This study recruited 81 right-handed elderly participants from Beijing Tiantan Hospital in 2023, comprising 35 patients with CSVD and 46 age-matched healthy controls (HC). The inclusion criteria for CSVD patients were: (1) age 45 to 75 years; (2) fazekas score $>$2 (range: 0–6), or Fazekas score = 2 with at least two cerebrovascular risk factors (e.g., smoking, drinking, hypertension, hyperlipemia, or hyperglycemia). For the HC group, inclusion criteria were: (1) age 45 to 75 years; (2) no history of stroke; (3) absence of major diseases affecting vital organs (e.g., liver, heart, or lungs); (4) no history of neurological or psychiatric disorders; (5) fazekas score 2, or Fazekas score = 2 with no more than one cerebrovascular risk factors.

Exclusion criteria for all participants included: (1) acute ischemic stroke with high signal intensity on diffusion-weighted imaging (DWI) and lesions diameter $>$20 mm; (2) acute intracerebral hemorrhage; (3) acute subarachnoid hemorrhage or a history of subarachnoid hemorrhage due to vascular malformation or aneurysm, or the presence of an untreated aneurysm ($>$3 mm in diameter); (4) confirmed diagnosis of neurodegenerative diseases (e.g., Alzheimer’s disease or Parkinson’s disease); (5) non-vascular white matter lesions (e.g., multiple sclerosis, adult-onset leukoencephalopathy, or metabolic encephalopathy); (6) psychiatric conditions diagnosed according to diagnostic and statistical manual of mental disorders, fifth edition (DSM-V) criteria; (7) contraindications for MRI (e.g., claustrophobia); (8) severe systematic diseases (e.g., malignant tumors with a life expectancy 5 years); (9) current participation in other clinical trials.

Due to incomplete resting-state fMRI data and the presence of old hemorrhagic lesions, one participant from the CSVD group and one from the HC group were excluded. Consequently, the final sample consisted of 34 CSVD patients and 45 HC individuals (see Table 1 for demographic details). This cross-sectional study was approved by the Ethics Committee of Beijing Tiantan Hospital, Capital Medical University (approval number: KY 2019-140-02; approval date: January 3, 2020). The study was carried out in accordance with the guidelines of the Declaration of Helsinki and written informed consent was obtained from all participants.

All participants underwent comprehensive cognitive assessments, including standardized neuropsychological tests and behavioral tasks. The neuropsychological battery comprised the mini-mental state examination (MMSE), Montreal cognitive assessment (MoCA), auditory verbal learning test (AVLT), digit span task (DST), alternative shape trail-making test (STT), Hamilton anxiety scale (HAMA), and Hamilton depression scale (HAMD). Participants subsequently completed three computerized cognitive tasks in a fixed order: a working memory task (526 s), a color-word Stroop task (466 s), and a number-arrow category switching task (466 s) (Fig. 1A). The tasks were presented in a block design and implemented using E-Prime 3.0 (Psychology Software Tools, Pittsburgh, PA, USA). Each of the three tasks consisted of one run containing six blocks. Prior to formal testing, participants completed a practice session requiring a minimum accuracy rate of 85% within one hour [25]. The images used in the practice session were not appear in the formal experiment to prevent learning effects.

Fig. 1.

Overview and performance of the three behavioral tasks. (A) The flow of a single trial for each task. (B) Performance of each sub-components. Updating was measured by the working memory task, inhibition by the color-word Stroop task, and shifting by the number-arrow switching task. ACC, accuracy; CSVD, cerebral small vessel disease; HC, health control; ITI, intertrial interval; ns, not significant; RT, reaction time. * p 0.05, ** p 0.01, *** p 0.001.

The tasks were designed to assess three core cognitive domains: updating, inhibition, and shifting [26, 27, 28, 29]. Specifically, the working memory task evaluate working memory capacity, providing measures of short-term memory retention and cognitive control. The color-word Stroop task assessed cognitive flexibility and inhibitory control through measures of interference processing and conflict resolution. The number-arrow switching task measured cognitive flexibility and attentional shifting by evaluating performance during transitions between different information types.

