Longitudinal Degeneration of Microstructural and Structural Connectivity Patterns Following Stroke

Abstract

BACKGROUND:

Stroke causes neurological impairment through local plasticity mechanisms, such as synaptic sprouting, and large-scale network reorganization. Although poststroke Wallerian degeneration is well established, whether white matter (WM) tracts’ remodel over time to support behavioral recovery remains controversial. We investigated whether longitudinal changes in WM microstructure and structural connectivity are coupled to poststroke behavior, and how structural connectivity reorganization relates to local disconnection microstructure.

METHODS:

In this prospective cohort study (Washington University, St. Louis; 2016–2018), patients with first-time stroke underwent diffusion-weighted magnetic resonance imaging and behavioral assessments at 2 weeks and 3 months poststroke; healthy controls were assessed twice. Latent factorial analysis was applied to behavioral and microstructural data from diffusion tensor imaging and neurite orientation dispersion and density imaging. Primary measures were whole-brain structural connectivity gradient divergences computed from tractography, microstructural diffusion parameters extracted from both WM and cortical regions disconnected by the lesion, and behavioral factors. Linear mixed models (adjusted for sociodemographic values) assessed longitudinal changes, linear regression assessed brain-behavior associations, and machine learning algorithms assessed differences between healthy controls and stroke.

RESULTS:

Seventy-nine patients (age, 60±12 years; 43% female; diffusion-weighted magnetic resonance imaging: 48 at 2 weeks and 26 at 3 months) and 33 age-matched healthy controls were included. At 2 weeks, patients with stroke exhibited diffuse whole-brain alterations in WM organization. These widespread changes were closely linked to microstructural abnormalities in WM pathways disconnected by the lesion. Notably, WM degeneration progressed between 2 weeks and 3 months poststroke. Whole-brain structural connectivity alterations were associated with cognitive deficits acutely but not at 3 months, whereas microstructural parameters in disconnected tracts were primarily related to motor impairment.

CONCLUSIONS:

In patients with predominantly mild-severity stroke, we observed widespread and progressive alterations in WM microstructure that persists beyond the lesion and continues to evolve over time. Although acute structural disconnection relates to behavioral deficits, chronic WM degeneration appears largely uncoupled from functional recovery.

Stroke triggers neural changes extending from the lesion to remote areas. In the perilesional region, hypoperfusion and hypoxia can cause reversible damage,^1,2 accompanied by changes in cellular composition, synaptic connectivity, and gliosis.³ Importantly, stroke-related changes also occur remotely; cortical and subcortical regions undergo functional and structural changes linked to behavior.^4,5

These alterations can be understood within the brain connectome, the large-scale functional (postsynaptic) and structural (axonal) connectivity architecture. Previous works have described connectional diaschisis,⁶ a pattern of poststroke functional connectivity (FC) disruption, including reduced interhemispheric coupling and decreased intrahemispheric segregation.^7–11 Such FC changes explain behavioral impairments and, when they normalize over time, predict recovery.^5,12,13 Notably, subacute FC changes are correlated with changes in structural connectivity (SC).^14–16 Stroke leads to progressive white matter (WM) degeneration beyond the lesion site, with implications for behavioral outcomes.^17,18 Egorova-Brumley et al¹⁹ documented progressive WM degeneration, associated with worsening cognition. However, how such degeneration relates to lesion-induced disconnection and large-scale SC reorganization, and how these changes over time contribute to recovery, remains unclear. Prior studies examined structural disconnection (SDC) and behavior or microstructural degeneration without relating it to whole-brain connectome reorganization.^16,17,19

Here, we integrate these levels within a unified longitudinal framework combining connectivity, microstructural modeling, and behavioral trajectories with 3 main innovations. First, we quantify whole-brain connectivity reorganization via tractography-based disconnection analysis. We compute separately SC patterns intrahemispherically and interhemispherically with respect to the lesion because previous work has documented correlated yet distinct FC alterations within the damaged hemisphere and across hemispheres.^8,16 Second, we measure WM/gray matter (GM) microstructural changes (diffusion tensor imaging and neurite orientation dispersion and density imaging), in regions disconnected by the lesion, based on a growing literature on the importance of disconnection in accounting for poststroke behavioral deficits.^14,20,21 Third, we relate structural and microstructural dynamics to recovery trajectories across cognitive and motor domains, measured with an in-depth behavioral battery at 2 critical stages: subacutely (<2 weeks) and chronically (3 months), when most recovery has occurred.^22,23 We specifically ask: (1) do SC and microstructural properties show coupled longitudinal changes? and (2) do these changes explain interindividual differences in recovery? While animal studies support WM plasticity as a substrate for recovery,²⁴ and human data support a role for FC normalization,¹² the direct link between SC remodeling, microstructural changes, and behavior in humans remains unknown. Demonstrating such a relationship would provide novel evidence for the contribution of WM reorganization to stroke recovery, with implications for prognosis and therapeutic targeting. Conversely, a lack of association would suggest that behavioral improvements may rely on mechanisms beyond structural remodeling, a key finding with critical implications for guiding future intervention strategies targeting nonstructural mechanisms of recovery. Figure 1 shows the study workflow.

Figure 1.

