One-hundred sixty-seven patients were enrolled in this retrospective study at the Memory Clinic of the Department of Neurology, Medical University of Vienna (MUV). As various patients with the main concern of subjective/objective cognitive decline are referred to our specialized memory clinic both by specialists and generalists, without a preselection, our patient cohort rather reflects a more heterogeneous study population and thus resembles more closely a real-world setting. Using two existing registries, the Dementia Registry RDA MUV (EK 1323/2018) and the BIOBANK MUV (EK 2195/2016), we identified 123 patients with a diagnosis along the clinical spectrum of cognitive decline, i.e., mild cognitive impairment (MCI, n = 63) due to AD and AD-dementia (n = 60). Additionally, 44 age-matched healthy controls (HC) were included. These participants were recruited from an unselected population of patients, that were administered to the Department of Neurology and received further neurological examination, including brain imaging and lumbar puncture, to rule out an underlying neurological disorder. The main diagnoses of these patient cohorts consisted of idiopathic cranial nerve palsies, headache syndromes, and somatic symptom disorders and showed no signs of a neurodegenerative disease or subjective/objective cognitive decline.

All 123 patients with an objective cognitive decline (MCI, AD) underwent a thorough standardized diagnostic examination including physical and neurological evaluation, neuropsychological testing, magnetic resonance imaging (MRI) of the brain, and basic laboratory testing. For a subset of patients, we extended our diagnosis with a biomarker-based approach. CSF analysis of established AD biomarkers [amyloid-beta 42 (Aβ42), total tau (tTau), and phosphorylated tau (pTau)] was available in 75 patients, amyloid-PET imaging was performed in 80 patients, and 60 patients underwent both diagnostic methods.

Diagnoses of MCI and dementia due to AD were based on the recommendation of the National Institute of Ageing and Alzheimer’s Association (NIA-AA; Albert et al., 2011; McKhann et al., 2011). All 167 study participants were required to have a plasma EDTA sample stored in the Biobank MUV, for 103 study participants CSF samples were available as well.

The project was approved by the Ethics Committee of the Medical University of Vienna (EK 1965/2019) on November 28th, 2019.

The Neuropsychological Test Battery Vienna (NTBV) was administered to assess cognitive function, including domains of attention, language, executive functioning, and episodic memory (Pusswald et al., 2013; Lehrner et al., 2015a). Adequate normative data from cognitively unimpaired individuals were available and z-scores for each variable were calculated and corrected for age, education, and sex. Screening of cognitive impairment consisted of Mini-Mental State Examination (MMSE), Global Deterioration Scale (GDS), and Wortschatztest (WST), a standardized vocabulary test providing an estimate of premorbid intelligence level (Schmidt and Metzler, 1992). Furthermore, the Vienna-Visuo-Constructional Test 3.0 (VVT-3.0) was applied to assess the visuo-constructive performance (Lehrner et al., 2015b). Depressive symptoms were measured via Beck Depression Inventory (BDI-II; Kühner et al., 2007).

Apolipoprotein E (APOE) genotyping was performed in 143 patients using quantitative polymerase chain reaction (qPCR) with TaqMan probes (Thermofisher) evaluating two single nucleotide polymorphisms (SNPs) in the APOE gene (rs429358 and rs7412). Each sample was tested for both SNPs in triplicates using 20 ng deoxyribonucleic acid (DNA). Allelic discrimination analysis was used to determine the APOE genotype of the study participants.

All patients underwent at least a T1-weighted MR sequence, a T2-weighted or a Fluid-attenuated inversion recovery (FLAIR) MR sequence, and a diffusion-weighted MR sequence within the routine diagnostic setting for the evaluation of the extent and pattern of atrophy, the presence and degree of vascular lesions and to exclude other underlying pathologies causing cognitive decline and diffusion restricted areas.

