Introduction

Alzheimer’s disease (AD) is characterized by a cascade of pathological events, starting with amyloid-beta (Aβ) and tau protein aggregation, which precede neurodegeneration and ultimately lead to cognitive decline1. In vivo detection of these pathological changes has been revolutionized by positron emission tomography (PET) imaging for Aβ and tau, allowing precise staging of disease progression. More recently, plasma biomarkers have emerged as a promising alternative for detecting AD pathology non-invasively2,3. Plasma phosphorylated tau 217 (pTau217), in particular, is closely associated with core AD pathology, Aβ and tau, providing a direct link to the primary pathological hallmarks of the disease4. In contrast, glial fibrillary acidic protein (GFAP) and neurofilament light chain (NfL) are associated with non-specific processes related to AD, with GFAP reflecting neuroinflammation and NfL indicating neurodegeneration5,6,7,8,9. These plasma biomarkers offer complementary insights into both the core pathologies and the downstream processes of AD, highlighting their potential to bridge gaps in understanding disease progression.

Diffusion tensor imaging (DTI) is a widely used MRI technique that measures the microstructural integrity of brain tissues in both white matter (WM) and gray matter (GM)10,11,12,13,14. Common DTI metrics, such as fractional anisotropy (FA) and mean diffusivity (MD), assess axonal integrity and tissue diffusivity15 while free water (FW) quantifies extracellular fluid and serves as a marker of neuroinflammation in AD15,16,17,18,19. Neuroinflammation plays a critical role in AD pathophysiology20,21,22 contributing to both microstructural disruptions and downstream macrostructural changes. Studies have demonstrated FW’s utility in capturing extracellular changes related to neuroinflammation, but its relationship with other DTI metrics like FA and MD, which may influence FW, has been less explored. This highlights the need for a comprehensive analysis of DTI metrics to better understand their combined role in AD pathology across WM and GM.

DTI metrics, which reflect microstructural changes, and structural MRI measures, such as hippocampal volume and cortical thickness, which capture macrostructural changes, provide complementary insights into AD progression. Microstructural disruptions, such as neuroinflammation and axonal degeneration, contribute to macrostructural alterations like hippocampal atrophy and cortical thinning, which in turn drives cognitive decline. Understanding the interplay between microstructural and macrostructural changes is crucial for elucidating the pathological mechanisms underlying AD and improving the prediction of cognitive outcomes.

In the present study, we aimed to investigate the role of FW on DTI in understanding the pathophysiology of AD, particularly its relationship with core AD pathologies, brain structural measures, and cognitive decline. Specifically, we examined the associations between plasma biomarkers and DTI metrics, the links between DTI metrics and brain structural measures, and how these structural measures mediate the relationship between DTI metrics and cognitive decline. This integrated approach provides a comprehensive understanding of the connections between molecular markers, micro- and macro-structural abnormalities, and clinical outcomes in AD.

Materials and methods

Study population

Participants were recruited from Samsung medical center between 2016 and 2023. We enrolled 968 participants aged ≥ 45 years who underwent Aβ PET, brain MRI including diffusion tensor image (DTI) and plasma AD biomarkers (217 with cognitively unimpaired (CU), 436 with mild cognitive impairment (MCI), and 315 with dementia of the Alzheimer’s type (DAT)23). All participants with CU status satisfied the following criteria: (1) no medical history that was likely to affect cognitive function based on Christensen’s health screening criteria20; (2) no objective cognitive impairment in any cognitive domain on a comprehensive neuropsychological test battery (above at least − 1.0 SD of age-adjusted norms on any cognitive test); and (3) independence in activities of daily living. All participants with amnestic MCI (aMCI) satisfied the criteria for aMCI with the following modifications24,25: (1) subjective cognitive complaints by the participants or caregivers; (2) objective memory impairment below − 1.0 SD on verbal or visual memory tests, (3) no significant impairment in activities of daily living, and (4) non-demented status. All participants with DAT satisfied the diagnostic criteria of the National Institute on Aging and Alzheimer’s Association (NIA-AA).