Overview of the study’s analytical framework: structural connectivity and latent-microstructural factors, cross-sectional comparisons with controls, longitudinal tracking of degeneration, and correlations between brain measures and behavioral recovery. Patients with stroke and age-matched controls completed an extensive behavioral assessment (A), which was subsequently reduced to a set of core components (B). Multishell diffusion magnetic resonance imaging (MRI) data were preprocessed (C) and used for 2 complementary analyses: (1) diffusion data were used to derive whole-brain structural connectivity matrices via tractography. A diffusion embedding algorithm was then applied to compute connectivity gradients for intra hemispheric and interhemispheric connections at both the parcel and network levels (gradient values were summarized into a score capturing how much each patient’s gradients diverged from the reference gradient derived from the control group [with a maximum value of 1 indicating maximal deviation and 0 indicating perfect similarity]) and (2) from the same processed diffusion data, microstructural models (diffusion tensor imaging and neurite orientation dispersion and density imaging) were computed, and the resulting maps were entered into a factorial analysis to quantify microstructural alterations in white matter (WM) tracts disconnected by the lesion and in contralesional homotopic gray matter (GM) regions (D). Several analyses were performed to assess: (1) differences between controls and patients with stroke at the subacute stage in both whole-brain tractography and the microstructural profile of disconnected tracts; (2) longitudinal changes between the subacute stage and the 3-month follow-up in both whole-brain tractography and disconnected tract microstructure; and (3) the relationship between microstructure and whole-brain connectivity, as well as between brain measures and behavior, both acutely and at follow-up, using the behavioral latent factors (E). 3-mos indicates 3 months; AD, axial diffusivity; FA, fractional anisotropy; ICVF, intracellular volume fraction; ISOVF, isotropic volume fraction; MD, mean diffusivity; ODI, orientation dispersion index; RD, radial diffusivity; SEM, structural equation modeling; SVM, support vector machine; UMAP, uniform manifold approximation and projection.

Methods

Data Availability

Data used for this study are publicly available at www.humanconnectome.org/study/hcp-young-adult/data-releases. The clinical data are available upon request by the corresponding author.

For this prospective longitudinal cohort study with a parallel healthy control (HC) group, patients with first-ever ischemic or hemorrhagic stroke were enrolled at Washington University in St. Louis (2016–2018) and followed at 2 weeks, 3 months, and 1 year (not reported). Inclusion required age ≥18 years, clinical stability, and objective neurological deficits; exclusions were multifocal lesions, reduced vigilance, major neurological/psychiatric, systemic comorbidities, or life expectancy <1 year (full criteria in Supplemental Section S1.1). Age-, sex-, and education-matched HCs were recruited among patients’ relatives during the same period and assessed twice (1 month apart). Matching was performed at the group level. Patients and HC underwent identical behavioral and magnetic resonance imaging (MRI) assessments (baseline and follow-up). Written informed consent was obtained from all participants per institutional review board procedures at Washington University. This article follows the STROBE reporting guidelines (Strengthening the Reporting of Observational Studies).²⁵ Figure S1 shows the enrollment diagram.

Participants underwent an extensive behavioral battery, administered typically on the same day as the imaging session (lasting ≈2.5 hours), representing an update of a validated battery previously used,²⁶ and assessing memory, language, visuospatial attention, executive, and motor functions (Supplemental Sections S1.2.1 through S1.2.5). A dimensional reduction approach was applied through a hierarchical factor analysis with orthogonal (varimax) rotation (Supplemental Section S1.2.6). The resulting behavioral factors were compared between stroke and HC using a 2-sample t test, and between the subacute and chronic stroke phases with a paired t test. Stroke severity was quantified using the National Institutes of Health Stroke Scale score both at hospital admission and at the time of the subacute assessment. A 3T Siemens Prisma scanner collected neuroimaging data. Acquisition imaging parameters and preprocessing details are reported (Supplemental Section S1.3).

SC Gradients

Whole-brain SC patterns were assessed through gradient analysis, a technique allowing to identify continuous axes of variation in brain connectivity and capture the hierarchical organization of brain networks.²⁷ Details are given in Supplemental Section S1.4.1. In line with the literature,²⁸ we selected the first 3 principal gradients to provide a balanced representation of the connectivity pattern.

To quantify variability in the gradients, we applied a gradient divergence (GD) measure for each brain parcel, normalized between 0 and 1, with higher values indicative of connectivity patterns that deviate more strongly from controls (the GD formula is reported in Supplemental Section S1.4.2). Briefly, GD captures how far each brain parcel’s connectivity profile departs from the control-derived gradient space, with higher values indicating greater divergence from typical SC patterns. For the networkwise analysis, GDs were averaged based on the correspondence between the Schaefer parcellation and the Yeo 7-network template.²⁹ To visualize key connectivity features, we examined the relationship between the first 2 gradients by identifying extreme parcels (from HC) along the Cartesian axes within each network. Structural pathways from these points, for both intrahemispheric and interhemispheric connectivity, were traced using a normative streamline map.³⁰

Microstructural Diffusivity

Diffusion processed data were used to estimate both diffusion tensor imaging and neurite orientation dispersion and density imaging measurements, fed into a latent factorial analysis (Supplemental Section S1.5). To confirm the latent factorial analysis microstructural patterns, the same method was applied in an independent sample of n=1065 individuals from the Human Connectome Project (Supplemental Section S1.6).

Microstructural analyses focused on (1) contralesional GM and (2) WM tracts structurally disconnected from the lesion. For the contralesional GM, the binarized and Montreal Neurological Institute-registered lesion mask was flipped along the y axis. These regions of interest were segmented into GM and WM according to the Harvard-Oxford atlas. WM tracts were based on SDC patterns embedding the lesion into a normative template to assess the probability of disconnections within the normative connectome.²⁰ To visualize SDC patterns, we applied a principal component analysis, and components >10% variance were visualized. We tested disconnection probability thresholds of 40%, 60%, and 80%. To minimize partial volume effects, a shell region, created by inflating lesion borders by 4 voxels, was excluded from tracts.

