Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Mar 1.
Published in final edited form as: Lancet Digit Health. 2026 Jan 22;8(1):100942. doi: 10.1016/j.landig.2025.100942

Associations between contralesional neuroplasticity and motor impairment through deep learning-derived MRI regional brain age in chronic stroke (ENIGMA): a multicohort, retrospective, observational study

Gilsoon Park 1, Mahir H Khan 1, Justin W Andrushko 1, Nerisa Banaj 1, Michael R Borich 1, Lara A Boyd 1, Amy Brodtmann 1, Truman R Brown 1, Cathrin M Buetefisch 1, Adriana B Conforto 1, Steven C Cramer 1, Michael Dimyan 1, Martin Domin 1, Miranda R Donnelly 1, Natalia Egorova-Brumley 1, Elsa R Ermer 1, Wuwei Feng 1, Fatemeh Geranmayeh 1, Colleen A Hanlon 1, Brenton Hordacre 1, Neda Jahanshad 1, Steven A Kautz 1, Mohamed Salah Khlif 1, Jingchun Liu 1, Martin Lotze 1, Bradley J MacIntosh 1, Feroze B Mohamed 1, Jan E Nordvik 1, Fabrizio Piras 1, Kate P Revill 1, Andrew D Robertson 1, Christian Schranz 1, Nicolas Schweighofer 1, Na Jin Seo 1, Surjo R Soekadar 1, Shraddha Srivastava 1, Bethany P Tavenner 1, Gregory T Thielman 1, Sophia I Thomopoulos 1, Daniela Vecchio 1, Emilio Werden 1, Lars T Westlye 1, Carolee J Winstein 1, George F Wittenberg 1, Jennifer K Ferris 1, Chunshui Yu 1, Paul M Thompson 1, Sook-Lei Liew 1, Hosung Kim 1
PMCID: PMC12949600  NIHMSID: NIHMS2147046  PMID: 41577565

Summary

Background

Stroke leads to complex chronic structural and functional brain changes that specifically affect motor outcomes. The brain predicted age difference (PAD) has emerged as a sensitive biomarker of both sensorimotor and cognitive function after stroke. Our previous study showed a higher global brain PAD associated with poorer motor function after stroke. However, the association between local stroke lesion load, regional brain age, and motor impairment is unclear. This study aimed to investigate the associations between focal lesion damage, regional brain PAD in both hemispheres, and motor outcomes in chronic stroke, and to identify key predictors of motor impairment.

Methods

In this multicohort, retrospective, observational study, we included individuals with chronic unilateral stroke (>180 days post stroke) from the ENIGMA Stroke Recovery Working Group dataset and used individuals from the UK Biobank cohort to train the regional brain age prediction model. Structural T1-weighted MRI scans were used to estimate regional brain PAD in 18 predefined functional subregions via a graph convolutional network algorithm. Lesion load for each region was calculated on the basis of lesion overlap. Linear mixed-effects models assessed associations between lesion size, local lesion load, and regional brain PAD. Machine learning classifiers predicted motor outcomes using lesion loads and regional brain PADs. Structural equation modelling examined directional relationships among corticospinal tract lesion load, ipsilesional brain PAD, motor outcomes, and contralesional brain PAD.

Findings

We included 501 individuals from the ENIGMA Stroke Recovery Working Group dataset (34 cohorts in eight countries) and 17 791 individuals from the UK Biobank dataset. Larger total lesion size was positively associated with higher ipsilesional regional brain PADs (older brain age) across most regions (β=0.5420 to 0.9458 across significantly correlated regions, false discovery rate [FDR]-corrected p<0.05), and with lower brain PAD in the contralesional ventral attention and language network region (β=−0.3747, 95% CI −0.6961 to −0.0534, FDR-corrected p<0.05). Higher local lesion loads showed similar patterns. Specifically, lesion load in the salience network significantly influenced regional brain PADs across both hemispheres. Machine learning models identified corticospinal tract lesion load (adjusted mean difference −0.0905, 95% CI −0.1221 to −0.0589, p<0.0001), salience network lesion load (−0.0632, −0.0906 to −0.0358, p<0.0001), and regional brain PAD in the contralesional frontoparietal network (0.9939, 0.4929 to 1.4950, p=0.0001) as the top three predictors of motor outcomes. Structural equation modelling revealed that higher corticospinal tract lesion load was associated with poorer motor outcomes (β=−0.355, 95% CI −0.446 to −0.267, p<0.0001), which were further linked to younger contralesional brain age (0.204, 0.111 to 0.295, p<0.0001), suggesting that severe motor impairment is linked to compensatory decreases in contralesional brain age.

Interpretation

Our findings reveal that larger stroke lesions are associated with accelerated ageing in the ipsilesional hemisphere and paradoxically decelerated brain ageing in the contralesional hemisphere, suggesting compensatory neural mechanisms. Assessing regional brain age might serve as a biomarker for neuroplasticity and inform targeted interventions to enhance motor recovery after stroke.

Funding

US National Institutes of Health.

Introduction

Stroke induces complex and distinct structural and functional changes in different brain regions and networks.1 Regional brain volumes rapidly decline within the first year after a stroke, beyond the atrophy rates observed in normal ageing.2 Disruptions to white matter integrity, especially in the corticospinal tract,3 strongly predict motor impairment at chronic timepoints. Moreover, people with stroke often have increased complexity in functional connectivity compared with healthy controls, suggesting compensatory mechanisms and neural reorganisation to overcome damaged connections.4 Notably, in chronic stroke, typically defined as 6 months or longer after stroke onset, brain changes might result from lesion-induced damage or use-dependent reorganisation supporting adaptive behaviours.5 Quantifying regional variations in these changes could identify precise neural biomarkers or therapeutic targets to enhance rehabilitation and understand neuroplasticity after stroke by linking regional brain predicted age difference (PAD)—the difference between predicted brain age and chronological age—with motor impairment severity.