Eighty patients underwent an amyloid-PET scan with [¹⁸F] flutemetamol (n = 28) or [¹¹C] Pittsburgh compound-B (PiB, n = 52). Amyloid-PET imaging was performed on one of two possible PET scanner systems (Siemens Biograph 64 True Point, Erlangen, Germany or GE Advances PET, GE Healthcare Institute, Waukesha, Wisconsin, USA). All studies were performed under strictly controlled conditions. In short, either ~400 MBq of [11C] PiB (in-house production according to previously published recommendations; Philippe et al., 2011) or 185 MBq of [18F] flutemetamol (Vizamyl^®, GE Healthcare) were injected intravenously into a peripheral vein with starting image acquisition 40 min p.i. for [11C] PiB and 90 min p.i. for Vizamyl^®, where the tracer accumulation in the brain is reaching the maximum. Subsequently, the image acquisition was performed for about 20 min following a computed tomography (CT) acquisition for attenuation correction using Siemens Biograph 64 True Point.

Scans were rated visually as positive or negative for the presence of amyloid pathology in the cortex by an experienced nuclear medicine physician according to the guidelines of the tracer manufacturers.

CSF was obtained by lumbar puncture between the L3/L4, L4/5, or L5/S1 intervertebral space, collected in polypropylene tubes and further stored at −20°C until biomarker analysis (as for Aβ42, pTau 181, and tTau), or immediately at −80°C for future research purposes (Teunissen et al., 2014; Duits et al., 2015). Levels of Aβ42, pTau 181, and tTau were measured with commercially available ELISA (Innotest hTAU-Ag, Innotest phosphoTAU 181p, Innotest beta-amyloid 1–42; Vanmechelen et al., 2000; Vanderstichele et al., 2009). The cut-off for these biomarkers were based on the manufacturer’s recommendation (Aβ42 < 500 pg/ml, pTau 181 > 61 pg/ml, tTau > 300 pg/ml). From these measurements, Innotest Amyloid Tau Index (IATI) was calculated for each patient (measured as Aβ42/(240+1.18xtTau), reference values <1 pg/ml indicative of AD pathology, >1 pg/ml—normal; Hulstaert et al., 1999; Tabaraud et al., 2012).

EDTA plasma was collected through venepuncture and stored at −80°C in our local biobank. Concentrations of NfL and GFAP were quantified with an ultrasensitive single molecule array (SIMOA^®) using the Neurology 2-Plex B kit from Quanterix in CSF and plasma. Detailed analyses are described elsewhere (Altmann et al., 2020). In short, equilibrated calibrators, samples, and controls were diluted (1:4 for plasma and 1:100 for CSF) and incubated with detector and paramagnetic reagents provided by the manufacturer. Streptavidin ß-galactosidase was added to each well before samples were transferred to the Quanterix SR-X analyzer for measurement of protein levels. All samples were analyzed as duplicates and all assay materials were obtained from the same kit lot. Intra-assay coefficient of variation (CV) was <12% for GFAP Plasma, <13% for GFAP CSF, <9% for NfL Plasma and <8% for NfL CSF. Inter-assay CV for two samples measured repeatedly on 10 plates was well acceptable (<12% for GFAP Plasma, <14% GFAP CSF, <8% NfL Plasma and <10% NfL CSF). Five patient samples were excluded from further analysis due to a high CV (>20%) and therefore not included in this study.

Amyloid positivity was defined by CSF (IATI < 1) and/or amyloid-PET imaging. In cases where both examinations were available or discordant results were obtained, amyloid status was determined by PET.

Data are presented as n (percent) or median (interquartile range) as appropriate. Testing for differences between groups was performed using the chi-square test, the Mann-Whitney-U-test, or the Kruskal-Wallis-test. The correlation was assessed using Spearman’s rank correlation coefficient. To evaluate the discriminatory performance of the biomarkers assessed herein, the cohort was split into pairs of two diagnoses (e.g., AD and HC) and the response variable was coded as existing for the more severe diagnosis (i.e., MCI when assessing MCI vs. HC). Next, a baseline model consisting of sex, age, and APOE4 status was constructed using logistic regression. A receiver operating characteristic (ROC) curve was plotted and the area under the curve (AUC) was measured. Optimal cutoffs were calculated using Youden’s J-Statistic (Youden, 1950). The baseline model was then supplemented by either level of plasma GFAP, plasma NfL, or both, and the AUC of each model was compared using DeLong’s test for correlated AUC curves (DeLong et al., 1988). A p-value of <0.05 was interpreted as statistically significant. All calculations were performed in R (Version 4.0.4) and the pROC package was used for ROC calculations (Robin et al., 2011).

