Background
Prodromal Alzheimer’s disease (AD) is a common cause of mild cognitive impairment (MCI), but it is very difficult to clinically differentiate it from other causes of MCI [
1]. Generally, between 20 to 40% of MCI patients will progress to AD dementia within a few years [
2,
3]. Improving the identification of patients at greater risk of further cognitive decline is thus important for clinical practice for patients and families, in clinical trials to enroll patients having AD pathology, and in the future for selecting patients for treatment with disease-modifying drugs. The key protein causing AD, beta-amyloid (Aβ) and tau, can be measured either in the cerebrospinal fluid (CSF) with lumbar puncture or in the brain with positron emission tomography (PET). These can be used to help in determining the risk of progression to AD dementia in individuals with MCI [
4‐
7]. However, CSF collection may be regarded as invasive and PET scans are costly and have limited availability, which hampers the use of these methodologies in clinical practice from a global perspective. With the recent advent of blood-based biomarkers, we can now measure a variety of proteins related to AD in a time- and cost-effective manner and investigate how well such markers can inform disease diagnosis and prognosis [
6]. The molecular pathways that can be investigated with plasma biomarkers also now extend beyond Aβ and tau. These include for example neurodegenerative markers such as neurofilament light (NfL) and glial activation biomarkers such as glial fibrillary acidic protein (GFAP) [
6,
8]. Very promising results in non-demented patients suggest that plasma tau phosphorylated at threonine 217 (p-tau217), in combination with cognitive performance and apolipoprotein E (
APOE) genotype, was the best marker to predict conversion to AD dementia within 4 years, with very high accuracy [
1]. These recent results were derived from only two cohorts, and thus, we still need to validate the optimal markers of conversion in other independent cohorts and determine the most consistent results before we can implement such prognostic algorithms in clinical practice globally.
The current study focuses on a subset of amnestic MCI patients previously enrolled in a 3-year clinical trial, which was originally designed to determine the accuracy of [
18F]flutemetamol PET to predict subsequent conversion to dementia [
7]. We now investigated which combinations of key plasma biomarkers and other commonly used and accessible markers of AD were related to progression to AD dementia. First, we studied the accuracy of plasma biomarkers to identify MCI patients who are likely to progress to AD dementia. In this cohort, we quantified four plasma biomarkers: p-tau217, the ratio of Aβ42/Aβ40, as well as NfL and GFAP. Next, we considered whether combining the best performing plasma biomarkers with hippocampal volume,
APOE genotype, and a composite cognitive score would further improve the discrimination between MCI patients who progressed to AD dementia and those who did not.
Methods
Participants
Patients for this study were originally included from a completed clinical trial that aimed at investigating the efficacy of [
18F]flutemetamol Aβ-PET to predict conversion from MCI to probable AD dementia (NCT01028053, 2009-2014). All participants had amnestic MCI based on the Petersen and Morris criteria [
9], a Clinical Dementia Rating (CDR) of 0.5, were 60 years or older, had a Mini-Mental State Examination (MMSE) between 24 and 30, and a score on the Modified Hachinski Ischemic Scale equal or less to 4. The main exclusion criteria were other significant neurological or psychiatric conditions. Participants and trial outcomes have been described in greater details previously [
7]. The present study included a subset of the initial trial sample, namely participants who had available plasma for analysis, resulting in 110 of the original 232 participants.
Outcome
Participants underwent evaluation by trained personnel at each site that consisted of neuropsychological tests as well as the CDR, MMSE, and activities of daily living every 6 months for up to 36 months. After each visit, participant data was reviewed by members of the clinical adjudication committee, who were blinded to the biomarker data, to determine clinical diagnosis. Diagnosis of probable AD was based on the National Institute of Neurological and Communicative Disorders and Stroke–Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria [
10]. For the present study, the clinical outcome was conversion to AD dementia over 36 months. Only those that were classified as probable AD dementia (according to the NINCDS-ADRDA criteria) and had a positive Aβ-PET at baseline were coded as progressors to AD dementia (in accordance with the NIA-AA definition of AD [
11]). Aβ-positivity was defined based on a predefined threshold of 1.56 standardized uptake value ratio from a global [
18F]flutemetamol Aβ-PET region including precuneus, cingulate, frontal, and lateral temporal regions [
7]. Those who remained MCI (
n = 71) and those who were given a clinical diagnosis of dementia but were Aβ-PET negative (
n = 13) were coded as non-progressors to AD dementia, hereafter refer to as non-progressors.