Brain age estimation, based on neuroimaging features, offers a promising approach to quantify regional structural changes after a stroke.6 The brain PAD has emerged as a sensitive biomarker, with a higher global brain PAD associated with poorer post-stroke motor outcomes (appendix p 28).7 Whole-brain PAD might obscure localised variations in neural integrity that are crucial for recovery. Advances in neuroimaging and computational methods now enable the estimation of brain age for specific regions using cortical features, a technique previously applied to study cardiometabolic syndrome.8 Yet, its use in predicting stroke outcomes is unexplored. Investigating regional brain age offers a quantification of stroke damage patterns, both proximal and distal to the lesion site, and provides a unique window into how unaffected brain areas contribute to stroke recovery, including their possible compensatory mechanisms. This is particularly important given emerging evidence that networks beyond the traditional sensorimotor system—such as the frontoparietal network and default mode network in the contralesional hemisphere—play a pivotal role in recovery.9–14

In this study, we examined the relationship between focal stroke damage, motor outcomes, and regional brain age in individuals with chronic stroke. We used a graph convolutional network (GCN) algorithm to estimate regional brain PAD using cortical features of predefined functional subregions. We hypothesised that greater local lesion damage would be strongly associated with older regional brain ages in the ipsilesional hemisphere and with worse functional outcomes in chronic stroke. Building on the association between global brain age and motor outcomes, we further posited that elevated regional brain PAD near the lesion would serve as a key predictor of poor motor recovery. Conversely, we anticipated that a younger contralesional PAD might reflect compensatory neuroplasticity—facilitated by corticocortical pathways and internetwork communication—particularly in people with severe motor impairment. To elucidate these relationships, we performed mediation analyses to assess whether and how regional brain age mediates the relationship between lesion damage and motor outcomes.

Methods

Participants

In this multicohort, retrospective, observational study, we analysed data from two sets of participants for algorithm development and analysis. The regional brain age prediction algorithm, as described previously,8 was trained on a subset of data from the UK Biobank database, a large-scale biomedical repository containing genetic, imaging, and health information from more than 500 000 individuals in the UK.15 Briefly, among participants recruited from 2014 to 2023, we included participants with MRI brain imaging who were non-Hispanic White and excluded any individuals with self-reported or hospital-recorded history of any neurological disorders, resulting in 17 791 individuals (52.7% women and a mean age 63.2 years [SD 7.4]) for algorithm development.

We then used the ENIGMA Stroke Recovery Working Group dataset of individuals with stroke, a multisite repository containing retrospective studies of neuroimaging data, demographic information, and functional outcome measures.16 Data were frozen for this analysis on June 23, 2023. We included any individuals with a motor outcome measure who were defined to be in the chronic stage of stroke recovery (180 days or more since stroke), had a chronological age of 45 years or older, had the presence of a lesion in only one hemisphere (ie, unilateral lesion), and who had a successful extraction of cortical thickness (figure 1; appendix pp 2–3).

Figure 1:

Figure 1:

Open in a new tab

Data selection flowchart for people with stroke

The collection of ENIGMA stroke data followed the Declaration of Helsinki and was approved by local ethics boards at each respective institute. All participants in the ENIGMA cohort provided written informed consent for the study. This study was approved by the Institutional Review Board (approval number 00002881) at the University of Southern California Health Science Campus. The UK Biobank data used for our study were anonymous and de-identified, exempting the study from the requirement of obtaining informed consent. To assess the generalisability of our regional brain age prediction model, we additionally used the OpenBHB dataset (391 healthy controls) for external validation and sensitivity analysis. Details regarding the OpenBHB dataset characteristics are provided in the appendix (pp 23–24).

Behavioural data

Due to the heterogeneity of motor outcomes recorded by each site in the ENIGMA stroke dataset, we harmonised different motor outcome measures (appendix p 20). As done previously, we defined a primary motor outcome score as the raw score divided by the maximum attainable score for each test, yielding a 0–1 proportion that preserves the higher equals better interpretation and matches previous ENIGMA work.16,17 A score of 1.0 indicated no impairment and 0.0 indicated severe impairment.

MRI data processing

3D T1-weighted MRI images were acquired from the UK Biobank dataset. The imaging variables include an inversion time of 880 ms, a repetition time of 2000 ms, an echo time of 2.01 ms, a 1 mm isotropic voxel size, a matrix size of 208 × 256 × 256, and a sensitivity encoding factor of 2.

A full description about the acquisition of high-resolution T1-weighted images in the stroke dataset has been reported elsewhere.16 Briefly, stroke imaging was done according to standardised protocols, such as the Human Connectome Project or Alzheimer’s Disease Neuroimaging Initiative.16 Since protocols varied across sites, images were visually inspected by ENIGMA Stroke Recovery core team (S-LL and MHK) for quality control before further processing, as well as after each processing step.

All T1-weighted images in both the UK Biobank dataset and the ENIGMA Stroke dataset were processed using a modified CIVET pipeline (version 2.1.0) to extract cortical morphometry features.18 The reconstructed surfaces consisted of 40 962 vertices in each hemisphere. Cortical thickness measurements were obtained by computing the Euclidean distance between the vertices of the inner cortical surface and the corresponding vertices of the outer cortical surface. The grey matter to white matter intensity ratio was calculated using the intensities of grey matter and white matter 1 mm from the inner cortical surface. Further methodological details, including the validation of stroke lesion effect on image processing and harmonisation of MRI scans across all sites, are provided in the appendix (pp 2–12).

The cortical surface was partitioned into nine functional subregions as defined by Yeo and colleagues (appendix p 19).19 The subregions comprised the sensorimotor, frontoparietal, dorsal attention, ventral attention with language, default mode, salience, auditory, visual, and limbic networks. Since we intended to study the brain age of the ipsilesional and contralesional hemispheres, the nine functional subregions were further divided into left and right hemispheres, for a total of 18 regions of interest (ROIs). We provided details of these methods in the appendix (pp 13–14).

Lesion analysis

Stroke lesions were manually segmented by trained research team members from ENIGMA Stroke Recovery core team (S-LL and MHK), based on a previously published protocol.20 Each 3D T1-weighted image and lesion mask in the stroke dataset was co-registered to Montreal Neurological Institute space using linear and non-linear transformations generated by the CIVET pipeline, as described previously.20 We did not have measures to ascertain whether the stroke was ischaemic or haemorrhagic, given that our study focuses on the chronic phase of stroke.

To quantify focal lesion damage for the 18 ROIs, we calculated lesion load, which is measured by dividing the volume of the overlap between the lesion and each ROI by the respective ROI volume. We also calculated a corticospinal tract lesion load using a publicly available corticospinal tract template that includes primary and higher order sensorimotor regions.19 The corticospinal tract is recognised as an important predictor of motor outcome due to its importance as a motor pathway in the nervous system.7,19 By including corticospinal tract lesion load in our model, we were able to assess the influence of local lesion loads of each ROI on motor outcomes, independent of corticospinal tract lesion load on motor outcome.

