We used data from the Age, Gene/Environment Susceptibility (AGES)-Reykjavik Study. The AGESReykjavik Study is a longitudinal, population-based cohort study originating from the Reykjavik Study, as described previously in detail [ 21 ]. The Reykjavik Study was initiated in 1967 and included individuals born between 1907 and 1935 from the Reykjavik area. Between 2002 and 2006, 5764 randomly chosen surviving participants of the Reykjavik Study were examined for the AGES-Reykjavik Study. Among those participants with an MRI (n = 4811), a random sample of 1200 was selected for the measurement of NfL, GFAP, and t-tau as a part of the MarkVCID project [ 22 ]. Characteristics among individuals included in the substudy and those in the original cohort were comparable (Supplementary Table 1). The study was approved by the National Bioethics Committee in Iceland (approval number: VSN-00–063) and by the National Institute on Aging Intramural Institutional Review Board. All participants gave written informed consent.

Plasma tubes were inverted 5 to 10 times and centrifuged for 10 min at 2000 × g within 1 h of collection. Five hundred microliter aliquots were transferred to polypropylene tubes and samples were places into − 80° freezer within 2  h of collection. Plasma samples were shipped to the Laboratory for Clinical Biochemistry Research at the University of Vermont, which has a strong quality assurance program for assays and is equipped with Simoa HD-1 Analzyer (Quanterix). Plasma levels of NfL, GFAP, and t-tau were measured using the Simoa Neurology 4-Plex Kit on a Simoa HD-1 Analyzer (Quanterix). Analytical ranges and inter-assay coeficients of variance are provided in Supplementary Table 2. A certified laboratory technician, blinded to diagnostic and ethnic groups, performed all assays between November and December 2019 using a single batch of reagents.

All eligible participants were ofered high-resolution 1.5  T MRI (Signa Twin-Speed; General Electric Medical Systems). A standardized imaging protocol was used, as described previously [ 24, 25 ]. This protocol included the following sequences: 3-dimensional spoiled-gradient recalled T1-weighted, proton density/T2-weighted fast-spin echo, fluid-attenuated inversion recover (FLAIR), and T2a-weighted gradient-echo type echo-planer image (GRE-EPI). All images were acquired to give full brain coverage with slices angled parallel to the anterior commissureposterior commissure line to give reproducible image views in the oblique-axial plane. We evaluated the following four markers of SVD: WMHV, subcortical infarcts, cerebral microbleeds, and large perivascular spaces. The identification of these markers was made in accordance with expert guidance that provided definitions and neuroimaging standards for markers and consequences of SVD [ 26 ]. Total brain parenchyma volume (TBV) and WMHV were computed automatically with a previously described image analysis pipeline [ 27 ] and were expressed as the percentage of total intracranial volume. Quality checks were done after tissue classification, as described in detail previously [ 27 ]. In brief, quality control consisted of visual inspection of a verification image for each subject including 14 a priori selected slice locations from each of the pulse sequences (T1, PD, T2, FLAIR), evenly distributed across the entire brain in the axial, coronal, and sagittal planes. Unsuccessful tissue classification that  could not be rescued by repeated processing or manual editing occurred in 53 cases, mostly due to severe motion artifacts. These scans were excluded from the analytical sample. Other lesions were evaluated by trained radiographers using a standardized protocol [ 24, 25, 28 ]. Subcortical infarcts were defined as brain parenchyma defects not extending into the cortex, with a minimum diameter of 4 mm and a signal intensity equal to cerebrospinal fluid on all pulse sequences (T2-weighted, proton density–weighted, and FLAIR), and surrounded by an area of high intensity on FLAIR images and without evidence of hemosiderin on T2a-weighted GRE-EPI sequence [ 25 ]. Cerebral microbleeds were defined as focal areas of signal void visible on the T2a-weighted GRE-EPI sequence [ 24 ]. Large perivascular spaces were defined as defects on the subcortical area without a rim or area of high signal intensity on FLAIR and without evidence of hemosiderin on the T2aweighted GRE-EPI sequence [ 28 ]. The total number of large perivascular spaces was based on the presence in the basal ganglia complex, along the paths of the perforating lenticulostriate arteries, and in white matter along the paths of the perforating medullary arteries [ 28 ]. Information on reproducibility of the process, including the image acquisition and the automatic pipeline, is provided in detail elsewhere [ 27 ]. For the volumetric markers, reproducibility was performed in 32 subjects and yielded an interclass correlation 0.98 for both TBV and WMHV [ 27 ]. For the other markers, intra- and inter-observer reliability was based on 2 ratings within a 6-month interval and indicated good reliability. Intra-observer reliability was 0.89 and 0.93 for subcortical infarcts [ 29 ], 0.75 and 0.73 for cerebral microbleeds [ 30 ], and 0.88 and 0.93 for large perivascular spaces [ 28 ], respectively. Interobserver reliability was 0.76 for subcortical infarcts [ 29 ], 0.70 for cerebral microbleeds [ 30 ], and 0.66 for large perivascular spaces, respectively [ 28 ].