Plasma biomarkers
Plasma p-tau217 and NfL concentrations were measured at Lund University, Sweden, for all participants. P-tau217 was measured using an immunoassay on the Meso-Scale Discovery (MSD) platform developed by Eli Lilly as described previously [
12]. NfL was measured using the commercially available Simoa immunoassay [
13,
14]. Plasma Aβ40, Aβ42, and GFAP concentrations were measured at the Clinical Neurochemistry Laboratory, University of Gothenburg, Sweden, in 80 participants out of 110, those with enough plasma left. These three proteins were measured using the Simoa Human Neurology 4-Plex E (N4PE) assay (Quanterix®, Billerica, MA, USA).
Cognitive tests
Participants underwent different neuropsychological tests as part of the clinical evaluations. For this study, given the smaller sample size, we focused on a composite measure focusing on cognitive domains affected early in AD, rather than on multiple individual tests. Our measure of interest was a modified version of the Preclinical Alzheimer’s Cognitive Composite 5 (PACC) score [
15]. The tests included in the modified PACC (mPACC) were the MMSE, Logical Memory Scale II delayed recall, Digit Symbol Substitution Test, Category Fluency of animals, and vegetables (the sum of both categories formed the Category Fluency score). The original PACC includes two measures of memory recall (Logical Memory and the Free and Cued Selective Reminding Test); however, because only one was available, Logical Memory delayed recall was given twice the weight to maintain the same proportion of memory as in the original composite score as done previously [
16]. All tests were
z-scored based on the current sample and then averaged to generate the mPACC used in statistical analyses.
Other predictors
In line with the original clinical trial and considering important factors related to AD etiology, we also included
APOE genotype and hippocampal volume in analyses.
APOE genotype was available for 100 out or 110 participants. People with at least one
ε4 allele were considered
APOE4 carriers. Structural T1-weighted magnetic resonance imaging was also acquired at baseline from which the hippocampus was segmented using a local, patch- based label fusion approach [
17]. Hippocampal volume was then adjusted for total intracranial volume using a scaling factor related to the difference between individual subject and MNI152 template space. More details have been described in Wolk et al. [
7].
Statistical analysis
All analyses were performed using R version 4.0.5. Demographics, plasma biomarkers, and markers of interest were compared between patients with MCI who progressed to AD dementia vs. those who did not progress to AD dementia using
t-test or chi-square. The main analyses were then logistic regressions to determine which combinations of markers best discriminated progressors to AD dementia from non-progressors. Variables of interest were
z-scored prior to the logistic regressions, so that odds ratio between variables and models are easily comparable. We used the Multi-Model-Inference R package version 1.43.17 that generates models with the best combinations of biomarker and the pROC package version 1.17.0.1 to compare them to one another. Models were ranked based on model fit using corrected Akaike Information Criteria (AICc), appropriate for smaller sample sizes, where lower values denote better model fit. The model with the lowest AICc represented the best model fit and was compared to subsequent models with the goal to retain the most parsimonious models. A change in AICc lower than 2 between models implied that the two models had a similar fit. ANOVA was also used to compare the best model to subsequent models. Area under the curve (AUC) and its 95% confidence interval computed from the DeLong method were also calculated for each model, and AUCs between models were also compared with the DeLong method. This approach using multi-model inference to retain the most discriminant markers has been validated recently in two independent cohorts [
1].
We performed two sets of logistic regression analyses. Given that only 80 out of the 110 participants had all plasma biomarkers level (p-tau217, NfL, Aβ42/Aβ40, and GFAP), we first aimed at identifying which plasma biomarker(s) were most related to progression to AD dementia. In this first set of model comparisons, only the four plasma biomarkers were entered as predictors, with conversion to AD dementia as the outcome. Plasma biomarkers with odds ratios with a p-value < 0.10 were kept for further analysis. Second, the same approach of using AICc for model selection was repeated to distinguish among possible models combining the key identified plasma biomarker(s), APOE4 status, hippocampal volume, and mPACC with conversion to AD dementia as outcome. The overall goal was to determine the best (lowest AICc) and the most parsimonious models (similar fit and AUC as the best model) combining plasma and key AD markers in relation to clinical progression. Sensitivity analyses also included basic demographics (age, sex, and education) as additional predictors to assess whether they were important factors in assessing risk of progressing to AD dementia.
Data availability
Anonymized data can be shared to qualified academic researchers after request for the purpose of replicating procedures and results presented in the study. Data transfer must be in agreement with EU legislation regarding general data protection regulation and decisions by the Ethical Review Board of Sweden and Region Skåne, which should be regulated in a data transfer agreement.