We excluded participants with structural lesions including territorial infarction, lobar hemorrhage, brain tumors, WMH due to demyelinating disorders, vasculitis, leukodystrophy, radiation injury, and hydrocephalus on MRI. Additionally, participants with abnormal laboratory results for complete blood count, electrolyte, vitamin B12, folate levels, syphilis serology, and liver/kidney/thyroid function tests were excluded.

The Institutional Review Board of Samsung Medical Center approved this study (No. 2021-02-135). All participants provided their informed consent to participate in the study and the data were collected in accordance with to the Declaration of Helsinki.

Aβ PET acquisition and quantification using Centiloid values

All the participants underwent Aβ PET18 F-florbetaben (FBB) PET or18 F-flutemetamol (FMM) PET using a Discovery STe PET/CT scanner (GE Medical Systems, Milwaukee, WI, USA). For FBB PET or FMM PET, a 20 min emission PET scan in dynamic mode (consisting of 4 × 5 min frames) was performed 90 min after injection of a mean dose of 311.5 MBq FBB or 197.7 MBq18F-flutemetamol, respectively. Three-dimensional PET images were reconstructed in a 128 × 128 × 48 matrix with 2 mm × 2 mm × 3·27 mm voxel size using the ordered-subsets expectation maximization algorithm (FBB, iteration = 4 and subset = 2018; F-flutemetamol, iteration = 4 and subset = 20).

We used regional direct comparison centiloid (rdcCL) method, to harmonize the quantitative analysis of Aβ PET across different ligands26,27. Data processing was conducted utilizing Statistical Parametric Mapping version 8 (SPM 8). The acquired PET images were co-registered with the individual MRI scans. Subsequently, The PET and MRI images were spatially normalized using the transformation parameters of the SPM 8 unified segmentation method for T1-weighted MRIs. The whole cerebellum was used as a reference region to calculate the standardized uptake value ratio (SUVR) using regional masks obtained from the GAAIN website (http://www.GAAIN.org). The common cortical target region was obtained by comparing the scans of 25 Aβ PET-positive Alzheimer’s disease dementia patients and 18 Aβ PET-negative controls among patients who underwent both FBB and FMM-PET with ligand-to-head. The MRI-based regional volumes of interest (VOIs) were defined by applying the common cortical target VOI (FMM-FBB global CTX VOI) based on the automated anatomical labeling (AAL) atlas. Finally, individual rdcSUVR values in the FMM-FBB global CTX VOI were calculated for all PET images and rdcSUVR values of the FBB-FMM CTX VOI were converted to rdcCL units using a conversion equation28. A more detailed process for obtaining global rdcCL is discussed elsewhere27. For clarity and consistency, global rdcCL will be referred to as centiloid in the following sections.

MRI acquisition and processing

We acquired three-dimensional T1 Turbo Field Echo MRI scans for all participants using a 3.0 T MRI scanner (Philips 3.0T Achieva) with the following imaging parameters: sagittal slice thickness, 1.0 mm with 50% overlap; no gap; repetition time of 9.9 ms; echo time of 4.6 ms; flip angle of 8°; and matrix size of 240 × 240 pixels reconstructed to 480 × 480 over a field view of 240 mm. All T1 scans were processed by CIVET pipeline (version 2.1.0)29 developed at the Montreal Neurological Institute for fully automated structural image analysis. In brief, T1 scans were transformed to MNI-152 template30,31 and N3 algorithms31 was used for correction of intensity non-uniformity. Next, tissue classification32 was performed, and WM and GM surfaces were constructed using the Constrained Laplacian-based Automated Segmentation with Proximities (CLASP)33 algorithm. The cortical thickness was defined of the Euclidean distance between the linked vertices of the WM and GM surfaces. A more detailed process is described in elsewhere34.

To measure hippocampal volume (HV), we employed an automated hippocampus segmentation method using a graph cut algorithm combined with an atlas-based segmentation and morphological opening as described in an earlier study35.

