In this cross-sectional study, we included patients with MCI from two independent cohorts, based on complete availability of plasma p-tau217, CSF Aβ42/Aβ40, Aβ-PET and APOE ε4 genotyping. Our model training cohort, BioFINDER-1 (NCT01208675), recruited patients between January 2010 and January 2015 and our validation cohort, BioFINDER-2 (NCT03174938), started recruitment in May 2017. In both cohorts, the patients were consecutively recruited from secondary memory clinics in the southern part of Sweden, where most of the study participants were referred directly from primary care, as described below. In Supplementary Information, we demonstrate that the included BioFINDER-1 and BioFINDER-2 populations (that is, with full biomarker availability) were similar to the nonincluded participants due to lack of data for one or more biomarkers (Supplementary Tables 7 and 8)10,45,46.
The BioFINDER-1 inclusion criteria for enrolling participants with subjective cognitive decline or MCI were as follows: (1) having been referred owing to cognitive symptoms experienced by the participant or perceived by an informant; (2) age between 60 and 80 years; (3) MMSE score of 24–30 points at the baseline visit; (4) do not fulfill the criteria for any dementia; and (5) fluency in Swedish. The exclusion criteria were as follows: (1) a systemic illness or organ failure of substantial severity that would hinder participation in the study; (2) current substance misuse or alcohol abuse; (3) refusal of neuropsychological assessment or lumbar puncture; and (4) cognitive impairment at baseline that could, with high confidence, be explained by another condition or disease, such as major cerebral hemorrhage, normal pressure hydrocephalus, brain tumor, brain infection, epilepsy, multiple sclerosis, psychotic disorders, severe depression or ongoing use of medication that causes a reduction in cognitive functioning (such as high-dose benzodiazepines). The clinical diagnosis was delivered at baseline based on an extensive battery of neuropsychological tests evaluating verbal and episodic memory, visuospatial ability and attention/executive domains, as described in detail elsewhere46. In the whole BioFINDER-1 study, for which enrollment was completed, a thorough analysis on referral origin had been previously conducted as described by Petrazzuoli et al.46. Most of the BioFINDER-1 patients (80.8%) were referred from primary care, whereas 12.5% of referrals were made by other specialist clinics and 6.7% of patients were self-referrals46. The inclusion criteria for recruitment of patients with MCI for BioFINDER-2 were as follows: (1) aged 40–100 years; (2) referred to the memory clinics due to cognitive symptoms; (3) MMSE score of 24–30 points; (4) did not fulfill the criteria for any dementia (major neurocognitive disorder) according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edn (DSM-IV)47; and (5) fluent in Swedish. The BioFINDER-2 study also recruits patients who are CU, patients with AD dementia and patients with non-AD neurodegenerative conditions, and its general exclusion criteria were as follows: (1) unstable systemic illness that makes it difficult to participate in the study; (2) current alcohol or substance misuse; and (3) refusing lumbar puncture, MRI or PET. Out of the 212 MCI-included participants from BioFINDER-2 with readily available referral data, most were referred from primary care (n = 179; 84.4%), followed by hospital referrals (n = 31; 14.6%) and self-referrals (n = 2; 0.9%).
In both cohorts, a clinical diagnosis of MCI was made for those patients who did not meet the criteria for dementia (major cognitive disorder as in DSM-V48) but have lower scores than −1.5 s.d. in at least one cognitive domain such as memory, verbal, attention/executive or visuospatial function. In BioFINDER-1, a senior neuropsychologist made the diagnosis after a thorough neuropsychological battery to make this determination, as previously described46. In BioFINDER-2, the MCI diagnosis was based on a score <−1.5 z-scores in any cognitive domain, based on regression normative scores accounting for age, education and test performance in Aβ-negative controls49. The z-scores for each cognitive domain were calculated by averaging the z-scores of relevant tests, with further details on the derivation of such normative equations available elsewhere50,51. The domains included attention/executive function, verbal ability, memory and visuospatial function, and the tests used included Trail Making Test A, Trail Making Test B, Symbol Digit Modalities Test, verbal fluency animals, 15-word short version of the Boston Naming Test, 10-word delayed recall from the Alzheimer’s Disease Assessment Scale, and incomplete letters and cube analysis from the Visual Object and Space Perception battery.
In BioFINDER-1 and BioFINDER-2, we also evaluated the presence of comorbidities in the study population, evaluating for history of cardiovascular disease, diabetes or dyslipidemia36. Participants were considered to have cardiovascular disease if they presented with a history of either ischemic heart disease or hypertension, or if they were on anti-hypertensive/cardioprotective therapy. A history of dyslipidemia was considered when patients had such a diagnosis previously made or if they were on lipid-lowering therapy. Participants were considered to have CKD based on estimated glomerular filtration rate <60 ml min−1 per 1.73 m2, accepted as a functional criterion for CKD52.
