.007), and Aβ_1–40 (Rho 0.11, p = 0.06). When using ratios, the correlation was moderate for Aβ_1–42/Aβ_1–40 (Rho = 0.69, p < 0.001). The correlation between plasma pTau₂₁₇ and CSF pTau₁₈₁ was high (Rho = 0.75, p < 0.001). Both plasma pTau₂₁₇ and pTau₁₈₁ correlated with age (Rho = 0.42, p < 0.001 and Rho = 0.45, p < 0.001 respectively). Detailed correlations within clinical subgroups and partial correlations are shown in Supplementary Material (Figs. 4 and 5).

We assessed the differences in plasma biomarkers between CSF amyloid positive and amyloid negative individuals considering other variables in a multivariate model. We studied the effect of age, sex, APOE status (APOE ε4 +), renal function measured by the estimated glomerular filtration rate (eGFR), vascular risk factors (presence of at least one of the following: high blood pressure, diabetes mellitus, dyslipidemia, history of stroke, obstructive sleep apnea with CPAP) and clinical status (CU, MCI and Dementia). As shown in Fig. 1, the log transformed multivariate model confirmed that the A + group had higher plasma concentrations of pTau₂₁₇ (fold-change 4.42, p < 0.001) and pTau₁₈₁ (fold-change 1.72, p < 0.001) compared to the A- group. Similar results were seen using the ratios pTau₂₁₇/Aβ_1–42 and Aβ_1–42/Aβ_1–40. The plasma pTau₂₁₇/Aβ_1–42 ratio was higher in A + compared to A- (fold-change 4.08, p < 0.001). The plasma Aβ_1–42/Aβ_1–40 ratio was lower in A + compared to A- (fold-change 0.86, p < 0.001). We found lower plasma concentrations of Aβ_1–42 in A + compared to the A- group (fold-change 0.92, p < 0.001) and no differences in Aβ_1–40 concentrations between A status (not shown). To assess the influence of the values close to CSF Aβ_1–42/Aβ_1–40 cutoff, we conducted a sensitivity analysis by excluding participants with values close to the cutoff point (± 10%) (Supplementary Material, Figs. 1– 3). We found no significant influence of these observations in our analysis.

We assessed whether log transformed plasma pTau₂₁₇, pTau₁₈₁, Aβ_1-42 and Aβ_1-40 were affected by other variables in the multivariate model. As seen in Fig. 2, amyloid positivity was the variable with the largest effect on all plasma markers. We also observed that decreased renal function was associated with higher concentrations of pTau₂₁₇ (p = 0.019) and pTau₁₈₁ (p < 0.001) and higher Aβ_1–42/Aβ_1–40 ratio (p < 0.001). Male sex was associated with higher pTau₂₁₇ (p = 0.033) and pTau₁₈₁ (p = 0.0009), and age was associated with lower Aβ_1–42/Aβ_1–40 ratio_. Our model had and adjusted R² value of 0.62 for pTau₂₁₇, 0.52 for pTau₁₈₁ and 0.48 for Aβ_1–42/Aβ_1–40. All the presented coefficients were obtained from the model in which all the listed variables were included. Complete forest plot shown in Supplementary Material (Fig. 6).

To further investigate the association of pTau₂₁₇ and pTau₁₈₁ with renal function, we performed a subanalysis stratifying by estimated glomerular filtration rate. We found that pTau₁₈₁ concentration in plasma was higher as renal function decreased (< 60 vs 60-90 mL/min/1.73m² and < 60 vs. > 90 mL/min/1.73m², p < 0.001). Aβ_1–42 and Aβ_1–40 concentrations in plasma were also higher as renal function decreased (p < 0.001). We also observed marginally significant differences in pTau₂₁₇ concentrations in plasma samples from patients with low renal function (< 60 mL/min/1.73m²) compared to those with normal renal function (> 90 mL/min/1.73m², p = 0.047). However, those differences were lost when using the Aβ_1–42/Aβ_1–40 or the pTau₂₁₇/Aβ_1–42 ratios.

In the whole sample, the AUC to discriminate A + from A- participants were 0.94 (95% Cl 0.92–0.97) for pTau₂₁₇, and 0.88 (95% CI 0.84–0.92) for both pTau₁₈₁ and Aβ_1–42/Aβ_1–40 (Fig. 3). The diagnostic accuracy of pTau₂₁₇ to detect amyloid positivity was not outperformed by any other individual plasma biomarker, their ratios or their combinations. Aβ_1–42 and Aβ_1–40 individually had poor diagnostic accuracy, yielding AUCs below 0.70. Detailed two-by-two comparisons can be found as Supplementary Table 4. Sensitivity, specificity and Youden indices yielded by individual plasma markers are shown in Supplementary Fig. 7.