Regional brain age prediction

We developed an in-house regional brain age prediction model using GCNs,21 trained on the UK Biobank dataset (figure 2). The model used cortical thickness and grey matter to white matter intensity ratio at each vertex as features (figure 2A). Vertices and their connections were defined as nodes and edges in the graph structure, respectively. In this study, we focused exclusively on cortical features to investigate the lesion influence on cortical ageing, which allowed for maintaining a large and homogeneous sample, although many participants had subcortical or thalamic strokes. Signals at each node were projected into the spectral domain via the graph Fourier transform, filtered, and then projected back into the spatial domain.22 The GCN architecture included a graph convolutional layer, rectified linear unit activation, graph max pooling, and a fully connected layer for prediction (figure 2B). Model performance was evaluated using five-times cross-validation, a standard technique chosen to provide a reliable estimate of model performance with reasonable computational cost, and an ensemble of the five models was used to improve the prediction performance on the target dataset. Detailed model architecture and training procedures are provided in the appendix (pp 15–16).

Figure 2: The training flowchart for predicting regional brain age.

Figure 2:

Open in a new tab

(A) Flowchart for data generation to pass into GCNs for predicting regional brain age. We built a cortical surface model from a 3D T1-weighted image using the CIVET pipeline and extracted cortical thickness and grey matter to white matter intensity ratio. The cortical surface and cortical features were divided into 18 ROIs. (B) One GCN model per ROI was trained to predict regional brain age. The cortical surface was used to define nodes and edges for a graph structure, and cortical features were used as signals for each node. All models had the same GCN structure, which consists of a graph convolution, rectified linear unit for activation function, max-pooling layer for graph pooling, and a fully connected layer. To obtain regional brain PAD, we took the difference between the predicted regional brain age and chronological age. GCN=graph convolutional network. PAD=predicted age difference. ROI=region of interest.

Brain PAD analysis

Brain PAD was calculated as the difference between the predicted brain age and the chronological age of the individual. A higher, positive brain PAD is indicative of an older appearing brain, whereas a lower, negative brain PAD is indicative of a younger appearing brain. Regional brain PAD is defined as the brain PAD obtained from a particular functional subregion or ROI. To mitigate the regression dilution bias observed with brain PAD, we used the linear trend removal method proposed by Smith and colleagues.23 This adjustment ensures that brain PAD reflects the relative brain health status of individuals, independent of their age.

Statistical analysis

We first evaluated the prediction performance of the regional brain age models trained on the UK Biobank dataset. Performance was assessed using the mean absolute error and Pearson correlation coefficients (R) between the predicted brain age and chronological age. To further address generalisability, we conducted an external validation (sensitivity analysis) using the OpenBHB dataset, a multisite collection of MRI data.

After model validation, we investigated the relationship between regional brain PAD and total lesion volume. Linear mixed-effects models were used with each regional brain PAD as the dependent variable; total lesion volume as the independent variable; age at scan time, sex, and days since stroke as covariates; and cohort as a random effect.

Next, we investigated the relationship between lesion load of each ROI, including corticospinal tract lesion load and regional brain PAD using linear mixed-effects models, with lesion load as the dependent variable, regional brain PAD as the independent variable, and the same covariates and random effects as described earlier. To focus on relationships between each lesion load and regional brain PAD, we included mean regional brain PAD as an additional covariate to account for overall brain ageing patterns. For both analyses, if a lesion load was greater than 20% of the total volume of a ROI for which brain PAD was computed, we excluded that regional brain PAD from our analysis. This 20% cutoff was established empirically to avoid potentially false brain age predictions due to lesions encroaching on the region. False discovery rate (FDR) correction was applied to account for multiple comparisons. The t values reported corresponded to the t statistics associated with the FDR-corrected p value thresholds.

We then predicted motor outcomes using different machine learning methods: random forest, gradient boosting, AdaBoost, and XGBoost. Based on the previous literature7,24 a threshold of 0.636 was used to categorise motor outcomes into good (mild to no impairment; n=278) and poor (moderate to severe impairment; n=223). Performance metrics such as accuracy and area under the curve were used to evaluate the predictive quality of each method. Input features for these methods included local lesion loads, corticospinal tract lesion load, regional brain PAD, age at scan time, sex, days since stroke, and intracranial volume. The goal of this prediction was to identify the effect of each feature on motor outcomes. To do this, we calculated the Gini index, which estimates the impurity, or the probability of incorrectly labelling a randomly chosen element in the dataset, at each node of the decision tree. A Gini index of 0 suggests complete purity (ie, only one label in the set), whereas a Gini index of 0.5 suggests complete impurity (ie, equally likely to obtain both labels in the set). To establish feature importance, we identified how much each feature increased the purity by decreasing the Gini index. In other words, the features with the highest importance contributed to the greatest magnitude decrease in the Gini index, which we refer to as the feature importance score. To derive a statistically robust importance score, we performed bootstrapping with 5000 iterations where the dataset was randomly divided into training and test sets with 8:2 splits. We examined the mean and SD of the feature importance scores obtained in the 5000-iteration bootstrap. To quantify the effect of key features on motor outcomes, we calculated adjusted mean differences between good and poor motor outcome groups using a linear mixed-effects model. This model adjusted for age, sex, days since stroke, intracranial volume, and mean regional brain PAD as fixed-effect covariates, with site included as a random effect.

In our final analysis, we examined directionality within the relationships between corticospinal tract lesion loads, regional brain PADs, and motor scores using structural equation modelling. Based on our hypotheses, we evaluated the mediating effects of ipsilesional regional brain PAD and motor impairment severity on mean contralesional regional brain PAD. Models were fitted to all data by an optimiser with a Wishart log likelihood, and model fit was assessed using a χ2 distribution, comparative fit index, root mean square error of approximation, and adjusted goodness-of-fit index, with 95% CIs obtained through 5000 bootstrap iterations. Implementation details for all models are available in the appendix (p 17). All statistical analyses were implemented using MATLAB version R2022b for linear mixed-effects modelling (fitlme). Machine learning analyses were conducted in Python (version 3.10) using scikit-learn (version 1.0.2) for the Random Forest, Gradient Boosting, and AdaBoost classifiers, and XGBoost (version 1.6.2) for the XGBoost classifier. Structural equation modelling was performed in Python using the semopy package (version 2.3.9).