DTI acquisition

Axial diffusion-weighted images were acquired using a single-shot echo-planar imaging (EPI) sequence with b-values of 0 and 600 s/mm2. We applied 45 directions of motion probing gradients, and other imaging parameters for the diffusion MRI were as follows: repetition time (TR), 9300 ms; echo time (TE), 91 ms; section thickness, 2 mm; flip angle, 90°; field-of-view, 220 × 220 mm2; acquisition matrix, 128 × 128; scan time, 7 min and 26 s. All axial sections were acquired parallel to the anterior commissure-posterior commissure line.

Calculation of DTI metrics

All diffusion MRI data were processed using the functional magnetic resonance imaging of the brain (FMRIB) software library tool (FSL) (www.fmrib.ox.ac.uk/fsl)36. Briefly, correction for head movements and eddy current distortions were performed and brain masks were defined for each subject using the Brain Extraction Toolbox (BET) for the single B0 images. Single-tensor fractional anisotropy (FA) and mean diffusivity (MD) maps were generated using the DTIFIT tool implemented in FMRIB Software. We used FW imaging to estimate fractional volume of freely diffusing extracellular water molecules. The FW and FW-corrected maps (FAt and MDt) were then calculated using the open-source software package Diffusion Imaging in Python (DIPY, https://dipy.org/)37. A more detailed description of the bi-tensor model is extensively discussed elsewhere16.

Plasma collection and processing

We obtained 8 mL of blood from each participant and placed in a 0.5 M EDTA-containing tube and mixed it for 5 min. Plasma was extracted from the blood sample after a 10-min centrifuge (1300g) and dispensed into 5 or 10 vials at a volume of 0.3 mL each. All plasma samples were kept frozen at -75 °C until analysis. This process complied with the manual for human resource collection and registration of the National Biobank of the Republic of Korea. The median interval between plasma collection and Aβ PET scans was 4 days (interquartile range [IQR], 0–69 days).

Plasma samples were shipped in a frozen state at − 70 degrees to the Department of Psychiatry and Neurochemistry, University of Gothenburg for analysis. Plasma samples were thawed on wet ice, centrifuged at 500×g for 5 min at 4 °C. Plasma pTau 217 concentration was meaused using the commercial ALZpath pTau217 immunoassay on a Single molecule array HD-X instrument (Quanterix, Billerica, MA), a paramagnetic microbead-based sandwich enzyme-linked immunosorbent assay system. Plasma GFAP and NfL were quantified using the commercial Neurology 4-Plex E kit on the same instrument (Quanterix).

Clinical and vascular assessments

Of the 968 participants, 592 underwent follow-up Clinical Dementia Rating–Sum of Boxes (CDR-SOB) assessments. The number of follow-up visits varied by individual (mean ± SD, 1.79 ± 0.96 visits), with a mean interval of 3.06 ± 1.59 years between the baseline MRI and the final follow-up. Annual CDR-SOB change rates were then calculated using linear regression. Diabetes mellitus (DM) was defined as self-reported diagnosis by specialist or current use of diabetic medications or insulin, and hypertension (HTN) was defined as self-reported diagnosis or current use of antihypertensive medications.

Voxel-wise whole brain analysis

A tract-based spatial statistics (TBSS)38 and GM-based spatial statistic (GBSS)39 were used to elevate the relationship between plasma and DTI metrics for WM and GM, respectively.

Analysis of WM microstructural changes were performed using the skeleton projection of TBSS with the following procedure. First, FA maps of all subjects were aligned to the Montreal Neurological Institute (MNI 152) standard space using the FMRIB non-linear registration tool36. Next, a mean FA image was generated and thinned to create the mean FA skeleton, which represented the centers of all tracts common to the groups. FA > 0.20 was used as the threshold of the mean FA skeleton to include major WM pathways and exclude peripheral tracts and GM. Finally, the aligned FA map of each subject was projected onto the FA skeleton and other maps (FAt, MDt, and FW) were projected on the mean FA skeleton.