Incident all-cause dementia was assessed at the follow-up examination (2007–2011) using a 3-step procedure, as described previously [ 31 ]. This was the same assessment used for the ascertainment of prevalent all-cause dementia at the baseline examination performed by the same panel of professionals [ 32 ]. In brief, the Mini-Mental State Examination and the Digit Symbol Substitution Test were administered to all participants. Individuals who screened positive based on a combination of these tests (< 24 on the Mini-Mental State Examination or < 8 on the Digit Symbol Substitution Test) were administered a diagnostic battery of neuropsychological tests. Based on performance on the Trails B and the Rey Auditory Verbal Learning Test, a subset of these individuals (Auditory Verbal Learning test ≤ 18 or Trails B ≥ 8 for the ratio of time taken for Trails B/Trails A corrected for the number correct: (time trails B/number correct Trails B)/(time Trails A/number correct Trails A)) underwent a proxy interview and were examined by a neurologist. A consensus diagnosis, based on the Diagnostic and Statistical Manual of Mental Disorders 4th edition criteria, was made by a panel of experts including a geriatrician, a neurologist, a neuropsychologist, and a neuroradiologist. In addition, all participants were continuously followed up for dementia diagnosis through vital statistics, hospital records, and the nursing and home-based Resident Assessment Instrument [ 33 ]. Follow-up for dementia ended October 4, 2015.

Education level (primary, secondary, and college/ university) and smoking history (never, former, current) were assessed by questionnaire. Medication use was assessed by questionnaire and from medication bottles brought to the clinic. Blood pressure, body mass index, and lipid levels were measured using standardized protocols [ 21 ]. We defined diabetes as a self-reported history of diabetes, use of blood glucose–lowering drugs, or a fasting blood glucose level of ≥ 7.0 mmol/l. Stroke (i.e., symptomatic brain infarct or hemorrhage) prevalent at baseline was obtained from medical records. Incident strokes that occurred between the baseline and follow-up examination were adjudicated by a dementia neurologist, a stroke neurologist, and a neuroradiologist.

Of the 1200 individuals in this biomarker substudy, 6 participants had missing data on one or more of the plasma biomarkers and another 8 did not have specific MRI images needed for assessment of cerebral microbleeds. Missing data on plasma biomarkers was due to technical reasons, including missing sample (n = 1), insuficient volume available (n = 3), and invalid result (n = 2). In addition, we excluded 10 participants with missing data on covariates. In the remaining 1176 participants, 107 were excluded because of a diagnosis of dementia at baseline (n = 47) or because of missing data on incident dementia (n = 60). The ifnal study sample included 1069 participants (Supplementary Fig.  1). Participants excluded from the biomarker substudy sample were older, less educated, and were more likely to have hypertension or type 2 diabetes compared to those included in the analysis (Supplementary Table 3).