Discussion
In this study, we investigated which combinations of plasma biomarkers and markers typically used in AD prognosis provided the best discrimination between MCI patients who progressed to AD dementia over 3 years compared to those who did not. Focusing first on four key plasma biomarkers, p-tau217 was the most predictive marker of clinical progression, both when used alone and in combination with the other plasma measures (AUC of 0.84). There was no meaningful improvement of combining p-tau217 with other plasma biomarkers. However, the discrimination between the two MCI groups was further improved when incorporating a score of global cognition, hippocampal volume, and APOE4 genotype as predictors along with plasma p-tau217 (AUC of 0.89). Aiming for a parsimonious model that would have a similar model fit as the one with all variables, we found that including only plasma p-tau217 and a global cognitive score yielded comparable results.
Aβ plaques, tau tangles, and neurodegeneration are the core pathophysiological alterations of AD, as conceptualized in the biomarker-driven AT(N) classification [
18]. However, other pathophysiological pathways (X) are being investigated as potentially important in AD, resulting in the proposition of new ATX(N) classification, to be able to incorporate and adapt to new biomarkers [
19]. One such pathway is neuroinflammation and glial activation, which can be tracked with novel fluid biomarkers like GFAP, YKL40, and TREM2 [
20,
21]. We thus applied an ATX(N)-like framework in the MCI cohort to investigate which plasma biomarkers were most related to conversion to AD dementia. We selected the Aβ42/Aβ40 ratio (A), p-tau217 (T), NfL (N), and the increasingly studied astrocytic marker GFAP (X). Using a data-driven approach allowing all combinations of plasma biomarkers to derive the best models with conversion to AD dementia as outcome, it was clear that p-tau217 was consistently the best biomarker to discriminate MCI progressors from non-progressors. In fact, adding other plasma biomarkers in combination with p-tau217 did not result in improved model fit or better discrimination (no change in AICc or AUC, Table
2). This result further adds to the growing literature of plasma p-tau217 (or p-tau181) as important markers to track AD progression [
1,
22,
23].
We did not observe added value of plasma Aβ42/Aβ40 when combined with plasma p-tau217 for either clinical progression or Aβ-PET status. However, we cannot exclude that this result is assay-specific, since Aβ42 and Aβ40 were measured using Simoa immunoassays, which have been shown to be less accurate than certain mass spectrometry-based Aβ assays [
12]. Still, a recent study where the outcome was brain amyloidosis also found that at the MCI stage, plasma p-tau217 was the best biomarker to identify Aβ-positive participants, with or without plasma Aβ42/Aβ40 (quantified with mass spectrometry-based assay) as an additional predictor [
12]. These results also align with the AD pathophysiology, with Aβ proteins starting to change and plateauing earlier in the AD continuum making them less informative for predicting much later cognitive decline, while p-tau continues to increase through the prodromal stage of the disease to the dementia stage, shown both in studies using CSF [
24‐
26] and plasma [
27,
28] biomarkers. Accumulating evidence also suggested close relationship between Aβ and GFAP rather than between GFAP and p-tau [
29‐
32]. While plasma Aβ and GFAP might be more closely associated with Aβ pathology in earlier stages of the disease, p-tau is more associated with clinical progression.
With the idea of implementing a cost-effective predictive model, we also focused on combining p-tau217 with other easily accessible measures in AD. We used measures in line with those investigated in the initial clinical trial in combination with Aβ-PET [
7], i.e., hippocampal volume,
APOE4 genotype and a global score of cognition. Adding these three variables with p-tau217 resulted in the best identification of MCI progressors to AD dementia, with an AUC of 0.89. Previous studies with a similar outcome but using CSF or PET biomarkers rather than plasma also often found an added value of such additional non-biomarker measures [
33‐
35]. However, when comparing the best model that included all variables to the best subsequent combinations of variables, we found that p-tau217 and the global cognitive score mPACC were largely comparable to the full model. Model fits were similar, but the AUC was slightly lower (0.89 for full model vs. 0.87 for p-tau217 + mPACC,
p = 0.07). Across all models combining p-tau217 with other variables, we should note that mPACC was always a significant contributor, while hippocampal volume or
APOE4 genotype were often at trend-level with
p-values around 0.1. Our approach for this study and the main results corroborate the findings from a recent large-scale study from our group where p-tau217, memory score, executive function, and
APOE4 genotype was the best combination to determine conversion to AD-dementia within 4 years in cognitively normal older adults or MCI patients [
1]. As with the current study, NfL, structural measures from MRI, and basic demographics only had little influence on model performance. Notably, model accuracy was similar in both studies, with an AUC of 0.91 in the large-scale study and 0.89 here. Overall, across very different datasets, there is converging evidence that plasma p-tau217 in combination with easily accessible AD markers have the highest potential to help detect individuals at risk of progression to dementia. To move the field forward and get closer to implementing the most promising markers more widely in clinical practice, it is important to validate results evaluating risk of conversion to AD in multiple samples. Further, in cases of smaller sample size as the current study where only minor differences between models existed, we propose that p-tau217 and global cognitive score would be sufficient predictors for a parsimonious, most easily accessible model predicting progression to AD dementia.