In a secondary analysis, we included a subset of 84 cognitively impaired participants with available plasma p-tau217, CSF Aβ42/Aβ40, APOE ε4 genotype and Aβ-PET from the TRIAD cohort, recruited from a tertiary care memory clinic specializing in the diagnosis and management of neurodegenerative diseases44. All clinical diagnoses were made blinded to biomarker results. All participants had clinical assessments including Clinical Dementia Rating (CDR), MMSE and cerebrovascular disease risk using the Hachinski Ischemic Scale. Participants were excluded from the present study if they had systemic conditions that were not adequately controlled through a stable medication regimen. Other exclusion criteria were active substance abuse, recent head trauma, recent major surgery or MRI/PET safety contraindications. The included participants had MCI as defined based on a CDR of 0.5 and an MMSE between 24 and 30 (n = 63), and patients with dementia who had CDR of ≤1 (n = 21).
All BioFINDER and TRIAD patients gave their written informed consent to participate in the study and participation was voluntary. The BioFINDER studies were approved by the Ethical Review Board in Lund, Sweden, which is part of the Swedish Ethical Review Authority. TRIAD was approved by the Montreal Neurological Institute PET working committee and the Douglas Mental Health University Institute Research Ethics Board.
Imaging and fluid biomarkers in BioFINDER-1 and BioFINDER-2
Aβ-PET was quantified using [18F]flutemetamol on a Philips Gemini TF 16 scanner in BioFINDER-1 and a digital GE Discovery MI scanner in BioFINDER-2. Scans were acquired 90–110 min after the injection of ~185 MBq of [18F]flutemetamol. The standardized uptake value ratio (SUVr) was obtained by normalizing the neocortical composite values to the whole cerebellum as a reference region. FreeSurfer (v.5.3) parcellation of the T1-weighted MR scan was used to transform the PET data to the participants’ native T1 space, so as to obtain mean regional SUVr values in predefined neocortical regions of interest, including prefrontal, lateral temporal, parietal, anterior cingulate and posterior cingulate/precuneus53. Aβ-PET data were binarized into normal and abnormal using cutoffs derived from Gaussian mixture modeling (GMM), with a threshold of ≥1.138 for BioFINDER-1 and ≥1.033 for BioFINDER-2.
CSF samples were collected and described based on previously described protocols54. CSF Aβ42/40 was measured using the fully automated Roche Elecsys NeuroTool Kit for the entirety of BioFINDER-1 and for 75% (n = 161) of BioFINDER-2 participants55,56. Abnormal CSF status was defined based on previously derived cutoffs determined using GMM, with a threshold of ≤0.066 for BioFINDER-1 and ≤0.080 for BioFINDER-2 (the higher cutoff in the latter study is due to use of LoBind tubes in BioFINDER-2, according to more recent protocols that prevent Aβ42 from binding to the tube walls57,58). For the 25% (n = 51) of BioFINDER-2 participants for whom the Elecsys measurements were not available, an abnormal CSF Aβ42/40 status was determined using the FDA-approved Lumipulse G assay, with a GMM-derived threshold of ≤0.06 (ref. 59). All CSF Aβ42/40 measurements were performed at the Clinical Neurochemistry Laboratory, Sahlgrenska Academy.
EDTA plasma samples were collected, handled and processed as previously described10,45. Plasma p-tau217 was quantified using the Mesoscale Discovery platform with an assay developed by Lilly Research Laboratories. Biotinylated-IBA493 was used as a capture antibody and SULFO-TAG-4G10-E2 (anti-tau) as the detector antibody, with sample and antibody dilution at 1:2, as previously described23. APOE ε4 was genotyped using a TaqMan allelic discrimination assay60.
Imaging and fluid biomarkers in TRIAD
Individuals were evaluated with plasma p-tau217, CSF Aβ42/40 and amyloid-PET using [18F]AZD4694. Plasma concentrations of p-tau217 were measured using a Simoa assay developed by Janssen R&D by scientists blinded to clinical, demographic and biomarker information as described previously16, using the PT3 antibody as capture and HT43 as detector, and samples and detector were diluted 1:2. CSF concentrations of Aβ42/40 were quantified using the fully automated Lumipulse G1200 instrument (Fujirebio), with an Aβ-positivity threshold of 0.068, by scientists blinded to clinical and biomarker information as described previously61. A [18F]AZD4694 amyloid-PET-positivity threshold of 1.55 was employed (centiloid ≥ 24), validated based on GMM, CSF thresholds and visual assessments62. Blood and CSF collections took place on the same day.