MRI data were acquired using a 3T scanner (Siemens MAGNETOM Prisma, Erlangen, Germany) at the Tiantan neuroimaging center, with a total acquisition time of 23 minutes 8 seconds. Participants wore noise-reducing earplugs and foam padding to minimize head motion and scanner noise. They were instructed to remain awake with eyes open, avoid systematic thinking, and maintain head position throughout the scan. Structural images were obtained using a sagittal 3D magnetization-prepared rapid gradient-echo (MPRAGE) T1-weighed sequence with the following parameters: repetition time (TR) = 2300 ms, echo time (TE) = 2.26 ms, inversion time (TI) = 900 ms, flip angle = 8^∘, matrix size = 256 $\times$ 256, voxel size = 1 $\times$ 1 $\times$ 1 mm³, slice thickness = 1 mm, and 192 slices. Resting-state fMRI data were acquired using a T2*-weighted gradient-echo planar imaging (EPI) sequence with the following parameters: TR = 750 ms, TE= 30 ms, flip angle = 54^∘, field of view (FOV) = 222 mm, matrix size = 74 $\times$ 74, voxel size = 3 $\times$ 3 $\times$ 3 mm³, slice thickness = 3 mm, 48 slices, and 480 volumes.

To ensure objective assessment, experienced neurologists evaluated CSVD markers on MRI scans. The evaluation included: WMH from FLAIR images, brain atrophy from T1-weighted images (T1-WI) using a five-point scale (0–4), cerebral microbleeds (CMB) from susceptibility-weighted imaging (SWI), perivascular space (PVS) from T2-weighted images (T2-WI), and lacunes from T2-WI and FLAIR images. WMH severity was rated using the Fazekas scale (0–6 points), with separate scores for periventricular WMH (pWMH, 0–3 points) and deep WMH (dWMH, 0–3 points) [31]. PVS were categorized as total PVS (tPVS), centrum semiovale PVS (cPVS), and basal ganglia PVS (bPVS), while lacunes and CMB were quantified using standardized counting methods.

T1-WI images were preprocessing using the Computational Anatomy Toolbox (CAT12, https://www.nitrc.org/projects/cat) implemented in Statistical Parametric Mapping (SPM12, https://www.fil.ion.ucl.ac.uk/spm) on MATLAB R2021b (MathWorks, Natick, MA, USA). The preprocessing pipeline included: (1) Tissue segmentation using the “Segment” tool in CAT12, which performed skull-stripping segmented 3D T1 images into white matter (WM), grey matter (GM) and cerebrospinal fluid (CSF) using an adaptive maximum a posterior (AMAP) technique that dose not require priori tissue probabilities information [32]. (2) Spatial registration of segmented images to the standard montreal neurological institute (MNI) template. (3) Modulation to account for volume changes during spatial normalization [33]. Total intracranial volume (TIV) and regional volumes (WM, GM, CSF) were extracted using the ‘get TIV’ tool in CAT12. Gray matter volume percentage (GMV%) was calculated as (grey matter volume/TIV) $\times$ 100%. WMH volumes, including total WMHs (tWMH), pWMH, and dWMH, were extracted using the FSL-based UBO detector (https://cheba.unsw.edu.au/research-groups/neuroimaging/pipeline).

Resting-state fMRI data were preprocessed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm) and the Data Processing & Analysis for Brain Imaging toolbox (DPABI 5.1, http://rfmri.org/dpabi) [34]. The preprocessing steps included: (1) removal of the first 10 volumes to account for magnetic field stabilization; (2) slice timing correction and realignment for head motion correction, with motion parameters quantified using framewise displacement (FD) [35]; (3) segmentation T1-WI into GM, WM, and CSF using the New Segment and DARTEL tool in DPABI, followed by co-registration of T1-weighted and functional data; (4) nuisance regression, including (A) motion parameters (Friston-24 model) [36], (B) polynomial trend, (C) WM and CSF signals; (5) scrubbing of time points with FD $>$0.5, as well as the preceding time point before and subsequent two time points; (6) spatial normalization to MNI space using DARTEL (resampled to 3 $\times$ 3 $\times$ 3 mm³); (7) spatial smoothing with a 6 mm full width at half-maximum Gaussian kernel; (8) bandpass filtering for three frequency bands: typical band (0.01–0.08 Hz), slow-5 (0.01–0.027 Hz), and slow-4 (0.027–0.073 Hz) [22, 37, 38]. Filtered data were used for subsequent analyses.