Limitations
There are a few limitations to consider to this study. Unfortunately, the four biomarkers of interest were not available for all participants, due to limited amount of plasma to analyze for some individuals. We tried to circumvent this aspect by first selecting the plasma markers most related to conversion to AD, which allowed us to conduct further analyses in the full sample, in which p-tau217 level was measured in all participants. Still, the limited plasma quantity precluded us from measuring other p-tau isoforms or Aβ42 and Aβ40 using the most accurate mass spectrometry-based methods (which are superior to plasma Aβ immunoassays used here) [
36]. Future studies should evaluate if combing p-tau isoforms with Aβ42/Aβ40 measured using mass spectrometry-based methods would offer improved performance in different stages of AD. Only plasma and no CSF was available in this sample; therefore, we were not able to test how well plasma p-tau levels reflect CSF level. Still, we hypothesize that CSF p-tau217 would have been a key marker related to progression to AD [
37]. Given the somewhat small sample size, we also aimed to restrict the number of variables included in logistic regression models and opted for a global score of cognition instead of multiple neuropsychological tests. With memory and executive function being both important cognitive domains to predict AD dementia in the previous large-scale study [
1], we derived a modified PACC (we were missing the Free and Cued Selective Reminding Test included in the original version), analogous to the PACC5, which encompassed both domains and is widely used [
15,
16]. However, we acknowledge that the Free and Cued Selective Reminding Test might have provided sensitive memory measure to the composite score. Ten participants had missing
APOE4 genotype, but we replicated the main results when restricting the variables of interest to p-tau217, cognition, and hippocampal volume. We should also mention the current study focused on determining conversion to AD dementia where all converters had brain amyloidosis, while the outcome of the original clinical trial was probable AD, relying on the clinical status from the clinical adjudication committee. Lastly, all participants were categorized as amnestic MCI, and thus generalization of the results to more diverse MCI patients should be determined.
Declarations
Competing interests
Henrik Zetterberg has served at scientific advisory boards and/or as a consultant for Abbvie, Alector, Eisai, Denali, Roche, Wave, Samumed, Siemens Healthineers, Pinteon Therapeutics, Nervgen, AZTherapies, CogRx, and Red Abbey Labs, has given lectures in symposia sponsored by Cellectricon, Fujirebio, Alzecure and Biogen, and is a co-founder of Brain Biomarker Solutions in Gothenburg AB (BBS), which is a part of the GU Ventures Incubator Program. Oskar Hansson has acquired research support (for the institution) from AVID Radiopharmaceuticals, Biogen, Eli Lilly, Eisai, GE Healthcare, Pfizer, and Roche. In the past 2 years, he has received consultancy/speaker fees from AC Immune, Alzpath, Biogen, Cerveau and Roche. Sebastian Palmqvist has served on scientific advisory boards and/or given lectures in symposia sponsored by F. Hoffmann-La Roche, Biogen, and Geras Solutions. Kaj Blennow has served as a consultant, at advisory boards, or at data monitoring committees for Abcam, Axon, Biogen, JOMDD/Shimadzu. Julius Clinical, Lilly, MagQu, Novartis, Prothena, Roche Diagnostics, and Siemens Healthineers, and is a co-founder of Brain Biomarker Solutions in Gothenburg AB (BBS), which is a part of the GU Ventures Incubator Program. Gill Farrar and Christopher Buckley are full time employees of GE Healthcare who sponsored the Wolk et al (2018) study from which these plasma samples were derived. Dr. Wolk has received research support (for the institution) for Eli Lilly, Biogen, and Merck. He has also received consulting fees from GE Healthcare and Neuronix and Honoria for DSMB participation from Functional Neuromodulation.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.