Statistics and reproducibility
First, we developed a logistic regression model using Aβ-PET status as the outcome with plasma p-tau217, age and APOE ε4 status as predictors in BioFINDER-1. Age and APOE ε4 were considered as predictors due to their inclusion in recently published, blood-based biomarker models and due to their well-described associations with Aβ positivity23,24,25,40,41. Plasma p-tau217 was log-transformed due to its skewed distribution and age was modeled with a linear term. Variables such as cognitive tests may be of more relevance to prognostic models (that is, predicting cognitive worsening) than in diagnostic models for Aβ positivity, given the poor association between Aβ load and symptoms63. To examine whether a simpler model would be preferred to this full model with age, APOE ε4 and p-tau217, backward variable deletion was performed during bootstrapped internal validation (n = 1,000), with the stopping criterion set at α = 0.157, recommended for model development scenarios such as ours64. The model most frequently chosen during this procedure was externally validated in BioFINDER-2. For model performance, we used the receiver operating characteristic’s AUC. In BioFINDER-1, the optimism-corrected AUC is reported, a metric recommended to account for overfitting-related optimism at model development65. Model calibration at external validation was assessed visually66. For goodness of fit, we report Nagelkerke’s pseudo-coefficient of determination (R2) and Akaike’s information criterion65,67.
Based on the blood biomarker, model-derived probabilities of Aβ-PET positivity and further testing with CSF Aβ42/Aβ40, we evaluated a two-step diagnostic workflow. In the first step, different thresholding strategies were explored to classify participants into low-, intermediate- and high-risk groups based on the plasma p-tau217 model-derived probabilities of Aβ-PET positivity. These strategies were defined based on lower probability thresholds with 90%, 95% and 97.5% sensitivity and higher probability thresholds with 90%, 95% and 97.5% specificity, with the same sensitivity and specificity always being tested together (for example, 90% sensitivity with 90% specificity). For each of the strategies, we calculated the prevalence of Aβ-PET negativity in the low-risk group along with the prevalence of Aβ-PET positivity in the high-risk group. For the second step, we tested the scenario in which further testing would be carried out with CSF Aβ42/Aβ40 measurements only in intermediate-risk participants from the first step. In this group, we reported the concordance between CSF and Aβ-PET status. Furthermore, we computed the overall workflow accuracy, represented by the proportion of correct Aβ-PET status classifications in both plasma and CSF steps, as well as the reduction in number of further confirmatory tests by the blood-biomarker-based risk stratification. In a secondary exploratory analysis, we further evaluated the robustness and generalizability of the two-step workflow using z-scored plasma p-tau217 values. The z-scores were obtained based on the distribution of this reference CU Aβ-negative sample as follows: (plasma p-tau concentration − mean p-tau concentration in CU Aβ negatives)/(s.d. of plasma p-tau concentration in CU Aβ-negatives). In BioFINDER-1, z-scored plasma p-tau217 (Lilly) values were obtained based on 283 CU Aβ-negative older adults with a mean (s.d.) plasma p-tau217 concentration of 0.153 (0.077) pg ml−1. In BioFINDER-2, based on 316 CU Aβ-negative participants, the mean (s.d.) concentrations were 0.156 (0.064) pg ml−1 for plasma p-tau217 (Lilly). In TRIAD, z-scores were calculated based on 111 Aβ-negative CU older adults with a mean (s.d.) plasma p-tau217 (Janssen) concentration of 0.052 (0.026) pg ml−1. Such a procedure enables application of the risk-prediction model for different plasma p-tau217 assays, because when z-scored they can be obtained from internal reference samples from clinical chemistry labs and memory clinic services. Briefly, the same original BioFINDER-1 model was re-fitted with z-scored plasma p-tau217 with the Lilly assay. Then, it was validated in two other cohorts: BioFINDER-2, based on z-scored Lilly plasma p-tau217 immunoassay and, in TRIAD, based on z-scored plasma p-tau217 measured with a different p-tau217 immunoassay (Janssen R&D). The whole workflow was re-evaluated for overall accuracy and reduction in the number of advanced tests for all of these secondary analyses, with the same risk thresholds from the original main analysis model. The z-scored model was developed in BioFINDER-1 in the exact same MCI population as that in the main analysis (n = 136). When validating the z-scored model in BioFINDER-2 with z-scored Lilly plasma p-tau217, we evaluated it in the exact same BioFINDER-2 MCI population as used in the main analysis (n = 212). In TRIAD, the z-scored model was applied in the n = 84 patients with cognitive impairment with key demographic characteristics shown in Supplementary Information. Our sample size was based on complete biomarker availability (for plasma, genetic, CSF and imaging data) rather than on statistically predetermined numbers, but our sample size was similar to those reported in previous publications evaluating risk-prediction models in AD23,24,25. When applicable, a two-sided α of 0.05 was used and 95% CIs are reported. No data exclusion was performed. Data collection and analysis were not randomized or performed blind to the experimental groups. All statistical analyses were performed in R v.4.1.1 (www.r-project.org), mainly using the ‘rms’ package68.
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.