Preprocessed time series data were transformed to the frequency domain using Fast Fourier Transform (FFT). The ALFF was calculated as the averaged square root of the power spectrum for each frequency band at the voxel level [39]. Individual zALFF maps were converted to z-scores for standardization and used in subsequent analyses. Whole-brain FC was computed for each frequency band using unsmoothed preprocessed data. Network nodes were defined according to the 246-region Human Brainnetome Atlas (BNA246; http://atlas.brainnetome.org/) and Pearson correlation coefficients between all nodes paires were calculated to generate 246 $\times$ 246 FC matrices for the three frequency bands. Network topological properties were analyzed using the Graph Theoretical Network Analysis toolbox (GRETNA, https://www.nitrc.org/projects/gretna/), including small-worldness (Sigma), global efficiency (Eglob), local efficiency (Eloc), and hierarchy [40]. Given that negative correlations exhibit greater variability [41] and lower test-retest reliability [42], with ambiguous biological explanations [40, 43], only positive correlations were retained for analysis. The area under the curve (AUC) for each network metric was calculated across a sparsity range of 0.03 to 0.4 (increment = 0.01) to provide threshold-independent measures. Brain networks are routinely compared against random networks to determine whether their topological organization significantly deviates from random configuration [40]. These randomly generated networks serve as null models for hypothesis testing in brain network analyses. We selected the weighted network mode and generated 100 random networks using the Markov chain wiring algorithm, ensuring that the generated networks had the same number of nodes, edges, and degree distribution as the real brain network [40, 44]. For each participant, we obtained AUC values corresponding to each network topological parameter, which were subsequently used for between-group comparisons.

Table 1 summarizes the demographic, clinical and cognitive characteristics of participants. No significant differences were observed between CSVD and HC groups in age, education level, hyperlipemia prevalence, tPVS, cPVS, or GMV%. However, the groups differed significantly in gender and mean FD (p 0.05), leading to the inclusion of age, gender, and mean FD as covariates in subsequent analyses. CSVD patients showed significantly higher prevalence of four cerebrovascular risk factors (smoking, drinking, hypertension, and hyperglycemia) compared to HC. In terms of imaging biomarkers, the tWMH Fazekas score (t₍₇₇₎= 9.422, p 0.001), pWMH (t₍₇₇₎ = 6.614, p 0.001), dWMH (t₍₇₇₎ = 8.708, p 0.001), brain atrophy (t₍₇₇₎ = 2.290, p = 0.025), CMB (t₍₇₇₎ = 2.588, p = 0.012), and bPVS (t₍₇₇₎ = 2.370, p = 0.020) were significantly higher in the CSVD group compared to the HC group. The Fazekas score showed significant correlation with WMH volume (Supplementary Fig. 3).

In the three cognitive tasks, the efficiency z-scores for the working memory task (updating; t₍₇₇₎ = –3.593, p 0.001), color-word Stroop task (inhibition; t₍₇₇₎ = –2.897, p = 0.005), and the number-arrow category switching task (shifting; t₍₇₇₎ = –2.576, p = 0.012) were significantly lower in the CSVD group compared to the HC group (Fig. 1). In the cognitive questionnaires, the MoCA (t₍₆₂₎ = –2.619, p = 0.011), AVLT-immediate memory (t₍₆₁₎ = –2.021, p = 0.048), AVLT-recognition (t₍₆₁₎ = –2.308, p = 0.024), and STT-B (t₍₆₁₎ = 2.527, p = 0.014), were significantly lower in the CSVD group compared to the HC group. These findings suggest that behavioral tests may be more sensitive than traditional questionnaires in detecting cognitive impairment in CSVD patients.

In the typical band, the CSVD group exhibited decreased ALFF in the right supra-marginal gyrus (SMG), and increased ALFF in the left inferior/middle temporal gyrus (ITG/MTG) and the right inferior temporal gyrus/fusiform gyrus (ITG/FFG) (corrected p 0.05). In the slow-4 band, the CSVD group exhibited decreased ALFF in the right SMG and increased ALFF in the left ITG and right FFG (corrected p 0.05). In the slow-5 band, the CSVD group exhibited increased ALFF in the left ITG, right ITG/MTG, and right FFG (corrected p 0.05). The spatial distribution of these changes is illustrated in Fig. 2, with detailed statistical results provided in Supplementary Table 1.

Fig. 2.

Regional functional changes (ALFF) in the typical band, slow-4 band, and slow-5 band in CSVD compared to HC (5000 permutations, TFCE corrected, p 0.05, two tailed). ITG, inferior temporal gyrus; SMG, supramarginal gyrus; FFG, fusiform gyrus; ALFF, amplitude of low frequency fluctuations; TFCE, threshold-free cluster enhancement.