Demographic and clinical characteristics are listed in Table 1.

We observed no significant difference in sex distribution between the groups, while HC were significantly younger than the two patient groups (p < 0.01 for HC vs. MCI and AD). MMSE decreased significantly with progressing disease with the lowest score in the AD group (p < 0.001). Data of APOE4 carriership (carriers of at least one APOE4 allele) was available for 143 patients, with the highest occurrence of APOE4 alleles in AD patients (33 of 53 patients, 62.3%), compared to 22 out of 54 patients in the MCI group (40.7%). A chi-square test of independence was performed to examine the relationship between the APOE4 status and the diagnosis. As can be seen by the frequencies cross-tabulated in Table 1, there was a significant relationship between APOE4 status and diagnosis (X²_{(2, N = 167)} = 8.5078, p < 0.05.

For a subset of patients (n = 75) CSF analysis of established AD biomarkers was available (Aβ42, tTau, pTau). While CSF tTau and pTau levels increased significantly with progression from MCI to AD (p < 0.001 and p < 0.05, respectively), the difference in Aβ42 concentration between MCI and AD reached no statistical significance. Accordingly, the IATI value was significantly lower in the AD group than in the MCI group (p < 0.001).

Amyloid-PET imaging was performed in 39 of 63 patients with MCI (61.9%) and 41 of 60 patients with AD (68.3%). Positive amyloid-PET imaging was significantly higher in AD patients with a total of 39 (95.1%) positive subjects in AD, compared to 23 patients with MCI (59%, p < 0.001). Taken together, biomarker data (CSF analysis or PET imaging) was available for 95 patients (56.9%), which demonstrated signs of amyloid pathology in a total of 76 patients (80%), determined by CSF IATI and/or amyloid-PET imaging as outlined previously.

Plasma GFAP displayed a gradual increase along the three cohorts, with the highest concentration in patients with AD (median 181.9 pg/ml, IQR 129.6, 269.6, Table 1 and Figure 1A). While plasma levels were significantly higher in patients with MCI vs. HC and AD vs. HC (p < 0.001), we observed no significant difference of plasma GFAP levels between MCI vs. AD.

Plasma NfL performed similarly to GFAP regarding the difference in concentrations between MCI vs. HC and AD vs. HC (p < 0.001, Table 1 and Figure 1B). In contrast to plasma GFAP, NfL levels showed a significant discrimination between MCI vs. AD (p < 0.05).

For 103 patients, CSF samples in our local biobank were available. Levels of CSF NfL increased gradually, with the lowest concentration in the HC group (median 584.1 pg/ml, IQR 449.6, 832.8) and the highest concentration in the AD group (median 1,559 pg/ml, IQR 1,026.6, 2,513.9).

On the contrary, CSF GFAP presented the lowest concentration in the MCI group (median 8,946.2 pg/ml, IQR 7,028.8, 13,842.7), followed by HC (median 11,145.3 pg/ml, IQR 6,980.5, 14,373.8) and AD (median 13,663.5 pg/ml, IQR 9,945.4, 21,059.1).

Both CSF biomarker levels allowed a good distinction between AD vs. HC and MCI vs. AD (NfL p < 0.001 for both measurements, GFAP p < 0.05 and p < 0.01, respectively), while a significant differentiation between HC vs. MCI could not be demonstrated.