Analysis of GM microstructural changes were performed using GBSS, a GM along of TBSS, using the following steps. First, the Brain Extraction Tool was used to remove non-brain tissues from 3D T1-weighted images. Next, each skull-stripped 3D T1-weighted images was linear and non-linearly aligned to MNI152 1 mm standard space using the FMRI registration tool. Next, filed bias was corrected, and skull-stripped images were segmented into GM, WM, and CSF using the FMRIB automated segmentation tool40. Next, GM images were used to create median GM skeleton with a threshold of 0.2 to minimize the contribution of voxels from WM and CSF. After then, B0 maps of each subject were aligned to 3D T1-weighted images. All DTI maps were linearly and non-linearly aligned into MNI152 1 mm standard space and aligned maps of each subject were projected on the median GM skeleton.

Statistical analyses

Clinical characteristics of the study participants are presented as mean ± standard deviation for continuous variables and as frequency (%) for categorical variables. The level of pTau 217, GFAP, NfL, and centiloid were transformed to log-transformed to satisfy the assumption of linearity.

For TBSS and GBSS analyses, a general linear model framework, with age, sex, BMI status (only for plasma analyses), and APOE4 genotype as covariates, was used. The randomize tool (FSL) with 5000 permutations was applied to test the relationships between DTI metrics and plasma biomarkers. Results were then corrected for multiple comparisons by controlling the family-wise error (FWE) at a 0.01 significance level and applying threshold-free cluster enhancement (TFCE).

To examine the relationship between DTI metrics and brain structural measures, mean cortical thickness and hippocampal volume, linear regression was performed, with age, sex, education, APOE4 genotype, and intracranial volume (ICV) as covariates. The regression analysis was conducted using standardized variables, with DTI metrics (FAt, MDt,, and FW) calculated as the mean within the skeleton mask.

To investigate whether the brain structural measure mediates the effect of DTI metrics on longitudinal cognitive decline, we conducted mediation with DTI metrics and FW as predictor, brain structural measures (mean cortical thickness or hippocampal volume) as a mediator, and CDR-SOB score as an outcome after controlling age, sex, education, APOE4 genotype, and ICV. Bootstrapping was used to verify the significance of the direct and indirect effects.

Results

Characteristics of participants

Table 1 shows the characteristics of the 968 participants. The age (mean ± standard deviation [SD]) was 70.6 ± 8.6 years, and 600 out of 968 (62.0%) participants were females. The frequencies of diagnosis were 217 (22.4%) for CU, 436 (45.0%) for MCI, and 315 (32.5%) for ADD. The frequency of BMI status was 34 (3.5%) for underweight, and 263 (27.2%) for obese. The frequency of the APOE4 genotype was 429 (44.3%) for ε4 carriers.

Table 1 Demographic and clinical characteristics of study population.
Full size table

Relationships between AD biomarkers and DTI metrics

Figure 1 show the results of the TBSS analyses for FAt (Fig. 1A), MDt, (Fig. 1B), and FW (Fig. 1C). The voxel-wise linear regression indicated that lower FAt in widespread brain WM regions were associated with increased levels of NfL and GFAP, while lower MDt were linked to increased NfL, GFAP, and centiloid. The results are TFCE-enhanced and reported at p < 0.01, with FWE correction. There were no regions showing an association between FAt and pTau217 or centiloid, nor between MDt and pTau217. Furthermore, higher FW in widespread WM regions was found to be associated with higher levels of NfL, GFAP, and centiloid, while increased FW in specific brain regions, including the corpus callosum, corona radiata, and posterior thalamic radiation, showed a significant correlation with elevated pTau217 levels. However, several projection fibers, including the posterior limb of internal capsule and the anterior limb of internal capsule, did not show significant associations.

Fig. 1
Fig. 1
Full size image

Association between plasma biomarkers, centiloid, and WM integrity. TBSS showing the correlations between FA (A), MD (B), and FW (C) and plasma biomarkers (NLF, GFAP, pTau217, and centiloid). Results are TFCE enhanced, reported at p < 0.01, FWE corrected. The green voxels indicate the white matter skeleton used for statistical testing, and voxels with significant correlations (corrected p < 0.01) are shown in red to yellow. TBSS, tract-based spatial statistics; WM, white matter; FA, fractional anisotropy; MD, mean diffusivity; FW, free water, NfL, neurofilament light chain; GFAP, glial fibrillary acidic protein.