In the typical band, the CSVD group exhibited decreased aSigma (t₍₇₇₎ = –2.827, p = 0.006). In the slow-4 band, the CSVD group exhibited decreased aSigma (t₍₇₇₎ = –2.726, p = 0.008) and AUC of hierarchy (t₍₇₇₎ = –2.327, p = 0.023). No significant differences were found between the groups for other graph theory measures (AUC of Eglob and Eloc) across any of the three frequency bands. The results are shown in Fig. 3.

Fig. 3.

Topological properties changes in CSVD compared to HC. Areas-under-the-curve (AUC) were plotted according to sparsity (ranging from 0.03 to 0.4, interval = 0.01). Significant group differences were observed in small-worldness in the (A) typical band and (B) slow-4 band, and (C) hierarchy in the slow-4 band. The overall AUC value for (D) small-worldness and (E) hierarchy across three frequency bands between the two groups are shown. ns, not significant. * p 0.05, ** p 0.01.

Significant group differences were identified in three categories of measures through independent sample t-tests: (1) CSVD markers (tWMH, pWMH, dWMH, brain atrophy, CMB, bPVS), (2) functional indicators (ALFF in three frequency bands, aSigma in the typical and slow-4 band, AUC of hierarchy in the slow-4 band), and (3) global cognition (updating, inhibition, shifting). These measures were subsequently included in pairwise correlation analyses. The correlation coefficient matrix for all participants is presented in Fig. 4, with detailed coefficients and corresponding p-values provided in Supplementary Tables 2–4. Supplementary Table 2 displays correlations between CSVD markers and functional indicators, Supplementary Table 3 displays correlations between CSVD markers and cognitive performance, and Supplementary Table 4 displays correlations between functional indicators and cognitive performance. Scatter plots with fitted regression lines are shown in Supplementary Fig. 4, while the correlation matrix specific to the CSVD patients is presented in Supplementary Fig. 5. These findings suggest a close relationship among WMHs (tWMH, dWMH, pWMH), ALFF across the three frequency bands, and cognitive performance, as well as dWMH, aSigma in the slow-4 band, and cognitive performance.

Fig. 4.

Pearson’s correlations among CSVD markers, functional indicators, and global cognition. Significance was set at p 0.05 (uncorrected). Red indicates positive correlations, while blue indicates negative correlations. aSigma, AUC of small-worldness; GC, global cognition.

As shown in Fig. 5 and Supplementary Table 5, ALFF in the typical band (indirect effect = –0.225, 95% CI = [–0.432, –0.090], p = 0.001) (Fig. 5A), slow-4 band (indirect effect = –0.249, 95% CI = [–0.484, –0.105], p = 0.001) (Fig. 5B), and slow-5 band (indirect effect = –0.200, 95% CI = [–0.543, –0.043], p = 0.009) (Fig. 5C) significantly mediated the relationship between dWMH and GC. Similar mediation patterns were observed for tWMH and pWMH. However, WMH (tWMH, pWMH, dWMH) did not mediate the relationship between ALFF in any of the three frequency bands and GC. These results suggest that WMH may influence GC through regional functional brain activities.

Fig. 5.

Simple mediation analyses between dWMH, ALFF/aSigma, and global cognition. The association between dWMH and global cognition was significantly mediated by ALFF (A) in the bilateral ITG and right SMG in the typical band; (B) in the left ITG, right SMG, and right FFG in the slow-4 band; (C) in the bilateral ITG in the slow-5 band; and (D) by aSigma in the slow-4 band. * p 0.05, ** p 0.01. Statistically significant path coefficients and p-values are highlighted in bold.

Additionally, aSigma in the slow-4 band mediated the relationship between dWMH and GC (indirect effect = –0.076, 95% CI = [–0.211, –0.011], p = 0.018) (Fig. 5D). Conversely, the relationship between aSigma in the slow-4 band and GC cannot be mediated by dWMH. These results suggest that dWMH may affect GC through large-scale functional brain network organization in the slow-4 band. These simple mediation analyses collectively show that WMHs is associated with GC through frequency-specific mechanisms involving both regional brain activity and network topological property, particularly in the slow-4 band.