Role of the funding source

The funder of this study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Results

As of June 23, 2023, since the first recruitment beginning in Jan 15, 2010, the ENIGMA Stroke Recovery Working Group dataset contained data from 883 individuals who matched our initial eligibility criteria. Cross-sectional data from 501 individuals with stroke from 34 cohorts across eight countries (Australia, Brazil, Canada, China, Germany, Norway, the UK, and the USA) met the final eligibility criteria and were included in this analysis (figure 1; appendix p 21). There were 318 men (63%), 162 women (32%), and 21 people with unknown sex (4%), with a median age of 63 years (IQR 56–70). 260 (52%) of 501 individuals were from the USA. Stroke severity (ie, primary sensorimotor score) ranged from 0 to 1 (median 0.6516, IQR 0.4175–0.9091). Days after stroke ranged from 180 to 8915 days (704, 262–1521). Lesion volume ranged from 0.013 mL to 339.64 mL (7.35, 1.62–50.66) and a lesion overlap map was produced (appendix p 18).

Using the UK Biobank dataset, we assessed the performance of the brain age prediction models for each ROI. The mean absolute errors ranged from 2.94 to 3.13 years, and correlation coefficients (R values) ranged from 0.88 to 0.90, indicating robust predictive performance across all models (appendix p 22). The sensitivity analyses using the OpenBHB dataset showed consistent prediction accuracy across diverse cohorts (mean absolute errors 1.53–4.98 years). Details on data characteristics and prediction accuracy are found in the appendix (pp 23–24).

We first examined the relationship between total stroke lesion size and regional brain PAD of the 18 functional ROIs (table). Larger total lesion size was positively correlated with regional brain PADs for all ipsilesional ROIs except for the auditory and visual networks (β=0.5420 to 0.9458 across positively correlated ROIs, FDR-corrected p<0.05). Larger total lesion size was negatively correlated with brain PAD in the contralesional ventral attention and language network ROI (β=−0.3747, 95% CI −0.6961 to −0.0534, FDR-corrected p<0.05). No significant associations were observed between total lesion size and other contralesional ROIs. Effects of covariates on brain PADs are reported in the appendix (p 25).

Table:

Association between total stroke lesion size and regional brain-PAD across functional regions

Ipsilesional Contralesional
β (95% CI) p value β (95% CI) p value
Sensorimotor 0.7431 (0.3535 to 1.1327) 0.0002* 0.0655 (−0.2659 to 0.3969) 0.6978
Frontoparietal network 0.7690 (0.4076 to 1.3040) <0.0001* −0.0611 (−0.3760 to 0.2539) 0.7032
Dorsal 0.5498 (0.2199 to 0.8797) 0.0011* 0.003 (−0.3193 to 0.3253) 0.9853
Ventral attention and language 0.7733 (0.2412 to 1.3053) 0.0045* −0.3747 (−0.6961 to −0.0534) 0.0224*
Default 0.5420 (0.1986 to 0.8855) 0.0021* −0.1306 (−0.4364 to 0.1753) 0.4019
Salience 0.7159 (0.1770 to 1.2548) 0.0094* −0.1048 (−0.4625 to 0.2529) 0.5651
Auditory 0.2807 (−0.2001 to 0.7615) 0.2518 −0.0598 (−0.3870 to 0.2674) 0.7195
Visual 0.3157 (0.0334 to 0.5980) 0.0285 0.0730 (−0.2210 to 0.3670) 0.6258
Limbic 0.9458 (0.5494 to 1.2409) <0.0001* −0.0804 (−0.4102 to 0.2495) 0.6323
Open in a new tab

The β coefficient and significance of the effect of total stroke lesion size on each region of interest are shown. PAD=predicted age difference.

*

Statistical significance (p<0.05) after applying a false discovery rate correction.

The second analysis examined the relationship between regional brain PAD and local lesion loads of the 18 functional networks, as well as the corticospinal tract lesion load (figure 3). Higher local lesion loads were correlated with higher ipsilesional regional brain PAD and lower contralesional regional brain PAD. Regional brain PADs of the ipsilesional frontoparietal network, default mode, salience, and visual networks, as well as the contralesional frontoparietal network, dorsal attention, ventral attention and language, and salience networks were associated with at least three lesion load metrics each. Notably, higher lesion loads in the salience network showed widespread significant associations in regional brain PADs across both ipsilesional (sensorimotor, frontoparietal network, ventral attention and language, default mode, and visual networks) and contralesional (frontoparietal network, dorsal attention, ventral attention and language, default mode, salience, and auditory networks) hemispheres. Further details on the β values and their 95% CIs are provided in the appendix (p 26).

Figure 3: Results of the association analysis between lesion loads and regional brain PAD.

Figure 3:

Open in a new tab

Each square represents the association between regional brain PAD (x-axis) and lesion load (y-axis). A darker red colour indicates higher brain PAD (older-appearing brain), and a darker blue colour indicates lower brain PAD (younger-appearing brain). Because the brain age in lesional regions (ie, a lesion load in the region of interest for which brain PAD was computed to be >20%) was not computed due to a potentially false brain age prediction, the correlation analyses in the diagonal cells were not performed and they are coloured in grey. FDR=false discovery rate. PAD=predicted age difference. *FDR-corrected p<0.05 (FDR threshold p=0.01320; t value=2.23). **FDR-corrected p<0.01 (FDR threshold p=0.00110; t value=3.08). ***FDR-corrected p<0.005 (FDR threshold p=0.00038; t value=3.39).

We evaluated features important for predicting motor outcome. The random forest method showed the most robust performance in terms of both accuracy (median 0.6634, IQR 0.6337–0.6931) and area under the curve (0.6572, 0.6281–0.6864) and was used to evaluate the importance of all input features (appendix p 27). Based on mean importance scores, the top three predictive features were corticospinal tract lesion load (mean importance score 0.0646, SD 0.0073), salience network lesion load (0.0593, 0.0074), and contralesional frontoparietal network brain PAD (0.0491, 0.0065; figure 4). Subsequent linear mixed-effects modelling confirmed that the adjusted mean differences between good and poor motor outcome groups for these features were all significant: corticospinal tract lesion load (adjusted mean difference −0.0905, 95% CI −0.1221 to −0.0589, p<0.0001), salience network lesion load (−0.0632, −0.0906 to −0.0358, p<0.0001), and contralesional frontoparietal network brain PAD (0.9939, 0.4929 to 1.4950, p=0.0001). Notably, only contralesional—and not ipsilesional—regional brain PADs were significant predictors of motor outcomes, with younger contralesional brain PAD associated with worse motor impairment.