We summarized the four markers of SVD into a composite sum score (range 0–4) to reflect bur den of SVD (SVD burden score) as done previously [ 34 ]. One point per SVD marker was assigned based on the following cut-ofs: for WMHV highest quar tile versus lowest three quartiles; and for subcortical infarcts, cerebral microbleeds, and large perivascular spaces, presence (i.e., ≥ 1 lesion(s)) versus absence). In all analyses, the SVD burden score was analyzed on a continuous scale to enhance the statistical power of our analysis, as done previously [ 34, 35 ]. We also evaluated the SVD burden score on an ordinal scale. Plasma biomarkers were transformed using a natural logarithm (i.e., base-e log) to normalize their skewed distribution.

The statistical analysis proceeded in three stages. First, to evaluate the relation between plasma biomarkers and the SVD burden score, we used linear regression to estimate regression coecfiients (betas) and 95% condfience intervals (95%CIs) for the association of plasma NfL, GFAP, and t-tau with the SVD burden score. Second, to evaluate the association between plasma biomarkers and incident dementia, we used Cox regression to estimate hazard ratios (HRs) and 95% CIs for the association of plasma NfL, GFAP, and t-tau with incident dementia using time-in-study as the time scale. Follow-up time was calculated from the AGESReykjavik baseline examination (2002–2006) to incidence of dementia, death, or end of follow-up (October 4, 2015), whichever came rfist. The proportional hazard assumption was assessed by visual inspection of Kaplan–Meier curves (Supplementary Fig.  2). Third, to investigate whether the SVD burden score explained the association of plasma NfL, GFAP, and t-tau with incident dementia (if any), we entered the SVD burden score as a covariate in the biomarker-dementia models. We did not consider the SVD burden score to be on the putative causal pathway of the plasma biomarkers leading to dementia, and we, therefore, did not do a formal mediation analysis. To quantify the degree to which the SVD burden score attenuated the association of plasma biomarkers with incident dementia, we calculated the explained eefcts. The explained eefcts were calculated as the multiplied eefcts of plasma biomark ers and brain MRI markers and brain MRI markers and incident dementia, adjusted for the plasma biomarkers [ 36 ]. The calculation of the explained eefct is summarized in Supplementary Fig. 3. We used bootstrapping (10,000 samples) to calculate bias-corrected 95% CIs for the explained eefcts.

All analyses were adjusted for age and sex (model 1) and additionally for education level, smoking history, diabetes status, body mass index, total cholesterol-to-HDL cholesterol ratio, use of lipid-modifying medication, systolic blood pressure, and use of antihypertensive medication (model 2). These covariates were selected on the basis of their biological plausibility, since they are known to be associated with SVD [ 37 ] or dementia [ 38 ]. Data on the association between plasma NfL, GFAP, and t-tau and the covariates included in model 2 is still limited [ 39 ]. However, these covariates are known to be associated with the neurodegenerative mechanisms that are presumed to be reflected by plasma levels of NfL, GFAP, and t-tau [ 40, 41 ].

We performed several sensitivity analyses. First, we repeated the analyses for each of the individual SVD markers separately. Second, to minimize potential confounding or mediating efects by TBV or stroke, we repeated the analysis additionally adjusting for TBV and baseline stroke or incident stroke during follow-up. Third, to investigate the efect of the definition of WMHV, we evaluated WMHV expressed on a continuous scale and WMHV expressed as higher versus lower than the median.