Based on the well-fitted simple mediation models, we constructed a chain mediation model to examine the relationship among dWMH, aSigma and ALFF in the slow-4 band, and GC (Supplementary Fig. 2A, model 5; parameters in Supplementary Table 6). The chain mediation results (model 5; Fig. 6) revealed three significant pathways: (1) dWMH negatively predicted aSigma ($\beta$ = –0.300, p = 0.005), (2) aSigma negatively predicted ALFF ($\beta$ = –0.259, p = 0.024), and (3) ALFF negatively predicted GC ($\beta$ = –0.440, p = 0.001). Bootstrap estimates confirmed a significant chain mediation effect through aSigma and ALFF ($\beta$ = –0.034, 95% CI = [–0.094, –0.010], p = 0.005), indicating that the indirect effect (dWMH$\to$aSigma$\to$ALFF$\to$GC) was significant. These results suggest that dWMH influences GC through sequential effects on large-scale network organization and regional brain activity in the slow-4 band.

Fig. 6.

A chain mediation model with aSigma and ALFF as mediators of the association between dWMH and global cognition (GC). ns, not significant. * p 0.05, ** p 0.01. Statistically significant path coefficients and p-values are highlighted in bold.

We further developed three alternative multiple mediation models (Supplementary Fig. 2B–D, models 6–8) to validate these findings. In model 6, the alternative pathway (dWMH$\to$ALFF$\to$aSigma$\to$GC) was not significant ($\beta$ = –0.024, 95% CI = [–0.104, 0.009], p = 0.127), supporting the unidirectional nature of the aSigma$\to$ALFF relationship observed in model 5. Model 7 revealed a significant alternative pathway (aSigma$\to$dWMH$\to$ALFF$\to$GC; $\beta$ = 0.059, 95% CI = [0.012, 0.171], p = 0.004). The parallel mediation model (model 8) revealed a significant indirect effect through ALFF (dWMH$\to$ALFF$\to$GC; $\beta$ = –0.228, 95% CI = [–0.468, –0.087], p = 0.001), but not through aSigma (dWMH$\to$aSigma$\to$GC; $\beta$ = –0.047, 95% CI = [–0.170, 0.013], p = 0.131). These results collectively highlight ALFF as a critical mediator in the pathway linking dWMH to cognitive impairment. All models showed good fit indices (Supplementary Table 7).

By dividing the typical band into slow-5 and slow-4, this study revealed distinct patterns of functional changes in CSVD patients across all three frequency bands. Specifically, decreased ALFF was observed in the right SMG, while increased ALFF was found in the bilateral ITG and right FFG. This pattern aligns with recent study reporting decreased ALFF in the posterior cortex and increased ALFF in deep gray matter nuclei and temporal lobe in CSVD patients [51]. Notably, regions showing increased ALFF (bilateral ITG and right FFG) demonstrated positive correlations with WMH and negative correlations with cognitive performance. Conversely, the right SMG, which showed decreased ALFF, exhibited negative correlations with WMH and positive correlations with cognitive performance. These changes suggest that both ALFF increases and decreases may represent either decompensatory mechanisms or functional impairment resulting from CSVD pathology [52].

The observed regional alterations have important functional implications. The SMG, a key node in the ventral attention network, plays critical roles in visual word recognition [53, 54, 55], and is specifically activated during executive function updating tasks in healthy young individuals [56]. The ITG, associated with higher-order cognitive functions including visual and language processing, verbal fluency, and emotional regulation [57, 58], shows synaptic deficits that correlate with cognitive test performance [57]. The FFG, critical for advanced visual functions such as face perception, object recognition, and reading [59], also demonstrated altered activity. These findings suggest that regions showing altered functional activity in CSVD are closely linked to cognitive domains, potentially serving as both biomarkers of cognitive impairment mechanisms and targets for neuromodulation interventions.

Small-worldness reflects the communication efficiency of functional brain networks, with higher values indicating greater clustering coefficients and shorter characteristic path lengths [60, 61]. CSVD-induced local structural damage exerts widespread effects on distant brain regions, impacting both functionality and network connectivity [2, 14]. Previous studies have demonstrated that high WMH burden leads to disrupted topological organization of intrinsic functional networks [62, 63, 64], with pWMH and dWMH exhibiting distinct functional, microstructural, and clinical correlates [65]. While the slow-5 band primarily reflects cortical neuronal activity, the slow-4 band is mor representative of basal ganglia function [22, 66]. This study found significantly reduced small-worldness in CSVD patients in both the typical and slow-4 band. Notably, small-worldness in the slow-4 band showed negative correlations with dWMH and positive correlations with cognitive performance, suggesting that reducted network efficiency in CSVD patients in particularly pronounced in this frequency band. These findings highlight the slow-4 band as a major contributor to variability within the typical band and emphasize the potential significance of dWMH in the overall WMH-related cognitive impairment [67].