Figure 4: Ranking of feature importance scores in predicting motor outcome on the basis of 5000 bootstrap iterations of the Random Forest method.

Figure 4:

Open in a new tab

The bars represent the mean feature importance scores, and the error bars represent the SDs for each predictor of motor outcome. Of the top 20 significant predictors, eight were contralesional regional brain PAD and none were ipsilesional regional brain PAD. CBPAD=contralesional brain PAD. LL=lesion load. PAD=predicted age difference.

We used structural equation modelling to examine the directional relationships among corticospinal tract lesion load, motor outcomes, mean contralesional brain PAD, and ipsilesional brain PAD (figure 5). The structural equation model had acceptable model fit indices (χ2/degree of freedom=0.99, χ2 p=0.3717, comparative fit index=1.00, root mean square error of approximation=0.00, and adjusted goodness-of-fit index=0.95). Higher corticospinal tract lesion load was directly associated with worse motor outcomes (β=−0.355, 95% CI −0.446 to −0.267, p<0.0001) and higher ipsilesional brain PAD (0.262, 0.131 to 0.421, p<0.0001). Elevated ipsilesional brain PAD was directly associated with worse motor outcomes (−0.102, −0.190 to −0.013, p=0.0266) and higher contralesional brain PAD (older brain ageing; 0.213, 0.128 to 0.315, p<0.0001). Worse motor outcomes were directly associated with lower contralesional brain PAD (younger brain ageing; 0.204, 0.111 to 0.295, p<0.0001). These results suggest that motor outcomes mediate the effect of corticospinal tract lesion load on the contralesional brain PAD. Higher corticospinal tract lesion load contributed to lower contralesional brain PAD via its association with motor outcomes (indirect effect β=−0.072, 95% CI −0.113 to −0.037), whereas higher corticospinal tract lesion load was associated with lower motor outcome scores through its effect on ipsilesional brain PAD (−0.027, −0.058 to −0.002).

Figure 5: Structural equation model to establish the relationship between corticospinal tract lesion loads, regional brain PADs, and motor scores.

Figure 5:

Open in a new tab

This model shows the significant directional relationship between corticospinal tract lesion load, ipsilesional brain PAD, mean contralesional regional brain PAD, and motor outcome. All associations in the model were statistically significant (p<0.05). PAD=predicted age difference.

Discussion

In this study, we showed a complex relationship between total and focal lesion damage, regional brain age measures, and motor outcomes in people with chronic unilateral stroke. We found that larger total lesion sizes and higher local lesion loads were associated with higher ipsilesional brain PADs and lower contralesional brain PADs. Notably, lesion load in the salience network substantially influenced regional brain PADs across both hemispheres. Additionally, corticospinal tract lesion load, lesion load in the salience network, and regional brain PAD in the contralesional frontoparietal network emerged as the most significant predictors of motor outcomes. We also found that higher corticospinal tract lesion load was associated with older ipsilesional brain PADs and worse motor impairment, consistent with a previous study.7 Notably, more severe motor impairment was associated with younger contralesional brain PADs, suggesting that motor impairment might drive compensatory changes in the contralesional hemisphere.

Our regional brain age model confirmed our hypothesis that, in unilateral chronic stroke, larger stroke volumes and higher lesion loads were associated with higher brain PADs in ipsilesional brain regions. Conversely, higher lesion loads were associated with lower brain PADs in contralesional regions. Since mean regional brain age was included in the model, these findings indicate region-specific patterns independent of overall brain ageing patterns. This finding shows that regional brain age is sensitive to focal and network-specific damage after stroke in ways that global brain age might not fully capture.

Lesion overlap with crucial brain regions is one of the most important features in predicting motor outcomes after stroke. The most important feature in our prediction models was the corticospinal tract lesion load, as expected based on previous literature.7,25 Notably, the second most important feature was lesion load of the salience network. Lesion load in the salience network was strongly associated with older ipsilesional brain PADs and younger contralesional brain PADs in almost all regions. Damage to the salience network disrupts connections with other functional networks such as the default mode network and frontoparietal network.26 Given that imbalances in the salience network have been linked to cognitive and neuropsychiatric disorders,27 future studies could assess relationships between damage to the salience network, cognitive and emotional dysfunction, and motor impairment after stroke.

In addition to the lesion load features, the most important PAD feature to predict motor outcome was younger contralesional brain PAD in the frontoparietal network. Notably, younger brain PAD in the default mode network was also identified as the eighth most significant predictor. We discuss the role of these functional network regions in post-stroke recovery in the appendix (p 29).

Notably, among the 20 significant motor outcome predictors, eight were younger contralesional brain PADs, and none were ipsilesional brain PADs, suggesting that decelerated ageing in the contralesional hemisphere might occur in response to severe motor impairment. One explanation is increased reliance on the non-paretic side, leading to enhanced connectivity in the contralesional hemisphere controlling it.28 Such changes might manifest as a younger brain PAD in our study.29 Alternatively, severe impairment might necessitate more widespread brain activation in both hemispheres during paretic limb use.30 This recruitment might occur via the corpus callosum,31–33 and might compensate for damaged ipsilesional tissue. Future studies should use longitudinal designs to investigate the time course of regional brain age changes and their causal relationship with motor outcomes.

Without longitudinal data, we used structural equation modelling with cross-sectional data to investigate possible directional relationships between regional brain changes and motor outcomes. We found that motor impairment mediated the relationship between corticospinal tract damage, ipsilesional brain age, and regional brain age in the contralesional hemisphere. Corticospinal tract damage and increased brain age are thought to be measures of primary (focal) damage and secondary atrophy due to stroke, respectively; as corticospinal tract damage and ipsilesional brain age increased, we observed worse motor outcomes. In turn, we found that worse motor outcomes were associated with younger contralesional regional brain ages, which is consistent with our hypotheses and might suggest that the brain uses contralesional regions to compensate for severe motor impairment. These findings show the complex, bidirectional nature of the relationship between brain ageing and motor impairment after stroke, because ipsilesional brain age affects motor outcomes, which then drive changes in contralesional brain age. Other possible interpretations of our findings in the relationship between regional brain ages and motor outcome are discussed in the appendix (p 30).