Figure 2 shows the results of the GBSS analyses. The results indicated a widespread association between these DTI metrics and GFAP across various brain areas, with lower FAt (Fig. 2A) and MDt values associated with higher GFAP levels (Fig. 2B). FAt showed a strong association with NfL in certain brain regions, including the cerebellum, the left and right inferior temporal gyri, the left and right fusiform gyri, the left parahippocampal gyrus, the left middle temporal gyrus, and the left lingual gyrus, whereas no significant associations were observed in others. There was no region showing an association between FAt and pTau217 or centiloid, nor between MDt and NfL, pTau217, or centiloid. FW showed significant associations with plasma biomarkers and centiloid in most cortical regions, with higher FW associated with elevated plasma and centiloid levels (Fig. 2C).

Fig. 2
Fig. 2
Full size image

Association between plasma biomarkers, centiloid, and GM integrity. GBSS showing the correlations between FA (A), MD (B), and FW (C) and plasma biomarkers (NLF, GFAP, pTau217, and centiloid). Results are TFCE enhanced, reported at p < 0.01, FWE corrected. The green voxels indicate the gray matter skeleton used for statistical testing, and voxels with significant correlations (corrected p < 0.01) are shown in red to yellow. GBSS, gray matter based spatial statistics; GM, gray matter; FA, fractional anisotropy; MD, mean diffusivity; FW, free water, NfL, neurofilament light chain; GFAP, glial fibrillary acidic protein.

The results of the TBSS analyses with additional adjustment for HTN and DM are presented in Supplementary Fig S1, and those of the GBSS analyses with the same covariates are shown in Supplementary Fig. S2. These sensitivity analyses revealed comparable spatial patterns and statistical significance, supporting the robustness of our findings regardless of vascular risk factor adjustment.

Relationships between DTI metrics and brain structural measures on MRI

The FAt was positively associated with cortical thickness (β = 0.046, p < 0.0001) (Fig. 3A) and hippocampal volume (β = 159.6, p < 0.0001) in WM (Fig. 3B), and similarly, in GM, it showed significant associations with cortical thickness (β = 0.08, p < 0.0001) (Fig. 3A) and hippocampal volume (β = 251.5, p < 0.0001) (Fig. 3B) after controlling for age, sex, APOE4 genotype, education, and ICV. The MDt showed a positive association with cortical thickness (β = 0.032, p < 0.0001) and hippocampal volume (β = 109.2, p < 0. 0001) in WM, while in GM, it was positively correlated with cortical thickness (β = 0.099, p < 0.0001) and hippocampal volume (β = 300.8, p < 0.0001). Contrary to the findings for FA and MD, the FW showed a negative association with cortical thickness (β = − 0.043, p < 0.0001) and hippocampal volume (β = − 103.7, p < 0.0001) in WM, and similarly, in GM, it was positively correlated with cortical thickness (β = − 0.097, p < 0.0001) and hippocampal volume (β = − 292.1, p < 0.0001), after controlling for the same covariates. Forest plots of the beta coefficients from the full models with all covariates are shown in Supplementary Figs. S3 and S4, and those additionally adjusted for HTN and DM are presented in Supplementary Figs. S5 and S6.

Fig. 3
Fig. 3
Full size image

Forest plot of beta coefficients from linear regression of brain atrophy and DTI metrics. The forest plots display the standardized beta coefficients from linear regression analyses with 95% confidence intervals, adjusted for age, sex, BMI, APOE4 genotype, and ICV. The forest plots show each regressor’s standardized beta linear regression coefficients with 95% confidence interval. Plots displaying the beta coefficients from linear regression analyses of cortical thickness with FA, MD, and FW (A), as well as hippocampal volume with FA, MD, and FW (B). FA, fractional anisotropy; MD, mean diffusivity; FW, free water.