The diagnostic performance of pTau₂₁₇ was also high across different clinical categories. In the MCI group, the accuracy of pTau₂₁₇ (AUC = 0.92, 95% CI 0.87–0.98) was not significantly different than that pTau₁₈₁ (AUC = 0.86, 95% CI 0.79–0.93) or that of the basic model with Age, Sex and APOE ε4 (AUC = 0.81, 95% CI 0.74–0.88). However, the addition of pTau₂₁₇ to Age, Sex and APOE ε4 increased the accuracy of the model significantly from 0.81 (95% CI 0.74–0.88) to 0.94 (95% CI 0.9–0.99). pTau₂₁₇ also showed very high accuracies (AUC = 0.97; 95% CI 0.94–1.00) to discriminate A + patients with a diagnosis of AD from different A- clinical groups (CN, other dementias and other not degenerative, Supplementary Fig. 8). The performance of plasma biomarkers to discriminate A + T + from A-T- participants yielded similar results (Supplementary Fig. 9).

Table 2 shows the accuracy of different thresholds for pTau₂₁₇, pTau₁₈₁ and Aβ_1–42/Aβ_1–40 to detect amyloid positivity with a sensitivity and specificity of 97.5%, 95% and 90%. As pTau₂₁₇ was the individual biomarker with highest diagnostic accuracy, we assessed the performance of selected cutoffs for this marker in different clinical groups, stratifying by decade. We found that the cutoff that had a global sensitivity of 95% (0.186 pg/mL) yielded accuracies above 84% across all decades and clinical groups when the amyloidosis prevalence was above 20%. In groups with lower amyloidosis prevalence (i.e. CN) the cutoff with specificity of 95% (0.388 pg/mL) showed higher accuracies. We observed that the cutoff that had a specificity of 95% showed a progressive decrease in accuracy with age in the MCI group (accuracy 84% in 60–69, 75% in 70–79 and 62% over 80 years). The negative predictive value of this cutoff was low over 70 years (58%). Detailed information on the accuracy of cutoffs in distinct clinical groups and decades can be found in Supplementary Material (tables 5, 6, 7 & 8).

We conducted a supervised decision tree analysis to determine the potential of various biomarkers and demographic factors in correctly identifying individuals with amyloidosis. This analysis incorporated plasma pTau₂₁₇, pTau₁₈₁, ratio Aβ_1–42/ Aβ_1–40, Age, Sex, APOE ε4 allele presence, and clinical diagnosis group (cognitively unimpaired, mild cognitive impairment, and dementia). The most effective discriminators for amyloidosis were plasma pTau₂₁₇ followed by the Aβ_1–42/ Aβ_1–40 ratio when pTau₂₁₇ was high (Supplementary Fig. 10). Other variables were deemed less critical for amyloidosis detection and thus excluded from the decision tree. This algorithm exhibited a misclassification rate of 7.4%, with a sensitivity of 91%, specificity of 94%, overall accuracy of 93%, positive predictive value (PPV) of 94%, and negative predictive value (NPV) of 91%, accompanied by a false-negative rate (FNR) of 9.4% and a false-positive rate (FPR) of 5.6%.

We assessed the predictive capability of pTau₂₁₇ to properly classify participants with amyloidosis in our dataset, defined as CSF Aβ_1–42/Aβ_1–40 < 0.062. As the model that combined pTau₂₁₇ with Age, Sex and APOEε4 did not perform better than pTau₂₁₇ alone in any comparison, we chose the simplest predictive model with pTau₂₁₇. We bootstrapped to obtain robust predictive cuttoffs. We found no differences between the initial prediction and the mean of the 1000 predictive iterations, with a perfect correlation between them (Rho = 1, p < 0.001) thus reinforcing the robustness of our initial predictions.

To facilitate the clinical implementation of plasma biomarkers while ensuring accuracy, we applied a two-threshold approach to classify participants into three groups, those with high, medium, and low likelihood of being CSF amyloid positive. Following this approach, those with a medium risk would benefit from a confirmation with gold standard tests like CSF or Amyloid PET. Figure 4 shows the thresholds in two different scenarios based on two levels of restrictiveness in sensitivity and specificity (97.5% and 95%). Using a highly accurate combination of cutoffs (one for 97.5% Sens and another for 97.5% Spec), only 41.9% of patients in the whole sample would require an additional test, with a global misclassification rate of 4.2%. Using less restrictive cutoffs (95% Sens/Spec), the proportion could be reduced to 19% with a global misclassification rate of 6%. Similar results were found in the subgroup of patients with CU, MCI and Dementia.