A limitation of this work is its cross-sectional design and focus on the chronic phase (>180 days) after stroke. Future research should examine brain age from acute stages onward to understand when the relationship between contralesional brain age and motor outcomes is strongest, potentially identifying optimal windows for intervention. Additionally, investigating the relationship between contralesional brain age and non-paretic limb use, possibly using wearable sensors, could enhance our understanding of how behavioural adaptations influence regional brain ageing. The absence of standardisation, despite numerous methods developed for inferring brain age, complicates cross-study comparisons and inferences. Brain age estimates are inherently indirect measures of brain health that might reflect a range of structural and functional influences. Our approach—excluding ROIs where 20% or more of the cortical area is affected by a lesion—might reduce sensitivity to detecting perilesional brain ageing, especially in regions affected by extensive cortical lesions. We only assessed cortical brain age, despite many participants having subcortical or thalamic strokes. Although thalamo-cortical connectivity suggests that cortical features reflect subcortical pathology,17 future work should incorporate subcortical brain age measures. The auditory network brain PAD did not correlate with the total stroke lesion size (table). The limited relationship between auditory network brain age and stroke lesion load is discussed in the appendix (p 31). Other than age, sex, number of days after stroke, and intracranial volume, we did not consider possible covariates associated with brain age, including education or race and ethnicity, due to the fact that these data were only available in a small number of participants in this study. Despite a large multisite study, it is important to note that more than half of the participants (52%) were recruited from the USA. This geographical skewness toward high-income nations might limit the generalisability of our findings to global stroke populations with different genetic backgrounds, health-care access, or rehabilitation standards.

Our findings emphasise the importance of assessing regional brain age as a sensitive biomarker for neuroplasticity and motor recovery after stroke. The observed accelerated ageing in ipsilesional regions and paradoxically younger brain age in contralesional regions suggest different roles for each hemisphere in response to stroke-induced damage. Targeting specific neural networks, such as the contralesional frontoparietal network and salience network, might hold promise for developing personalised rehabilitation strategies aimed at enhancing motor recovery. However, given the absence of longitudinal data, causality cannot be definitively established, and the observed brain ages could naturally be associated with the motor outcome.

Supplementary Material

1

Research in context.

Evidence before this study

We searched PubMed and Embase from database inception to June 23, 2023, for articles published in English, using the terms “stroke”, “brain age”, “brain-predicted age difference”, “lesion load”, “motor impairment”, and “neuroplasticity”. Previous studies have shown that stroke leads to complex structural and functional brain changes, including rapid brain atrophy and disruptions in white matter integrity, particularly in the corticospinal tract, which are associated with motor impairment. The brain predicted age difference (PAD)—defined as the difference between predicted brain age and chronological age—has emerged as a biomarker, with higher global brain PAD linked to poorer motor outcomes after stroke. However, the relationship between local lesion load, regional brain age estimates, and motor impairment is unclear. There is a gap in understanding how focal lesion damage affects regional brain ageing and how this relates to functional recovery in people with chronic stroke.

Added value of this study

Our study is the first to investigate the relationships between focal lesion damage, regional brain age, and motor outcomes in a large cohort of 501 individuals with chronic unilateral stroke. Using advanced machine learning techniques, specifically graph convolutional networks, we estimated regional brain age in predefined functional subregions. We found that larger lesion sizes and higher local lesion loads are associated with accelerated ageing in ipsilesional regions and decelerated ageing in contralesional regions. Notably, we identified that regional brain age in the contralesional frontoparietal network, along with lesion load metrics, are significant predictors of motor outcomes. These findings provide new insights into the neuroplastic mechanisms after stroke and highlight the sensitivity of regional brain age to focal and network-specific damage.

Implications of all the available evidence

Our findings suggest that assessing regional brain age can serve as a biomarker for neuroplastic changes and motor recovery potential after stroke. Understanding the distinct roles of the ipsilesional and contralesional hemispheres in response to stroke-induced damage might inform the development of targeted rehabilitation strategies. Future research should focus on longitudinal studies to explore the progression of regional brain ageing from acute to chronic stages of stroke, identify optimal windows for intervention, and investigate whether targeting specific neural networks can enhance motor recovery. Integrating regional brain age assessments into clinical practice could lead to personalised therapeutic approaches aimed at improving functional outcomes for people who survive stroke.

Acknowledgments

This work was funded by the National Institutes of Health grant R01 NS115845 (to S-LL and MHK).