Relationships between DTI metrics brain structural measures, and clinical progression

Figures 4 and 5 illustrate the relationships between DTI metrics in WM and GM, brain structural measures, and cognitive decline.

Fig. 4
Fig. 4
Full size image

Diagram of mediation analyses among DTI metrics in WM, brain atrophy measures, and CDR-SOB. Mediation analyses were performed using DTI metrics as a predictor, brain atrophy measures as a mediator, and CDR-SOB as an outcome. Significance levels are follows: *p < 0.05, **p < 0.01, ***p < 0.001. FA, fractional anisotropy; MD, mean diffusivity; CDR-SOB, clinical dementia rating sum of boxes; WM, white matter.

Fig. 5
Fig. 5
Full size image

Diagram of mediation analyses among DTI metrics in GM, brain atrophy measures, and CDR-SOB. Mediation analyses were performed using DTI metrics as a predictor, brain atrophy measures as a mediator, and CDR-SOB as an outcome. Significance levels are follows: *p < 0.05, **p < 0.01, ***p < 0.001. FA, fractional anisotropy; MD, mean diffusivity; CDR-SOB, clinical dementia rating sum of boxes; GM, gray matter.

For the DTI metrics in WM (Fig. 4), the relationship between the FAt and MDt metrics and annual CDR-SOB change was fully mediated by cortical thickness (FAt: p for indirect effect < 0.001, p for direct effect = 0.38; MDt: p for indirect effect < 0.001, p for direct effect = 0.43) and hippocampal volume (FAt: p for indirect effect < 0.001, p for direct effect = 0.39; MDt: p for indirect effect < 0.001, p for direct effect = 0.93). In contrast, the relationship between the FW and annual CDR-SOB change was partially mediated by brain structural measures, including cortical thickness (p for indirect effect < 0.001, p for direct effect < 0.01) and hippocampal volume (p for indirect effect < 0.001, p for direct effect < 0.01).

For the DTI metrics in GM (Fig. 5), the relationship between the FAt and annual CDR-SOB change was fully mediated by cortical thickness (p for indirect effect < 0.001, p for direct effect = 0.15) and partially mediated by hippocampal volume (p for indirect effect < 0.001, p for direct effect = 0.02). The relationships between the MDt and FW metrics and annual CDR-SOB change were partially mediated by cortical thickness (MDt: p for indirect effect = 0.01, p for direct effect < 0.001; FW: p for indirect effect = 0.01, p for direct effect < 0.001) and hippocampal volume (MDt: p for indirect effect < 0.001, p for direct effect < 0.001; FW: p for indirect effect < 0.001, p for direct effect < 0.001). Supplementary Figs. S7 and S8 present the results of additional mediation analyses that included HTN and DM as covariates. The findings were largely consistent with those from the original models, indicating that the inclusion of HTN and DM did not substantially alter the observed mediation effects.

Discussion

In the present study, we investigated the role of FW on DTI in understanding the pathophysiology of AD, with a particular focus on its relationships with plasma biomarkers, structural changes on MRI, and cognitive decline. Our major findings were as follows. First, core biomarkers, including Aβ and pTau217, showed significant associations with FW, but not with FA or MD, in both WM and GM. In contrast, plasma biomarkers of non-specific AD-related processes, such as GFAP and NfL, were associated with all DTI metrics, including FW, FA, and MD, in both WM and GM. Second, in both WM and GM, all DTI metrics’ abnormalities, representing unspecific microstructural changes, were associated with macrostructural changes such as reduced hippocampal volume or cortical thinning. Finally, these macrostructural changes, in turn, predict cognitive decline, highlighting their critical role in linking microstructural integrity with clinical outcomes. Taken together, our findings suggest that FW on DTI reflects a continuum of pathological processes in AD, linking plasma biomarkers and microstructural changes to macrostructural alterations and subsequent cognitive decline. These findings underscore the utility of FW as a sensitive biomarker for detecting early pathological changes, predicting cognitive decline, and guiding targeted interventions through integration with structural MRI and plasma biomarkers.