Subacute Stroke Versus Controls

SC Gradients

To ensure the robustness of results, we performed different complementary analyses. First, we assessed whether the distributions of the relationship between the main gradients were different between cohorts by means of structural equation modeling (Supplemental Section S1.7.1). Specifically, we assessed whether group (HC; stroke) influenced the association between gradients at the network level. Intraconnectivity and interconnectivity SC gradients were analyzed independently and considered significant with a P_FDR<0.05. Second, we summarized gradient values across brain parcels, projecting data into a 2-dimensional space using the Uniform Manifold Approximation and Projection algorithm.³¹ For each participant, parcelwise gradient values were reduced to 2 components. A multivariate ANCOVA was performed to test group differences in this 2-dimensional space (controlling for sociodemographic variables). A support vector machine was used to classify stroke and controls based on Uniform Manifold Approximation and Projection embeddings (Supplemental Section S1.7.2).

Furthermore, we assessed GD parcel-level differences through a 1-sample t test in stroke (as the GD metric was derived from controls, thus preventing the possibility of an unbiased direct comparison between groups). Zero represents a perfect alignment with the control’s gradient templates, so positive t-scores indicate areas where the stroke group’s GD values differ from this baseline. For visualization, only parcels with t-scores exceeding T=5 were plotted, to highlight brain regions with the highest deviations from controls.

Finally, we performed a network-level analysis comparing stroke and controls using directly gradient values (in place of GD). For stroke, hemispheres ipsilesional to the lesion were flipped to the left, with the right hemisphere treated as contralesional. Each parcel was assigned to one of the Yeo network.²⁹ Absolute gradient values were computed per parcel to avoid cancelation of opposing values within the same network and then averaged at the network level. An ordinary least squares regression was used to compare stroke and controls, controlling for sociodemographic outcomes. We tested the main effects of diagnosis and network, as well as their interaction (diagnosis×network).

Microstructural Differences Between Stroke and Controls

Microstructural diffusion factors obtained from latent factorial analysis (hereafter termed latent microstructure) within SDC maps were compared between stroke and HC. We created a mean template from controls for the latent factor maps. For each patient, the corresponding SDC maps were used as a mask in the control’s template. Values within the individual and template maps were compared by means of a nonparametric independent test (Mann-Whitney U test; P<0.05). Moreover, we applied a voxelwise analysis comparing latent-microstructure factors between controls and subacute stroke (Supplemental Section S1.8).

Microstructural-Disconnection Correlates in Contralesional GM

We tested whether the disconnected GM also showed microstructural abnormalities. Pathological mechanisms in the disconnected GM tissue have been demonstrated in an animal model³² but never in humans. To achieve this aim, we assessed whether contralesional latent microstructure could predict the likelihood of the contralesional area being disconnected from the lesion. A linear regression analysis was applied, including as independent variables all the latent-microstructural values within the GM contralesional tissues and the structural degree of disconnection as the dependent variable. In addition, we investigated whether GM regions disconnected from the lesions exhibited worse microstructural values compared with unconnected regions (Supplemental Section S1.9).

Longitudinal Analysis

We used the linear mixed model to compare network-GD values longitudinally in stroke (subacute versus 3 months), with networks and individuals as random factors. Separate models were run for intrahemispheric and interhemispheric gradients in both ipsilesional and contralesional hemispheres (4 models). Similarly, we compared latent-microstructural factors over time within SDC masks. All LLM analyses included sociodemographic covariates. To assess the robustness of results, both sets of analyses were performed on the full data set (all-available-data analysis) and on patients with diffusion-weighted MRI data at both timepoints (complete-case analysis). The complete-case analysis was specifically conducted to verify whether the longitudinal effects were comparable when restricting the sample to patients who completed both baseline and 3-month scans, thereby reducing the possibility that the observed results were driven by systematic characteristics of participants who did not return for follow-up (eg, differences in disease severity).

Behavioral Correlates

We assessed the parcelwise relationships (Pearson) between behavioral latent factors with intrahemispheric and interhemispheric GD for subacute and chronic stages, separately. This analysis was performed independently for intraconnectivity and interconnectivity (P_FDR<0.05).

The relationship between latent-microstructural properties and behavioral performance was assessed by means of a linear regression (R² robustness was evaluated via 1000 bootstrap resamples [95% CI]). Latent-microstructural values within SDC and lesion masks were included as predictors for behavior. The analysis was repeated independently for subacute and chronic stages (P_FDR<0.05). Outliers exceeding 2× the interquartile ranges were removed to avoid inclusion of biased values. For SDC, we considered the more lenient threshold (40%).

Connectivity Organization and Microstructure

We examined the relationship (linear regression model) between gradient organization and microstructure, assessing whether deviations in GD from controls were linked to biological mechanisms in disconnected tracts. Microstructural factors within SDC tracts were related to GD separately for intrahemispheric and interhemispheric connectivity and for ipsilesional and contralesional hemispheres. A global GD value was calculated by averaging parcelwise GD. Latent-microstructural parameters in SDC-ROIs were used as predictors of global GD (P<0.05). We also conducted this analysis at the network level (P_FDR<0.05). All analyses were performed in patients with subacute stroke, using SDC maps 40% probability.

Sensitivity Analyses

Several sensitivity analyses were conducted to assess the robustness of the analyses, including analyses examining the potential differential impact of subcortical versus nonsubcortical (cortical or cerebellar) lesions on diffusion-derived metrics (Supplemental Section S1.10).