Declaration of interests

SCC is a consultant for Alevian, Astellas, Bayer, BlueRock Therapeutics, BrainQ, Constant Therapeutics, Medtronic, MicroTransponder, Myomo, Myrobalan, NeuroTrauma Sciences, Simcere, and TRCare. CAH served as a consultant to MagStim and Roswell Park Cancer Insitute, and is an employee of BrainsWay. AB has served on Scientific Advisory Boards for Eisai, Eli Lilly, Novo Nordisk, and Roche; was the honorary chair of the International Stroke Genomics Consortium 2025–27; was a medical advisor for Dementia Australia; and received support from Lilly for attendance at the Australian Dementia Research Forum, Perth, Australia (June, 2025). LTW is a shareholder and research advisor in baba.vision. BH reports a clinical partnership with Fourier Technology. FP received honoraria for an educational event for EISAI. NJ received a speaker honorarium from the University of Southern California and served on the National Advisory Mental Health Council on High Dementional Datasets. CJW receives royalties from the 2nd Edition of Stroke Recovery and Rehabilitation, Demos Medical Publishing; is a consultant for MicroTransponder and MedRhythm; and is a member of the Data Safety and Monitoring Committee of Enspire Deep Brain Stimulation Therapy. All other authors declare no competing interests. GP was supported by Micheal J Fox Foundation and The Parkinson’s Progression Markers Initiative (#MJFF-023385). LAB was supported by Canadian Institutes of Health Research operating grants (MOP-130269 and MOP-106651) and project grant (PTJ-148535). AB was supported by a National Health and Medical Research Council project grant (GNT1020526), the Australian Brain Foundation, Wicking Trust, Collie Trust, and Sidney and Fiona Myer Family Foundation; and fellowships from the National Heart Foundation (100784 and 104748). CMB was supported by a US National Institutes of Health (NIH) grant (R01NS090677). ABC was supported by Hospital Israelita Albert Einstein (2250–14) and the NIH (R01NS076348–01). NE-B was supported by the Australian Research Council Future Fellowship (FT230100235). FG was supported by the Wellcome Trust (093957) and partly supported by the National Institute for Health Research Imperial Biomedical Research Centre. CAH was supported by NIH and National Institute of General Medical Sciences (NIGMS) grant (P20GM109040). SAK was supported by an NIH grant (P30 GM154630) and Department of Veterans Affairs Rehabilitation Research and Development grants (1IK6RX003075 and 1I01RX001935). KPR was supported by an NIH grant (R01NS090677). NJS was supported by grants from the NIH and National Institute of Child Health and Human Development (R01HD094731) and NIH and NIGMS (P20GM109040). SRS was supported by European Research Council (ERC NGBMI 759370 and ERC BNCI2 101088715), Deutsche Forschun gsgemeinschaft grant (DFG SO 932/7–1), and Federal Ministry of Education and Research (Bundesministerium für Bildung und Forschung 01GQ0831) for data collection. GTT was supported by the Temple University sub-award of NIH R24 and the REACT Pilot Grant sub-award for “The effect of transcranial direct current stimulation (tDCS) on upper extremity use following a Cerebral Vascular Accident (CVA)” and the National Resource Center for High-Impact Clinical Trials in Medical Rehabilitation: Pilot Projects Component 2019–2021. DV was supported by the Italian Ministry of Health Ricerca Corrente 2024 (RC-24) at Istituto di Ricovero e Cura a Carattere Scientifico Santa Lucia Foundation, Rome, Italy. LTW was supported by the Research Council of Norway (249795 and 248238), the South-Eastern Norway Regional Health Authority (2014097, 2015044, 2015073, 2018037, 2018076, 2019107, and 2020086), and the Norwegian Extra Foundation for Health and Rehabilitation (2015/FO5146); and performed this work on the Services for Sensitive Data, University of Oslo, Norway, with resources provided by UNINETT Sigma2, the National Infrastructure for High-Performance Computing and Data Storage in Norway. GFW was supported by grants from the Department of Veterans Affairs (I01 RX001667) and the NIH (R01 HD061462). HK was supported by the NIH grant (U01 AG024904). NS was supported by a grant from the NIH and NINDS (R56NS126748). MHK acknowledges the support of an NIH grant (S10 OD032285). PMT was supported by grants from the NIH for the ENIGMA Bipolar Initiative (R01MH129742), “MEGA-OCD: A Global Data-Driven Initiative to Discover Biosignatures of OCD” (R01MH138569), ENIGMA Eating Disorders Initiative (U01MH136221), Federated Deep Learning to Accelerate Alzheimer’s Disease Research (R01AG081571), ENIGMA Parkinson’s Initiative (R01NS107513), ENIGMA World Aging Center (R01AG058854), CARE4Kids: Imaging Biomarker Core (U54NS121688), Worldwide Tractometry Initiative (Parkinson’s Disease) (RF1NS136995), and Ultrascale Machine Learning for Alzheimer’s Biobanks (U01AG068057). BH was supported by the Sylvia and Charles Viertel Charitable Foundation (VTL2016CI009). NJ was partly supported by NIH grants (R01AG087513 and R01MH134004) and acknowledges the use of the UK Biobank Resource under Application Number 11559. SCC was supported by the NIH, Patient-Centered Outcomes Research Institute, and Department of Veterans Affairs.

Footnotes

Data sharing

The de-identified participant data used for model training are available from the UK Biobank resource to bona fide researchers for health-related research in the public interest; applications for access can be made at https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. The full ENIGMA Stroke data are not publicly available in a repository as they might contain information that could compromise the privacy of research participants. Publicly sharing the full data is limited by data sharing restrictions imposed by some of the: (1) local ethical review boards of the participating sites, and consent documents; (2) national and trans-national data sharing laws; and (3) institutional processes, which might require a signed data transfer agreement for limited and predefined data use. However, we support data sharing among members of the ENIGMA Stroke Recovery Working Group who submit an analysis plan for a secondary project for group review. Following approval of the analysis plan, access to the relevant data will be provided, contingent on data availability, local principal investigator approval, and compliance with all supervening regulatory boards. More information about joining ENIGMA Stroke Recovery can be found here: https://enigma.ini.usc.edu/ongoing/enigma-stroke-recovery/. The source code for the regional brain age models developed in this study is openly available on GitHub at https://github.com/pks1207/regional_Brain_age.