The interpretation of altered small-worldness in CSVD remains debated, particularly regarding whether these changes represent functional compensation or impairment. Some studies suggested compensatory machanisms, with non-demented CSVD patients showing increased small-worldness compared to controls [68], potentially maintaining cognitive function through reducted modularity and increased network integration [69]. Conversely, other research demonstrated reduced small-worldness in CSVD structural networks, with these alterations correlating with deficits in executive function, attention, and processing speed [70]. Our findings support the latter perspective, indicating that reduced small-worldness in CSVD patients primarily represents functional impairment rather than compensatory reorganization.

WMH promote neurodegeneration, and understanding the WMH-cognition relationship provides crucial insights into cognitive impairment in CSVD [71]. However, the specific patterns of brain reorganization patterns and their relationship with cognitive impairment remain unclear. Our simple mediation analyses indicated that ALFF in all three frequency bands fully mediated the effects of WMH (including tWMH, pWMH, and dWMH) on EF, suggesting widespread association between WMH and regional brain activity throughout the brain and across multiple frequency bands. Furthermore, small-worldness in the slow-4 band completely mediated the relationship between dWMH and global cognition. Chain mediation analysis revealed that small-worldness and ALFF significantly mediate the effects of dWMH on global cognition, with the mediating pathway flowing from small-worldness to ALFF. While previous studies have primarily examined simple mediation models [13, 15, 16, 17, 18, 19, 20, 21, 72], our fingdings extend the WMH-brain function-cognition pathway, highlighting the complexity of WMH-related cognitive impairment mechanisms.

Two key insights emerge from these findings. Firstly, the mediation effects of regional brain activity and network topology suggest that WMH influence cognition indirectly through functional brain changes [10]. Pathological studies indicate that WMH may induce cognitive impairment through multiple mechanisms, including ischemia, hypoxia, hypoperfusion, immune activation, blood-brain barrier dysfunction, metabolic alterations, and glial injury [10, 73, 74]. These pathological changes may disrupt white matter integraty and alter blood-oxygen-level-dependent (BOLD) signal distribution [75], ultimately affecting cortical connectivity and function [76].

Secondly, the chain mediation effect occurs from global to local levels. Research indicates that changes in small-worldness may be associated with ALFF in specific brain regions. For instance, stroke patients showed decreased ALFF in the left superior frontal gyrus alongside reduced FC with other brain regions (right precentral gyrus, right postcentral gyrus, left inferior temporal gyrus, and right paracentral lobule), highlighting changes in network clustering and characteristic path length [77]. Our findings suggest that WMH primarily impaie global cognition by disrupting the overall organizational of functional networks, reducing communication efficiency, and subsequently causing abnormal increases or decreases local brain activity patterns that lead to cognitive deficits in CSVD.

This study has several limitations. First, our findings are based on non-demented CSVD patients (as determined by MoCA/MMSE score) with a relatively small sample size. The observed effects of dWMH on cognition and the mediating roles of small-worldness and ALFF in the slow-4 band require validation in larger cohorts and in CSVD populations with cognitive dysfunction or dementia [78]. Second, while we focused on three core components of global cognition using objective procedural tasks, our assessment did not encompass all cognitive domains. Given the extensive nature of CSVD-related cognitive impairment [79], future studies should implement more comprehensive and objective cognitive assessments. Third, this study identified macro-level neuroimaging evidence for the structure-function-behavior relationship in CSVD, the underlying pathological mechanisms remain unclear. Future research should integrate pathological processes to better differentiate WMH etiologies and uncover the mechanisms underlying clinical heterogeneity [11]. Fourth, the cross-sectional design limits causal inferences. Longitudinal studies are needed, particularly given that baseline WMH predicted cognitive impairment across diagnostic categories and dementia risk in mild cognitive impairment and post-stroke conditions [80], with WMH being one of the strongest predictors of cognitive function in CSVD patients over three-year follow-up [81]. Finally, while we identified a pathway from dWMH to global cognition through small-worldness and ALFF, the relationships are not merely unidirectional and potentially involve more complex interactions. Future studies should explore these intricate relationships in greater depth.