Results

Seventy-nine patients (age=60.1±11.5 years; education=13.3±2.5) and n=33 age-matched HC were included. National Institutes of Health Stroke Scale scores at evaluation indicated that most patients presented with mild impairment (mean, 4.5±4.5) although a minority of patients showed higher initial severity (Table). Patients and HC were matched for age (T=0.927; P=0.356), education (T=−0.897; P=0.372), and sex (X²=1.433; P=0.231). N=29 HC performed MRI exam (age=57±11 years; education=13.7±2.6), of whom n=22 repeated the MRI visit 1 month apart (session 2). A total of n=48 patients had available MRI data at the baseline visit (age=59±11 years; education=13.5±2.2), while n=26 repeated the exam at the 3-month follow-up (age=59±11 years; education=13.8±2.2). No differences were observed between the stroke baseline imaging cohort and the cohort of controls performing MRI at the first session (age: T=−0.55; P=0.58; education: T=0.38; P=0.71; and sex: X²=3.06; P=0.08). No significant differences were observed between patients who completed the 3-month MRI follow-up and those who did not in terms of age, sex, education, lesion volume, or baseline clinical severity (all comparisons P>0.39; Figure S2). Figure 2 shows the lesion frequency maps.

Figure 2.

Brain imaging maps used in the study. A, Lesion frequency map projected both at the surface and volume level. Most lesions mapped to the middle cerebral artery territory, with predominantly subcortical involvement, consistent with the mild-to-moderate severity of the cohort. B, Structural disconnection patterns (the first 2 main components are shown at the fiber level and 3-dimensional rendering). All analyses were performed on individually computed disconnection maps. These maps illustrate the main white matter pathways affected by the lesions. C, Intrahemispheric (left) and interhemispheric (right) connectivity gradients from the healthy control population (Washington University cohort). These gradient maps capture the principal axes of white matter organization and serve as the normative reference against which stroke-related deviations are measured. D, Microstructural latent diffusion maps at both volumetric (left) and surface (right) space (healthy controls of the Washington University cohort). Each latent factor summarizes a distinct aspect of tissue microstructure: latent diffusion factor (_dwiF) 1 reflects free-water content, _dwiF2 captures fiber orientation and anisotropy, and _dwiF3 relates to properties mainly associated with myelinated fiber architecture. dwi indicates diffusion-weighted imaging; and G, gradients.

Table.

Whole Cohort of Stroke Patients and Healthy Controls (Intention-to-Treat Sample)

Latent Behavioral Factors

Latent factorial analysis was used to assess the latent behavioral space. The Kaiser-Meyer-Olkin score (computed to evaluate the sampling adequacy for latent factorial analysis) showed a value suggestive of adequacy (0.813). We identified 5 main factors explaining ≈50% of the variance. The first factor loaded on verbal memory, language, and executive functions (13.3% variance; hereafter referred to as the language-verbal-memory). The second (12.2%) and third (8.1%) factors reflected right and left motor functions, respectively (right motor and left motor factor). The fourth (6.4%) captured visual memory, working memory, and general attention (visuospatial-memory-attention component); the fifth (6.38%) related to lateralized visuospatial performance on the Posner task (lateralized attention factor; Figure S3). Due to the left-right structure of this last factor, absolute values were used. The analysis repeated on the stroke-only (subacute) sample yielded similar results (Figure S4). These factors differed significantly between HC and patients with subacute stroke (P<0.05; Figure 3). At 3-month follow-up (n=33), patients improved in motor (left: P=0.046; right: P=0.002) and lateralized visual attention (P=0.030), while language-verbal memory and visuospatial-attention factors showed nonsignificant improvements (Figure S4).

Figure 3.

Latent factors of cognitive performance. A, Correlation matrix across all behavioral scores from the whole cohort of stroke and controls (a full description of the tests is reported in the Supplemental Material). The block-diagonal structure of the matrix reflects the clustering of behavioral tests into distinct but partially overlapping clinical domains (motor, language, memory, and attention). B, Differences between patients with stroke at the subacute stage (red) and controls (blue; session 1) in the behavioral latent factor scores. Asterisk marks significant group differences (P<0.05). Patients with stroke showed significant deficits across all 5 behavioral domains compared with controls, confirming cognitive and motor impairment. ARAT indicates Action Research Arm Test; and RT, reaction times.

Whole-Brain Patterns

Structural Connectivity

Figure 2 shows the surface representation of the SC gradients obtained in HC intrahemispherically (variance=52%) and interhemispherically (variance=71%). Gradients are data-driven axes of SC covariance that mathematically captures the principal organization of WM and describe variability along posteroanterior, ventral-dorsal, and callosal directions. In HC, the first intrahemispheric connectivity gradient (_intraG1) explained 22% variance, the second (_intraG2) 17%, and the third (_intraG3) 13%. For the interhemispheric connectivity, _interG1, _interG2, and _interG3 explained 32%, 24%, and 15% variance, respectively. Supplemental Section S2.1 provides an anatomic description of the gradients. Figure S5A shows the parcelwise top 10% gradient values.

The spatial patterns of these gradients were replicated in HC at both sessions (Figure S5B; parcelwise spatial correlation >0.992 across the gradients). Similar average-group patterns were found between patients with stroke (subacute) and controls, highlighting that gradient values captured similar connectivity axes in the 2 groups (Figure S6).

An intuitive understanding of the anatomic organization of the gradients emerges from their bivariate relationships. The relationship between _intraG1 and _intraG2 highlighted an SC pattern transitioning from a posterior-anterior axis (highest _intraG1-G2 values) to a superior-inferior pattern (lowest _intraG1-G2 values) and then returning to an anterior-posterior axis (intermediate _intraG1-G2 values), forming a virtual U-shaped trajectory in the embedded gradient space (Figure 4A) that reflects a continuous reorganization of dominant connectivity patterns rather than a discrete anatomic pathway. Tracts connected with the parcels with the highest _intraG1-G2 values include the geniculocalcarine tract or the inferior frontal-occipital fasciculus. In contrast, tracts connected with the lowest _intraG1-G2 values include the corticospinal tract.

Figure 4.