References

  • 1.Schlemm E, Schulz R, Bönstrup M, et al. Structural brain networks and functional motor outcome after stroke–a prospective cohort study. Brain Commun 2020; 2: fcaa001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Brodtmann A, Khlif MS, Egorova N, Veldsman M, Bird LJ, Werden E. Dynamic regional brain atrophy rates in the first year after ischemic stroke. Stroke 2020; 51: e183–92. [DOI] [PubMed] [Google Scholar]
  • 3.Pinter D, Gattringer T, Fandler-Höfler S, et al. Early progressive changes in white matter integrity are associated with stroke recovery. Transl Stroke Res 2020; 11: 1264–72. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Li W, Li Y, Zhu W, Chen X. Changes in brain functional network connectivity after stroke. Neural Regen Res 2014; 9: 51–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Katai S, Maeda M, Katsuyama S, et al. Cortical reorganization correlates with motor recovery after low-frequency repetitive transcranial magnetic stimulation combined with occupational therapy in chronic subcortical stroke patients. Neuroimage Rep 2023; 3: 100156. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Han LKM, Dinga R, Hahn T, et al. Brain aging in major depressive disorder: results from the ENIGMA major depressive disorder working group. Mol Psychiatry 2021; 26: 5124–39. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Liew S-L, Schweighofer N, Cole JH, et al. Association of brain age, lesion volume, and functional outcome in patients with stroke. Neurology 2023; 100: e2103–13. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Kang SH, Liu M, Park G, et al. Different effects of cardiometabolic syndrome on brain age in relation to gender and ethnicity. Alzheimers Res Ther 2023; 15: 68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Siegel JS, Ramsey LE, Snyder AZ, et al. Disruptions of network connectivity predict impairment in multiple behavioral domains after stroke. Proc Natl Acad Sci USA 2016; 113: E4367–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.De Bruyn N, Saenen L, Thijs L, et al. Brain connectivity alterations after additional sensorimotor or motor therapy for the upper limb in the early-phase post stroke: a randomized controlled trial. Brain Commun 2021; 3: fcab074. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Tscherpel C, Dern S, Hensel L, Ziemann U, Fink GR, Grefkes C. Brain responsivity provides an individual readout for motor recovery after stroke. Brain 2020; 143: 1873–88. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Zhao Z, Wu J, Fan M, et al. Altered intra- and inter-network functional coupling of resting-state networks associated with motor dysfunction in stroke. Hum Brain Mapp 2018; 39: 3388–97. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Olafson E, Russello G, Jamison KW, et al. Frontoparietal network activation is associated with motor recovery in ischemic stroke patients. Commun Biol 2022; 5: 993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Dahms C, Noll A, Wagner F, Schmidt A, Brodoehl S, Klingner CM. Connecting the dots: motor and default mode network crossroads in post-stroke motor learning deficits. Neuroimage Clin 2024; 42: 103601. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Sudlow C, Gallacher J, Allen N, et al. UK Biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med 2015; 12: e1001779. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Liew SL, Zavaliangos-Petropulu A, Jahanshad N, et al. The ENIGMA Stroke Recovery Working Group: big data neuroimaging to study brain–behavior relationships after stroke. Hum Brain Mapp 2022; 43: 129–48. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Liew S-L, Zavaliangos-Petropulu A, Schweighofer N, et al. Smaller spared subcortical nuclei are associated with worse post-stroke sensorimotor outcomes in 28 cohorts worldwide. Brain Commun 2021; 3: fcab254. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Zijdenbos AP, Forghani R, Evans AC. Automatic “pipeline” analysis of 3-D MRI data for clinical trials: application to multiple sclerosis. IEEE Transactions on Medical Imaging 2002; 21: 1280–91. [DOI] [PubMed] [Google Scholar]
  • 19.Yeo BTT, Krienen FM, Sepulcre J, et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiology 2011; 106: 1125–65. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Liew S-L, Lo BP, Donnelly MR, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Sci Data 2022; 9: 320. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Defferrard M, Bresson X, Vandergheynst P. Convolutional neural networks on graphs with fast localized spectral filtering. Adv Neural Inf Process Syst 2016; published online June 30. https://arxiv.org/abs/1606.09375 (preprint). [Google Scholar]
  • 22.Shuman DI, Narang SK, Frossard P, Ortega A, Vandergheynst P. The emerging field of signal processing on graphs: extending high-dimensional data analysis to networks and other irregular domains. IEEE Signal Process Mag 2013; 30: 83–98. [Google Scholar]
  • 23.Smith SM, Vidaurre D, Alfaro-Almagro F, Nichols TE, Miller KL. Estimation of brain age delta from brain imaging. Neuroimage 2019; 200: 528–39. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Woytowicz EJ, Rietschel JC, Goodman RN, et al. Determining levels of upper extremity movement impairment by applying a cluster analysis to the Fugl-Meyer assessment of the upper extremity in chronic stroke. Arch Phys Med Rehabil 2017; 98: 456–62. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Archer DB, Vaillancourt DE, Coombes SA. A template and probabilistic atlas of the human sensorimotor tracts using diffusion MRI. Cereb Cortex 2018; 28: 1685–99. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Jilka SR, Scott G, Ham T, et al. Damage to the salience network and interactions with the default mode network. J Neurosci 2014; 34: 10798–807. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Peters SK, Dunlop K, Downar J. Cortico-striatal-thalamic loop circuits of the salience network: a central pathway in psychiatric disease and treatment. Front Syst Neurosci 2016; 10: 104. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Mattos DJS, Rutlin J, Hong X, Zinn K, Shimony JS, Carter AR. White matter integrity of contralesional and transcallosal tracts may predict response to upper limb task-specific training in chronic stroke. Neuroimage Clin 2021; 31: 102710. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Tønnesen S, Kaufmann T, de Lange AG, et al. Brain age prediction reveals aberrant brain white matter in schizophrenia and bipolar disorder: a multisample diffusion tensor imaging study. Biol Psychiatry Cogn Neurosci Neuroimaging 2020; 5: 1095–103. [DOI] [PubMed] [Google Scholar]
  • 30.Ward NS, Brown MM, Thompson AJ, Frackowiak RS. Neural correlates of outcome after stroke: a cross-sectional fMRI study. Brain 2003; 126: 1430–48. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Varghese R, Chang B, Kim B, Liew S-L, Schweighofer N, Winstein CJ. Corpus callosal microstructure predicts bimanual motor performance in chronic stroke survivors: a preliminary cross-sectional study. Top Stroke Rehabil 2023; 30: 626–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Hayward KS, Ferris JK, Lohse KR, et al. Observational study of neuroimaging biomarkers of severe upper limb impairment after stroke. Neurology 2022; 99: e402–13. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Domin M, Hordacre B, Hok P, et al. White matter integrity and chronic poststroke upper limb function: an ENIGMA stroke recovery analysis. Stroke 2023; 54: 2438–41. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS2147046-supplement-1.pdf (1.8MB, pdf)

Data Availability Statement

The de-identified participant data used for model training are available from the UK Biobank resource to bona fide researchers for health-related research in the public interest; applications for access can be made at https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. The full ENIGMA Stroke data are not publicly available in a repository as they might contain information that could compromise the privacy of research participants. Publicly sharing the full data is limited by data sharing restrictions imposed by some of the: (1) local ethical review boards of the participating sites, and consent documents; (2) national and trans-national data sharing laws; and (3) institutional processes, which might require a signed data transfer agreement for limited and predefined data use. However, we support data sharing among members of the ENIGMA Stroke Recovery Working Group who submit an analysis plan for a secondary project for group review. Following approval of the analysis plan, access to the relevant data will be provided, contingent on data availability, local principal investigator approval, and compliance with all supervening regulatory boards. More information about joining ENIGMA Stroke Recovery can be found here: https://enigma.ini.usc.edu/ongoing/enigma-stroke-recovery/. The source code for the regional brain age models developed in this study is openly available on GitHub at https://github.com/pks1207/regional_Brain_age.

RESOURCES