Our first major finding was that core biomarkers such as Aβ and pTau217 were significantly associated with FW in both WM and GM, but not with FA or MD. This suggests that these biomarkers primarily reflect extracellular changes41,42,43,44 including neuroinflammation and tissue atrophy, rather than axonal or microstructural integrity, indicating their close link to extracellular fluid dynamics and neuroinflammatory processes. These findings align with previous studies, such as Ji et al.45 and Nakaya et al.46 which reported significant associations between FW, neuroinflammation, and tau pathology in AD45,46,47. Our study extends this understanding by demonstrating the specificity of these associations to FW, highlighting its sensitivity to extracellular alterations like neuroinflammation and synaptic dysfunction. While FW is sensitive to extracellular fluid changes associated with neuroinflammation, it may also capture neurodegenerative processes such as Wallerian degeneration and myelin breakdown16,48. Future work using multimodal imaging and region-specific analyses is needed to disentangle the contributions of inflammation versus degeneration to FW alterations. In contrast, FA and MD, which reflect axonal density and fiber organization, may remain relatively unaffected during early or extracellular-dominant stages of pathology. These results emphasize FW’s utility as a biomarker for detecting extracellular changes and clarifying the role of core biomarkers in AD.

In contrast, plasma biomarkers of non-specific processes involved in AD pathophysiology, such as GFAP and NfL, were associated with all DTI metrics, reflecting their relationship with microstructural changes, in both WM and GM. Their associations with all DTI metrics suggest that these biomarkers capture a broader spectrum of microstructural changes, encompassing both extracellular alterations measured by FW and intracellular disruptions reflected by FA and MD. Our findings align with previous studies, such as Ji et al.45 which demonstrated the sensitivity of FW to neuroinflammation mediated by GFAP, and Nakaya et al.46 which reported the association of NfL with axonal damage. Together, these results could be explained by the concept that GFAP and NfL represent secondary and unspecific processes following core biomarker-mediated pathology, with GFAP reflecting neuroinflammation and NfL indicating neurodegeneration and axonal damage. While consistent with existing findings, our study extends the current understanding by demonstrating that GFAP and NfL are associated with both extracellular changes, captured by FW, and intracellular disruptions, reflected by FA and MD, across WM and GM. This broader role of non-specific biomarkers highlights their ability to capture the full spectrum of microstructural changes, offering a more comprehensive view of AD pathophysiology beyond core biomarker-driven processes.

Our second major finding was that in both WM and GM, all DTI metrics’ abnormalities, representing microstructural changes49 are associated with macrostructural changes such as hippocampal volume loss or cortical thinning50,51 bridging the gap between cellular-level changes and large-scale structural damage. These findings suggest that microstructural disruptions, such as neuroinflammation and axonal degeneration, precede and potentially drive macrostructural changes in AD52. In WM, processes such as Wallerian degeneration (anterograde axonal degeneration following upstream injury)10 or the dying-back phenomenon (distal axonal loss preceding cell body loss)53 may propagate damage to connected GM regions. FW, reflecting neuroinflammation, indicates extracellular matrix disruptions, while FA and MD abnormalities suggest axonal degeneration and demyelination, which impair structural connectivity and contribute to secondary GM atrophy. In GM, FW highlights extracellular changes linked to neuroinflammation54 while FA and MD reflect intracellular disruptions49 such as neuronal shrinkage or synaptic loss, directly driving cortical thinning and hippocampal volume loss. Together, these findings underscore distinct but interconnected pathways: in WM, axonal degeneration impacts connected GM regions, while in GM, localized microstructural changes directly lead to volume loss. This highlights the importance of assessing both WM and GM to understand the cascading effects of AD pathology.