Structural connectivity gradients and microstructural factorial space. A, First (G1) and second (G2) gradient relation in controls (left) and stroke (right) for intrahemispheric connectivity. The U-shaped trajectory in controls describes a continuous transition of dominant connectivity patterns; in stroke, the wider scatter indicates disrupted connectome organization. B, Same as A, with G1-G2 interhemispheric interactions. Exemplary tracts are marked at extreme points of the Cartesian axes (left, right, up, and bottom). Colors denote the main functional networks analyzed in the study. C, Corpus callosum from the average template of the Human Connectome Project (HCP)-842 atlas superimposed on parcel-gradient values of controls from the Washington University cohort. This figure validates the biological plausibility of the interhemispheric gradients by demonstrating their correspondence with known callosal fiber anatomy. The _interG1-_interG2 relationship forms a spatial pattern overlapping the corpus callosum topography (transparent background behind black dots), aiding interpretation of interhemispheric structural gradients. D, Microstructural factorial analysis results in Washington University controls (session 1). E, Same as D, second session. Scatterplots (center) show voxel-level spatial correlations between session weights. The high cross-session correlations (r>0.94) confirm the test-retest reliability of the latent-microstructural space. F, Comparison between the third latent-microstructural factor (latent diffusion factor [_dwiF] 3) from the Human Connectome Project (HCP) cohort and myelin maps, supporting the interpretation of this factor as reflecting microstructural properties mainly associated with myelinated fiber pathways. Maps thresholded to retain 94% of the highest cortical values. DAN indicates dorsal-attention network; DMN, default mode network; FPN, frontoparietal network; LIM, limbic network; SMN, sensorimotor network; VAN, ventral-attention network; and VIS, visual network.

Notably, the relationship between _interG1 versus _interG2 captures the organization of callosal fibers from the anterior (highest _interG1 values) to the posterior (lowest _interG1 values) portion (Figure 4B). This interpretation was validated by comparing the _interG1 versus _interG2 values with the average corpus callosum fibers from the Human Connectome Project (Figure 4C).

In summary, the gradient structure allows for mapping in a compact way the whole-brain organization of the SC.

Microstructure

Diffusion tensor imaging and neurite orientation dispersion and density imaging maps from the 2 sessions of the HC showed high reproducibility (Figure S7A). From these maps, we reported 3 main latent diffusion factors (_dwiFs; Figure 2). The first factor (_dwiF1; 52% variance) loaded on axial diffusivity, radial diffusivity, mean diffusivity, and isotropic volume fraction. Higher values were predominantly observed in the cerebrospinal fluid, aligning with the loadings on diffusion metrics. The second factor (_dwiF2; 28% variance) loaded on fractional anisotropy and orientation dispersion index, with opposite directions, indicating a latent space capturing diffusivity orientation. Anatomically, divergent values were expressed in the GM and WM tissues, the latter showing lower values (higher anisotropy in the WM). The third factor (_dwiF3; 15% variance) loaded positively on both fractional anisotropy and intracellular volume fraction (Figure 4D and 4E). This factor showed a spatial distribution similar to the myelin map distribution from Glasser et al,³³ suggesting a microstructural pattern preferentially expressed along major WM pathways, which are enriched in highly myelinated fibers (Figure 4F). An additional analysis computing the latent space independently of the stroke cohort echoed these findings (Figure S7B).

The latent-microstructural space reported in the second session of HC was used to project microstructural data of patients with stroke. Due to the novelty of the reported patterns, we assess the robustness of this latent diffusion tensor imaging-neurite orientation dispersion and density imaging microstructural space in a large independent cohort (Human Connectome Project). The results were in line with the findings on the stroke data set and highly reproducible across the 2 independent data sets (Washington University and Human Connectome Project; Supplemental Section S2.2; Figure S8).

SC in Stroke

Cross-Sectional Analysis

First, we assessed whether the distribution of the relationship between gradients was similar between patients and HC. While, qualitatively, the pattern was similar between cohorts (Figure 4A and 4B), patients with stroke showed a more widespread distribution, with significant differences for both _intraG1-_intraG2 and _interG1-_interG2 mapping to different networks (Supplemental Section S2.3).

Differences between cohorts were confirmed by the multivariate ANCOVA analysis applied to the Uniform Manifold Approximation and Projection space for both intraconnectivity (ipsilesional hemisphere: F=6.70; P=0.0022; contralesional: F=3.43; P=0.0376) and interconnectivity (ipsilesional: F=8.51; P=0.0005; contralesional: F=4.01; P=0.0225). Furthermore, we assess the classification accuracy of the Uniform Manifold Approximation and Projection space, segregating controls and stroke. Support vector machine reached a classification accuracy of 89% and 96% for the interhemispheric and intrahemispheric SC gradients, respectively (Figure 5A). Details for the support vector machine hyperparameters’ validation are given in Supplemental Section S2.4.

Figure 5.

Cross-sectional and longitudinal structural connectivity gradient organization in stroke. A, Intrahemispheric and interhemispheric connectivity classifications (top) and parcel-level 1-sample t tests (bottom) are shown. The classification figures display the nth-dimensional gradient space projected into a 2-dimensional space, grouped by diagnostic labels (red: acute stroke; blue: controls-first session). Left, The grouping of the real data, while the right figures show the support vector machine classification accuracy for the same data distribution (accuracy is reported in the bottom box embedded in the plots). B, Parcelwise 1-sample t test highlights parcels with the highest values (dark red), indicating greater deviation of stroke data from the control template in terms of gradient divergence. The most affected parcels extend well beyond the middle cerebral artery territory, indicating that the structural impact of stroke is not confined to the region of the lesion. C, Longitudinal changes in the complete-case sample. Each dot represents the gradient divergence value at the network level for each individual included in the analysis. These longitudinal results show that structural connectivity continues to diverge from the normative pattern between 2 weeks and 3 months, indicating progressive white matter degeneration. UMAP indicates uniform manifold approximation and projection. **P<0.005; *P<0.01.