The analyses with the SVD burden score modelled on an ordinal scale showed a linear increase for the risk of dementia for a higher SVD burden score (Supplementary Table 6). In addition, plasma NfL, but not GFAP and t-tau, increased linearly for a higher SVD burden score (Supplementary Table  7). The non-linear association between plasma GFAP and t-tau and the SVD burden score is in accordance with the non-signicfiant nifding of the explained eefct by the SVD burden score of the associations between plasma GFAP and t-tau and incident dementia. Additionally adjusting the association between the plasma biomarkers and incident dementia for the SVD burden score on an ordinal scale yielded results similar to those obtained when we additionally adjusted for the SVD burden score on a continuous scale (Supplementary Fig.  4). A higher plasma NfL was associated with a higher WMHV and presence of subcortical infarcts; only WMHV explained part of the association of plasma NfL with Data are means (standard deviation) or median (interquartile range) Abbreviations: NfL, neurofilament light; HDL, high-density lipoprotein; TBV , total brain volume; SVD, cerebral small vessel disease; WMHV, white matter hyperintensity volume; GFAP, glial fibrillary acidic protein; t-tau, total tau aTBV and WMHV were expressed as percentage of intracranial volume bSVD burden score was calculated by assigning one point per cerebral small vessel disease marker based on the following cutofs (range 0–4): for WMHV highest quartile vs lowest three quartiles, and for subcortical infarcts, cerebral microbleeds, and large perivascular spaces presence vs absence incident dementia (Supplementary Fig.  5 and 6 and Supplementary Table 8). A higher plasma GFAP was only associated with WMHV, and similar to plasma NfL, WMHV explained part of the association of plasma GFAP with dementia (Supplementary Figs.  5 and 6 and Supplementary Table  8). Plasma t-tau was associated with a higher WMHV, but not with any of the other markers of SVD (Supplementary Fig.  5). cerebral small vessel disease; NfL, neurofilament light; GFAP, glial fibrillary acidic protein; t-tau, total tau. aSVD burden score was calculated by assigning one point per cerebral small vessel disease marker based on the following cut-ofs (range 0–4): WMHV highest quartile vs lowest three quartiles, and for subcortical infarcts, cerebral microbleeds, and large perivascular spaces presence vs absence Fig. 2 Associations between plasma NfL, GFAP, and t-tau and incident dementia with and without adjustment for the  SVD burden scorea. Hazard ratios for incident dementia are expressed per natural log-transformed pg/ml higher plasma NfL, GFAP, or t-tau. Model 1 adjusted for age and sex. Model 2 additionally adjusted for education level, diabetes status, smoking history, body mass index, total cholesterol-to-HDL cholesterol ratio, use of lipid-modifying medication, systolic blood pressure, and use of antihypertensive medication. Models 1 and 2 represent the total efect, and model 2 + adjustment for the SVD burden score represents the direct efect. Total and direct efect are defined in Supplementary Fig.  3. Abbreviations: NfL, neurofilament light; GFAP, glial fibrillary acidic protein; t-tau, total tau; SVD, cerebral small vessel disease. aSVD burden score was calculated by assigning one point per cerebral small vessel disease marker based on the following cut-ofs (range 0–4): WMHV highest quartile vs lowest three quartiles, and for subcortical infarcts, cerebral microbleeds, and large perivascular spaces presence vs absence scorea of the associations between plasma NfL, GFAP, Explained efects Hazard ratios for incident dementia are expressed per natural log-transformed pg/ml higher plasma NfL, GFAP, or t-tau. Total efect, direct efect, and explained efect are defined in Supplementary Fig. 3. The explained efect quantifies the degree to which the SVD burden scorea attenuated the association of plasma biomarkers with incident dementia. All analyses adjusted for age, sex, education level, diabetes status, smoking history, body mass index, total cholesterol-to-HDL cholesterol ratio, use of lipid-modifying medication, systolic blood pressure, and use of antihypertensive medication Abbreviations: SVD, cerebral small vessel disease; NfL, neurofilament light; GFAP, glial fibrillary acidic protein; t-tau, total tau aSVD burden score was calculated by assigning one point per cerebral small vessel disease marker based on the following cut-ofs (range 0–4): WMHV highest quartile vs lowest three quartiles, and for subcortical infarcts, cerebral microbleeds, and large perivascular spaces presence vs absence Results were similar when we additionally adjusted for TBV or prevalent or incident stroke (Supplementary Fig. 7 to 10 and Supplementary Table 9 and 10).

Results were similar using WMHV on a continuous scale, or using  WMHV expressed as higher versus lower than the median, instead of comparing those in the highest quartile of WMHV to those in the lowest three quartiles (Supplementary Figs.  11 and 12 and Supplementary Tables 11).