Our final major finding was that the relationship between DTI metrics and cognitive decline was mediated by brain structural measures, with distinct patterns in WM and GM. Especially, FW in both WM and GM exhibited both indirect and direct effects on cognitive decline, suggesting that its influence extends beyond its impact on structural changes. This direct effect of FW highlights the potential role of neuroinflammation in driving cognitive impairment, independent of macrostructural degeneration. This finding aligns with evidence that neuroinflammation disrupts neural function independently of structural degeneration by impairing synaptic integrity, altering extracellular fluid homeostasis, and inducing chronic gliosis55,56,57,58,59. In contrast, FA and MD primarily contributed to cognitive decline through their effects on brain structural measures. In WM, FA and MD were fully mediated by cortical thickness and hippocampal volume, indicating that axonal integrity and diffusivity influence cognition indirectly via macrostructural changes. In GM, FA was fully mediated by cortical thickness and partially by hippocampal volume, while MD was partially mediated by both cortical thickness and hippocampal volume. These findings suggest that while axonal and cellular integrity are key correlates of structural brain changes, FW uniquely captures neuroinflammatory processes that may be more proximally associated with cognitive decline. Microstructural alterations—captured by changes in FW, FA, and MD—reflect underlying processes such as neuroinflammation, axonal degeneration, and demyelination. These processes compromise extracellular matrix integrity and axonal transport, leading to neuronal loss and synaptic pruning. Over time, such cumulative microstructural damage manifests as measurable reductions in cortical thickness and hippocampal volume. Our mediation analyses thus link these cellular- and tissue-level pathologies to large-scale structural degeneration in AD.

A major strength of our study is the comprehensive integration of plasma biomarkers, DTI metrics, and structural MRI measures to elucidate the differential contributions of microstructural and macrostructural abnormalities to cognitive decline in AD. However, several limitations should be acknowledged. First, one limitation of this study is the use of a single non-zero b-value (600 s/mm2), which may reduce sensitivity to free-water components. Although multi-shell or higher b-value protocols are optimal16,60 our single-shell approach with spatial regularization produced plausible FW maps and significant group effects. Future studies should employ multi-shell acquisitions for improved free-water modeling. Second, we relied on global DTI metrics for association and mediation analyses, which, while robust, may obscure regional microstructural variations. Moreover, FW is a non-specific indicator of extracellular alterations, and future studies should apply voxel-wise or ROI-based approaches to capture anatomical specificity and disentangle distinct pathological processes. Third, although cognitive trajectories were modeled longitudinally, all plasma biomarkers and imaging measures—including DTI metrics—were obtained cross-sectionally. As a result, causal inferences linking neuroinflammation, structural brain changes, and cognitive decline remain limited. Fourth, we included participants aged 45 years and older to capture early midlife microstructural and inflammatory changes that may precede clinical onset of AD. While this broader age range increases the study’s generalizability to preclinical stages, it may also introduce heterogeneity compared to cohorts restricted to typical late-onset AD. Future studies could stratify analyses by age groups to confirm the reproducibility of our findings in older samples. Fifth, while FW is a widely used marker of neuroinflammation15,16,17,18,19 it is not specific to neuroinflammation alone and may also reflect other extracellular alterations such as edema or vascular dysfunction. In our sensitivity analyses additionally adjusting for vascular risk factors including hypertension and diabetes mellitus, the main results remained unchanged, suggesting that FW alterations are not solely attributable to vascular dysfunction. Nevertheless, the potential for residual confounding by unmeasured vascular or other systemic factors cannot be entirely excluded, and the non-specificity of FW should be considered when interpreting these findings. Finally, despite our efforts to control for confounders, unmeasured variables such as comorbid cerebrovascular disease may have influenced the observed associations. Nevertheless, our findings provide novel insights into the role of neuroinflammation in AD progression, highlighting FW as a potential biomarker that bridges microstructural disruption, structural measures, and cognitive decline, and underscoring its relevance for both disease monitoring and therapeutic targeting.

Conclusions

Our integrated analysis combining plasma biomarkers, microstructural DTI metrics, and macrostructural MRI measures provides a comprehensive framework for understanding the continuum of pathological changes in AD. Notably, our findings reveal that extracellular alterations—reflected by increased FW—are closely linked to core AD pathologies and cognitive decline, with brain structural measures mediating these relationships. These insights underscore the promise of multi-modal biomarker integration for advancing early diagnosis and targeted therapeutic strategies in AD.