Finally, we performed a parcelwise finer analysis by means of a 1-sample t test applied to the GD. Although the lesions were heterogeneous in location, most of them concentrated in the middle cerebral artery distribution (Figure 2A); yet, the cortical parcels showing the greatest divergence from HC also involved regions away from the middle cerebral artery distribution (Figure 5B). Group differences were further confirmed at the network level when analyses were applied directly to gradient values rather than GD scores (Supplemental Section S2.5; Figure S9).

These findings indicate that subacute stroke lesions induce widespread alterations of the structural connectome, affecting both the ipsilesional and contralesional hemispheres.

Longitudinal Analysis

For the all-available-data sample, the linear mixed model analysis for the ipsilesional hemisphere showed a significant time effect for both _intraGD (T=3.406; P=0.001) and _interGD (T=3.035; P=0.003). A significant effect was also observed when considering the contralesional hemisphere (_intraGD: T=3.431; P=0.001; _interGD: T=2.667; P=0.008; Figure 5C). Results were confirmed in the complete-case sample (Tables S1 and S2). These findings indicate a chronic degeneration of the structural connectome from 2 weeks to 3 months.

Microstructural Properties Differences

Cross-Sectional Analysis

First, a qualitative description of the SDC in patients with stroke was performed through a descriptive principal component analysis. Two main factors (>66% variance) were reported, loading within the left and the right hemisphere, consistent with the lesion frequency map (Figure 2). SDC maps were used as masks for microstructural analysis.

We found a significant group difference for _dwiF2 (loading on fractional anisotropy and orientation dispersion index). The results were consistent across multiple SDC thresholds (60%: U=1277; P=0.033; 80%: U=1135; P=0.024), while the 40% threshold showed a trend toward significance (P=0.06). No significant differences emerged for the other factors (P>0.6; Figure 6A).

Figure 6.

Cross-sectional and longitudinal microstructure. A, Microstructure differences in stroke vs controls (*significance). The selective involvement of latent diffusion factor (_dwiF) 2 indicates that, subacutely, microstructural disorganization in fiber orientation and anisotropy of disconnected tracts is the dominant abnormality. B, Linear model linking microstructure to subacute contralesional disconnection (observed vs predicted); subfigure shows bootstrap R². This result demonstrates a link between remote tissue damage and disconnection. C, Microstructure in disconnected vs not disconnected parcels, iteratively tested by randomly selecting the least disconnected parcel. Each plot shows the P value associated with each iteration. Significant P values are in the green zone, nonsignificant in the red. The bar plot shows the difference in _dwiF2 values (y axis) between the maximally disconnected parcel and the iteration-average minimally disconnected parcel for each patient (x axis). D, Longitudinal differences in microstructure between subacute and chronic stages; significance reported for all-available data and complete case (*significant; #trend). The progressive decrease of _dwiF3 from 2 weeks to 3 months indicates ongoing degeneration of microstructural properties, mainly associated with highly myelinated tracts in the disconnected white matter. dwi indicates diffusion-weighted imaging; and SDC, structural disconnection.

The specific involvement of _dwiF2 was confirmed by the exploratory voxelwise analysis, reporting higher _dwiF2 values in stroke, subacutely, compared with HC (p_FWE<0.05). No significant voxels were reported for _dwiF1 and _dwiF3 (Supplemental Section S2.6; Figure S10).

The linear regression analysis showed significant results at the subacute stage (_adjr2=0.625; P<0.0001). Specifically, higher disconnection probability was accompanied by lower _dwiF2 (coef: −0.244; P<0.001) and higher _dwiF3 (coef: 0.132; P=0.012), with a reversed (and expected) effect of these factors on GM compared with WM. Results were echoed by the bootstrapping procedure (F=21.65; P<0.0001; Figure 6B).

To generalize these findings, we ran a separate analysis in which we compared brain regions with the highest and the lowest probability of disconnections based on the SDC values. A significant effect was found for _dwiF2 (Figure 6C). The result was highly stable across the random selection of parcels with the lowest probability. Moreover, _dwiF2 showed the highest differences between parcels compared with _dwiF1 and _dwiF3 assessed with a 1-way ANOVA (F=347; P<0.00001; np²=0.32). Results were replicated when excluding the subcortical parcels from the analysis (F=26.3; P<0.00001; np²=0.15; Figure S11).

These results indicate a significant linear relationship between the degree of WM disconnection and the alteration of the diffusion properties of the disconnected GM.

Longitudinal Analysis

Longitudinal changes in SDC latent-microstructural parameters were significant for _dwiF3, consistently for the different probability thresholds in the all-available-data (40%: T=2.124 [95% CI, 2.5–97.5]=0.113–0.752; P=0.043; 60%: T=2.747 [95% CI, 2.5–97.5]=0.022–0.133; P=0.011; and 80%: T=2.128 [95% CI, 2.5–97.5]=0.006–0.150; P=0.043). SDC regions (independently of the threshold) showed, on average, lower values at the chronic stage compared with the subacute stage. Because _dwiF3 loads on the main myelinated tracts, its reduction chronically strongly supports longitudinal degeneration (Figure 6D; Table S1), potentially reflecting a combination of processes including demyelination, axonal degeneration, and secondary inflammatory or neurodegenerative mechanisms. Results were confirmed by the complete case (Table S2). A post hoc sensitivity analysis, evaluating whether microstructural latent factors computed at the individual level produced consistent results, confirmed the longitudinal degeneration findings (Supplemental Section S2.7).

Similar to the gradient analysis, the microstructural properties of the disconnected tracts further degenerate from the subacute to the chronic stage.