In a logistic regression model including GFAP or NfL as the dependent variable and age and diagnosis (with HC as the comparator) as predictors, age was significantly associated with GFAP (B = 4.2, p < 0.001) and NfL (B = 0.4, p < 0.001), while diagnosis remained significant in both models.

Using Spearman correlation coefficient, the correlation of NfL and GFAP in CSF and plasma were analyzed (Figures 2A,B). Correlation between NfL in CSF and plasma performed better (R = 0.64, p < 0.001, Figure 2A) than the correlation of GFAP in CSF and plasma (R = 0.4, p < 0.001, Figure 2B).

To assess the clinical utility of GFAP and NfL in plasma, particularly in distinguishing healthy controls from patients with cognitive complaints (MCI and AD) and potentially predicting cerebral amyloid status, ROC analyses were performed and adjusted for sex and age. We constructed a diagnostic panel, consisting of well-established risk factors such as age, sex (defined as female > male), and APOE4 carriership (defined as carrying at least one copy of the APOE4 allele; i.e., age+sex+APOE4 panel) and compared it with a panel of age, sex, APOE4 carriership added by plasma NfL and plasma GFAP, called combination panel (i.e., age+sex+APOE4+GFAP+NfL panel, Figures 3A–D). Additionally, we analyzed each biomarker separately to evaluate the potential benefit of GFAP or NfL alone (i.e., age+sex+APOE4+GFAP panel and age+sex+APOE4+NfL panel).

When using the age+sex+APOE4 panel alone, we calculated an AUC of 73.4% (95%CI = 0.63–0.84) for HC vs. AD (Figure 3A), AUC of 64.6% (95% CI = 0.53–0.76) for HC vs. MCI (Figure 3B) and an AUC of 66.4% (95%CI = 0.56–0.77) for MCI vs. AD (Figure 3C). Regarding the diagnostic accuracy in predicting amyloid status and the distinction of amyloid-negative (Aβ-) from amyloid-positive (Aβ+) individuals, the AUC was 75% (95% CI = 0.62–0.88, Figure 3D).

By adding NfL to the panel (age+sex+APOE4+NfL panel), the discrimination between HC vs. AD reached a significantly higher AUC of 84.5% (95%CI = 0.76–0.93, Figure 3A) compared to the age+sex+APOE4 panel alone (p = 0.003), while the other calculations failed to achieve significantly better results (HC vs. MCI AUC 68.8%, 95%CI = 0.58–0.80, MCI vs. AD AUC 72%, 95%CI = 0.63–0.82), amyloid positivity (AUC 76.5%, 95%CI = 0.61–0.87, Figures 3B–D).

The combination of GFAP with the age+sex+APOE4 panel (age+sex+APOE4+GFAP panel) obtained an AUC of 91.3% (95%CI = 0.85–0.97) for HC vs. AD (p < 0.001, Figure 3A) and AUC of 81.3% (95% CI = 0.72–0.90) for HC vs. MCI (p < 0.01, Figure 3B) compared to the age+sex+APOE4 panel. The prediction of amyloid positivity demonstrated an AUC of 86% (95%CI = 0.70–0.97, Figure 3D), but missed statistical significance as well as the differentiation between MCI vs. AD (AUC 66.7%, 95%CI = 0.57–0.77, Figure 3C).

When combining the two biomarkers with the age+sex+APOE4 panel (combination panel: age+sex+APOE4+GFAP+NfL), the AUC of HC vs. AD reached 91.6% (95%CI = 0.85–0.98, p < 0.001, Figure 3A), AUC of HC vs. MCI 81.7% (95%CI = 0.73–0.90, p < 0.01, Figure 3B) and for amyloid positivity 87.5% (95%CI = 0.73–0.96, p < 0.05, Figure 3D), therefore, significantly outperforming the age+sex+APOE4 panel alone. Similar to the other two panels (age+sex+APOE4+GFAP panel and age+sex+APOE4+NfL panel), the combination panel could not improve the distinction between MCI vs. AD (AUC 72.3%, 95%CI = 0.63–0.82, Figure 3C).