We finally assessed the robustness of those cutoffs in the whole sample, considering the increase of prevalence of amyloidosis with age in our population. In Fig. 5, we show the NPV, PPV, and global accuracy of pTau₂₁₇ to detect A positivity by decades using cutoff combinations with 95% and 97.5% sensitivity and specificity in different clinical groups. We found high global accuracy (75%-100%) of the two-threshold application in all the scenarios, with variations of PPV and NPV according to CSF amyloid positive prevalence in our sample.

The ethics committee of Hospital Sant Pau approved all procedures included in this study following the standards for medical research in humans recommended by the Declaration of Helsinki. All participants or their legally authorised representative gave written informed consent before enrolment in the study.

All authors revised the manuscript for content and provided critical feedback.

Daniel Alcolea is employed by Hospital de la Santa Creu i Sant Pau and received research grants from Pla Estratègic de Recerca i Innovació en Salut (PERIS SLT006/17/125), and from Instituto de Salud Carlos III (PI18/00435 and INT19/00016). He participated in advisory boards from Fujirebio-Europe and Roche Diagnostics and received speaker honoraria from Fujirebio-Europe, Roche Diagnostics, Nutricia, Zambon S.A.U., Esteve, and from Krka Farmacéutica S.L.

Javier Arranz is employed by Biomedical Research Institute Sant Pau. He is funded by a “Rio Hortega” research grant from the Institute of Health Carlos III. Declarations of interest: none.

Nuole Zhu is employed by Biomedical Research Institute Sant Pau. He is funded by a “Rio Hortega” research grant from the Institute of Health Carlos III. Declarations of interest: none.

Sara Rubio-Guerra is employed by Hospital de la Santa Creu i Sant Pau. Declarations of interest: none.

Iñigo Rodríguez-Baz is employed by Biomedical Research Institute Sant Pau. He is funded by a “Rio Hortega” research grant from the Institute of Health Carlos III. Declarations of interest: none.

Rosa Ferrer is employed by Hospital de la Santa Creu i Sant Pau. Declarations of interest: none.

María Carmona-Iragui is employed by Hospital de la Santa Creu i Sant Pau. Declarations of interest: none.

Isabel Barroeta is employed by Hospital de la Santa Creu i Sant Pau. Declarations of interest: none.

Dr. Illán-Gala is a senior Atlantic Fellows for Equity in Brain Health at the Global Brain Health Institute (GBHI), and is supported with funding from GBHI, Alzheimer’s Association, and Alzheimer’s Society (GBHI ALZ UK-21–720973 and AACSF-21–850193). Dr Illán-Gala was also supported by the Juan Rodés Contract (JR20/0018) and Fondo de Investigaciones Sanitario, (PI21/00791) from Instituto de Salud Carlos III.

Ignacio Illán-Gala reported receiving personal fees from Nutricia, Esteve, UCB, and Neuraxpharm Spain outside the submitted work.

Miguel Santos-Santos is employed by Hospital de la Santa Creu i Sant Pau. His research is supported by funding from the Spanish Institute of Health Carlos III (Juan Rodés contract JR18-00018; Fondo de investigación sanitaria grant PI19/00882), the Alzheimer’s Association clinician scientist fellowship (AACSF-22–972945), and the National Institutes of Health (R01AG080470).

Juan Fortea is employed by Hospital de la Santa Creu i Sant Pau and received research grants from Institute of Health Carlos III, National Institutes of Health, Fundació La Marató de TV3, and Pla Estratègic de Recerca i Innovació en Salut (PERIS). Dr. Fortea has served as a consultant for Novartis and Lundbeck, has received honoraria for lectures from Roche, NovoNordisk, Esteve and Biogen and served at advisory boards for AC Immune, Zambon and Lundbeck.

Alberto Lleó is employed by Hospital de la Santa Creu i Sant Pau and received research grants from CIBERNED, Institute of Health Carlos III, Generalitat de Catalunya (PERIS and AGAUR) and Fundación Tatiana and BBVA. He participated in advisory boards from Biogen, Eisai, Fujirebio-Europe, Novartis, NovoNordisk, Nutricia, Otsuka Pharmaceutical, and Zambón, and received speaker honoraria from Lilly, Biogen, KRKA and Zambon.

Mireia Tondo is employed by Hospital de la Santa Creu i Sant Pau and has received research grants from Instituto de Salud Carlos III (PI18/00164; PI21/00140) and has served as consultant in Araclon.

Raw anonymized data and code for statistical analysis are available upon reasonable request. All requests should be sent to the corresponding author detailing the study hypothesis and statistical analysis plan. The steering committee of this study will decide whether data/code sharing is appropriate based on the novelty and scientific rigor of the proposal. All applicants will be asked to sign a data access agreement.