Correlations Between WM Alterations and Behavior

At the subacute stage, stroke-related changes in intrahemispheric GD were significantly associated with deficits in language-verbal memory, visuospatial-memory, and attention-neglect. Motor deficits were not significantly associated with GD (Figure 7A). At the chronic stage, we found no significant behavioral association with the intrahemispheric GD. A weak association was found for the interhemispheric GD with visuospatial-memory-attention and lateralized-neglect components (extended data are reported in Figure S12).

Figure 7.

Cognitive relationship with structural connectivity patterns and microstructural factors. A, Parcelwise cognitive-gradient correlation at the subacute stroke stage. The linear correlation between behavioral scores and gradient divergence (GD)values at the parcel level was computed for each behavioral factor. Only parcels surviving P<0.05 FDR correction are shown. Results are reported for both intraconnectivity (left) and interconnectivity (right) connectivity gradients. The chronic stage showed marginal significance, and results are reported in Figures S12. Subacutely, whole-brain structural connectivity alterations were significantly associated with cognitive deficits but not with motor impairment. The weakening of these associations at 3 months suggests a decoupling between ongoing structural degeneration and behavioral factors. B, Microstructural latent space and behavioral deficits at acute and chronic stages. Top, Linear relationship between behavioral factors and the microstructural factors within the region-of-interest considered by the study (disconnection maps) at the acute stage. Each boxplot represents the r² of the linear model predicting behavioral scores by means of the latent-microstructural values. Bottom, The same analysis was repeated at the chronic stage. In contrast to the gradient-behavior associations, latent-microstructural parameters in disconnected tracts were primarily related to motor deficits at both timepoints, reflecting the predominant involvement of corticospinal tract disconnection in this cohort. These results reveal a double dissociation: changes in whole-brain structural connectivity relate to cognitive deficits subacutely, while local disconnection microstructure preferentially tracks motor impairment. FDR indicates false discovery rate. **P<0.01 surviving multiple comparison significance.

Contrary to SC gradients, subacute motor left deficits were associated with latent-microstructural parameters of disconnected WM (r²=0.37; P_FDR=0.015 [95% CI, 0.146–0.690]). The remaining behavioral factors did not show significant associations (Figure 7B). The relationship between SDC latent-microstructural parameters and left motor deficits remained significant at the chronic stage (r²=0.42; P_FDR<0.005 [95% CI, 0.237–0.639]; Figure 7B). Chronically, there was also a significant association between the SDC latent microstructure with the visuospatial-memory-attention (P_FDR=0.020) and language-verbal-memory factors (P_FDR=0.020) although these associations were of lower magnitude (r²=0.20 with 95% CI, 0.091–0.394, and r²=0.19 with 95% CI, 0.095–0.391, respectively).

Relationship Between Whole-Brain Gradients and Local Disconnection

Previous studies have documented widespread changes in the WM poststroke but have not related these changes to the site of lesion. Hence, it is important to establish if the degree of local disconnection relates to the degree of global SC abnormalities.

For this analysis, only _dwiF2 and _dwiF3 were considered, given the marginal contribution of _dwiF1 to the results. Microstructural outcomes within the SDC ROI were significantly predictive of the global GD value (intrahemispheric contralesional: R²=0.186; P=0.020; intrahemispheric ipsilesional: R²=0.224; P=0.008; interhemispheric contralesional: R²=0.228; P=0.011; and interhemispheric ipsilesional: R²=0.167; P=0.041). As GD increased (higher dissimilarity from HC), in the disconnected WM tracts, _dwiF2 increased, and _dwiF3 decreased, indicating microstructural alterations compatible with axonal disorganization, while potentially also reflecting contributions from other biological processes such as edema or inflammatory responses (Figure S13; Table S3). At the network level, these results were largely confirmed (Figure S13).

This analysis links for the first time the degree of WM disconnection caused by stroke lesions to widespread changes in the structural connectome organization.

Sensitivity Analysis

These findings were confirmed by multiple sensitivity analyses (Supplemental Section S2.8; Figure S14).

Discussion

One of the most influential results in the past decade has been realizing the importance of brain networks’ disconnection caused by focal lesions.^12,34 Here, we demonstrated for the first time 3 related findings: progressive SC degeneration, relations between microstructural degeneration in disconnected tracts and connectome reorganization, and stage-specific dissociations in brain-behavior relationships between the subacute and follow-up stages.

Through gradient mapping,²⁷ we identified 3 SC gradient maps. In patients, these gradients showed, subacutely, widespread differences, in line with previous studies that found poststroke SC alterations^17,35 and FC changes.^5,12,34 The structural changes were large enough that a patient with stroke could be clearly distinguished from HC (accuracy around 90%). When the SC gradients were remeasured after 3 months, we showed a growing deviation from the HC’s template. This is interesting as local synaptic sprouting and remodeling of the WM have been proposed as a mechanism of recovery in animal models,^24,36 and most recovery has already occurred at 3 months.³⁷ However, our results are in line with other human studies, showing progressive poststroke structural degeneration.^18,19,38 The progressive structural WM degeneration observed here contrasts with prior studies showing FC normalization correlating with recovery of function.^13,39 Since, at the subacute stage, changes in FC are closely aligned with alterations of SC,¹⁶ it follows that structural and FC trajectories progressively diverge during recovery, with structure degenerating, while function normalizes. These findings are consistent with the idea that adaptive FC reorganization could contribute to behavioral recovery, especially in more redundant cognitive networks where the link between structure and function is weaker.⁴⁰ However, because FC was not directly assessed, this comparison should be interpreted as a contextualization within the existing literature. Similarly, it is important to highlight that gradients represent descriptive SC-covariance axes. Thus, rather than being interpreted as specific anatomic pathways, gradients provide a principled low-dimensional framework to summarize large-scale structural reorganization, enabling the tracking of widespread and progressive deviations in poststroke connectome organization.

Beyond SC, we described a latent-microstructural space showing high reliability in condensing complex diffusion metrics into stable components. Three _dwiFs were identified robustly in independent cohorts. When these latent diffusion microparameters were mapped to GM and WM regions identified as disconnected,²⁰ we found, subacutely, disorganization in disconnected regions. This pattern is consistent with decreased fractional anisotropy in WM, hence fiber disorganization, and decreased orientation dispersion index in the GM, hence loss of neurite dispersion suggestive of cortical deafferentation. Disconnected regions showed further degeneration from 2 weeks to 3 months, as indicated by a decrease in a measure suggestive of myelination. Interestingly, changes of latent microparameters in disconnected tissue partially but significantly explained the SC distance between stroke and HC. Microstructural latent factors predicted GD values in stroke, suggesting a complex relationship between SC and microstructural alterations. These findings suggest a cascade, whereby focal lesions drive microstructural degeneration in disconnected tracts and subsequent connectome reorganization. Despite ongoing structural decline, behavioral recovery occurred, indicating reliance on functional adaptation (likely via synaptic reweighting of surviving connections) rather than structural restoration.

We measured behavior using a large battery, finding 5 factors that explained about 50% of the behavioral variance, consistent with previous studies in focal and neurodegenerative diseases.^41–43 While, in stroke, they likely reflect both the correlated topography of vascular lesions⁴⁴ and the low dimensionality of FC abnormalities,^5,8 more generally, they are thought to reflect loss of neural entropy⁴⁵ caused by any neurological disorder. Interestingly, different behavioral deficits showed a dissociated pattern of behavioral correlation with SC. Cognitive functions robustly correlated at the subacute stage with SC gradients. The topography was highly specific with left hemisphere alterations related to verbal-language-memory, bilateral alterations for visuospatial-memory-attention, and lateralized attention. Interestingly, motor deficits did not correlate with changes in whole-brain structural organization but rather with the microstructural latent factors of disconnected tracts that included, commonly, the corticospinal tracts. Importantly, our sample predominantly comprised patients with low-severity, mainly subcortical strokes, although consistent with previous literature.^46,47 This lesion profile likely emphasizes the role of focal WM disconnection in motor impairment.

The correlation between WM changes and behavior weakened at 3 months, a time when behavioral recovery is nearly complete.³⁷ SC gradient 3-month changes were weakly or not related to cognitive tasks, which indicates a decoupling between SC and cognitive performance. Given that normalization of FC correlates with the behavioral recovery,³⁹ our results show that behavioral recovery may depend on synaptic reweighting of residual structural connections in the context of longitudinal structural degeneration, at least in mild-to-moderate stroke. This view is also consistent with recent work in nonhuman primates highlighting the role of transmodal association cortex in poststroke recovery.¹³ In contrast to the cognitive domains, motor deficits (subacute and chronic stages) correlated with latent-microstructural parameters of disconnected regions. Because the lesion distribution was mainly located in the central WM underlying brain motor regions, and the main patterns of disconnection were correspondingly centered around the central sulcus, including the corticospinal tracts, the association between motor deficits and the microstructure of disconnected regions indicates the importance of local damage/disconnection in accounting for motor and sensory deficits.^8,26,39,42

These results should be interpreted considering several limitations. Gradients depend on methodological choices (eg, tractography algorithm, seeding, and filtering) and parcellation scheme, which shape the connectivity matrices. Although these decisions were guided by established literature and validated in HC, replication across alternative pipelines would further support the generalizability of the results. Interpretation of gradients is also more complex than established methods, with no consensus on their meaning. Consequently, gradients should be viewed primarily as descriptive tools for summarizing connectivity organization rather than as direct biological markers. Despite the absence of differences between patients who completed the MRI follow-up and those who did not, sample attrition remains a limitation. Future investigations with larger longitudinal cohorts will be necessary to confirm these findings. In addition, lesion heterogeneity represents a limitation; although sensitivity analyses addressed cortical versus subcortical lesions, the sample size did not allow for more detailed stratified analyses. Furthermore, the dissociation between chronic SC degeneration and behavior may, at least partly, reflect ceiling effects in behavioral measures or limited sensitivity of the structural indices used in this study. Nonetheless, several strengths stand out. Multishell diffusion data enhanced tractography reliability. The longitudinal design offered insights into stroke evolution, and integrating microstructural, connectivity, and behavioral data revealed novel latent patterns.

ARTICLE INFORMATION

Acknowledgments

The authors wish to thank the participants who participated in the study.

Author Contributions

Drs Pini and Aarabi were involved in conceptualization. Dr Pini was involved in conceptualization and formal statistical analysis. Drs Pini and Aarabi were involved in conceptualization methodology. Drs Pini and Aarabi were involved in conceptualization and investigation. Dr Pini was involved in conceptualization and visualization. Dr Corbetta was involved in conceptualization and funding acquisition. Drs Corbetta and Shulman were involved in conceptualization and project administration. Dr Corbetta was involved in conceptualization supervision. Drs Pini and Aarabi were involved in conceptualization and writing-original draft. A.M. Colletta, N.V. Metcalf, and Drs Griffis, Carter, and Shulman were involved in conceptualization and writing-review and editing.

Disclosures

None.

Supplemental Material

Supplemental Materials and Methods

STROBE Checklist

Tables S1 – S3

Figures S1 – S14

References 48–79

Funding Statement

This work was supported by the Horizon 2020 European School of Network Neuroscience-European School of Network Neuroscience, H2020-SC5-2019-2 (grant agreement 860563); HORIZON-ERC-SyG (grant 101071900); HORIZON-INFRA-2022 SERV (grant 101147319) EBRAINS 2.0: A Research Infrastructure to Advance Neuroscience and Brain Health; and Fondazione Cariparo 2023 to 2024 (agreement 68076; AWARENET project).