TL;DR: A 2026 study of 501 chronic stroke survivors in The Lancet Digital Health found a deep-learning brain-age model mapped both lesion-related aging in the damaged hemisphere and a younger-looking opposite hemisphere linked to worse motor impairment.
Key Findings
- The damaged side aged upward, the spared side aged downward: Larger lesions raised ipsilesional regional brain age (β = 0.5420 to 0.9458) and lowered contralesional ventral attention/language brain age (β = −0.3747; 95% CI −0.6961 to −0.0534).
- “Younger” tissue marked impairment, not recovery: Worse motor scores tracked younger-appearing contralesional networks — interpreted as compensatory recruitment, not normalized function.
- Three top motor predictors emerged: Corticospinal tract lesion load, salience-network lesion load, and contralesional frontoparietal brain age — combining direct damage with network reorganization.
- 17,791 UK Biobank scans trained the model: A graph convolutional network learned regional brain age from a large independent cohort before being applied to the stroke sample.
- 18 functional subregions instead of one number: The whole-brain age summary would have washed out the recovery biology. Regional decomposition exposed it.
- 34 cohorts, 8 countries, all 180+ days post-stroke: ENIGMA-scale data made the finding less likely to be a single-scanner artifact.
Source: The Lancet Digital Health (2026) | Park et al.
Stroke recovery depends on both the damaged tissue and the brain regions that reorganize around that damage.
The first half of that story has been imageable for decades; the second half has been harder to capture. This study uses a deep-learning brain-age model to make the reorganization visible — and the finding is more specific than “younger is better.”
Regional MRI Brain Age Captured Stroke Recovery Patterns
Brain-predicted age difference (brain-PAD) compares how old a brain region looks on MRI with the patient’s actual chronological age. Positive values mean older-looking tissue; negative values mean younger-looking tissue.
Most uses of brain-PAD report one whole-brain number, which can be informative for aging or neurodegeneration but is the wrong granularity for stroke.
Stroke is focal, recovery is networked, and a single global brain age would smear the core biology. The team got around this with a graph convolutional network trained on 17,791 UK Biobank scans, then applied it to 501 chronic stroke survivors across 18 predefined functional subregions.
Each region got its own brain-age estimate. The geography of recovery emerged from the differences between them.
Stroke Lesions Made the Damaged Hemisphere Look Older
Larger total lesion size was associated with higher regional brain age in the ipsilesional hemisphere — the side with the stroke looked biologically older across multiple regions, with betas ranging from 0.54 to 0.95. That fits the obvious intuition.
A large lesion disrupts tissue, connectivity, and network health, and the model captured that disruption as accelerated regional aging.
If the evidence stopped there, it would be a confirmatiin the model. The evidence does not stop there.
Younger-Looking Opposite Hemisphere Tracked Worse Motor Impairment
Larger lesions were also associated with lower brain-PAD in a contralesional ventral attention and language network region. The opposite hemisphere looked younger relative to age — and the structural-equation modeling tied that younger-appearing tissue to worse motor impairment, not better recovery.
That asymmetry complicates the brain-age vocabulary. “Younger” sounds reflexively positive, but in this clinical context, younger-appearing contralesional tissue marks the brain’s attempt to adapt to severe damage — compensatory recruitment of undamaged networks when the injured motor system cannot carry the load.
It is a sign of strain, not of healing.
The clinical reading needs to absorb that nuance. Compensation can be adaptive, inefficient, temporary, or even maladaptive depending on timing, circuit, and patient.
A biomarker that captures compensatory recruitment helps identify which patients need therapies aimed at strengthening spared networks versus restoring activity in damaged pathways. It does not say “this patient is recovering” on its own.
Three Brain-Age Features Predicted Motor Outcome
The machine-learning models identified 3 top predictors of motor outcomes after stroke:
- Corticospinal tract lesion load: direct damage to the descending motor pathway predicted poorer motor scores.
- Salience-network lesion load: attention and control circuitry added network-level information.
- Contralesional frontoparietal brain-PAD: younger-looking spared networks captured compensatory recruitment.
The corticospinal tract result is no surprise. The descending motor pathway remains the single clearest predictor of motor recovery, and damage to it predicted worse outcomes (β = −0.355; 95% CI −0.446 to −0.267; P<0.0001).
The salience-network finding is more network-level. Motor recovery is not only about muscle commands.
Planning, error monitoring, attention, and the selection of which actions to attempt all influence how movement is relearned. Damage to attention and control circuitry shows up here as a predictor independent of the corticospinal hit.
The third predictor — contralesional frontoparietal brain age — is the conceptually new one. It says that MRI-derived regional age captures something conventional lesion maps miss: the active reorganization of undamaged networks.
2 patients with similar lesion volumes can show very different recovery biology depending on what their spared networks are doing.
ENIGMA Stroke Scale Strengthened the 501-Patient Brain Age Analysis
Single-center stroke imaging studies are hard to generalize. Scanner protocols, segmentation methods, rehabilitation settings, and outcome scoring all vary across hospitals.
The analysis draws strength from ENIGMA scale — 34 cohorts across 8 countries, plus a large independent UK Biobank set for training the age model.
Harmonization is not perfect across that footprint, but the breadth makes the regional brain-age pattern less likely to be an artifact of one scanner or one rehabilitation program.
The retrospective, observational design limits causal claims because the MRI model measured associations after stroke rather than assigning rehabilitation strategies. It cannot prove that a younger-looking contralesional region causes compensation, and it cannot say which therapy should be assigned to a specific patient.
It does identify the regional brain-age biology that future therapy-selection trials would need to test.
Stroke Imaging May Combine Lesion Maps With Brain-Age Estimates
Traditional lesion mapping is not going anywhere. The corticospinal tract result confirms it. Regional brain-age modeling should pair with lesion analysis, not replace it.
The realistic future is combination — lesion load, tract disruption, regional brain age, functional connectivity, and clinical performance evaluated together.
2 patients with similar lesion volumes might show very different recovery biology: one with severe corticospinal disruption and little compensatory pattern, another with active contralesional recruitment that targeted therapy could try to harness.
The current analysis does not yet deliver that combined clinical tool.
It gives the field a clear reason to build it.
AI-derived brain age becomes more useful when it separates damaged tissue from compensatory networks after stroke.
The next study should ask whether those regional patterns predict which therapies actually work for which patients while there is still time to shape plasticity.
Citation: DOI: 10.1016/j.landig.2025.100942; Park et al; Associations between contralesional neuroplasticity and motor impairment through deep learning-derived MRI regional brain age in chronic stroke (ENIGMA): a multicohort, retrospective, observational study; The Lancet Digital Health; 2026.
Study Design: Multicohort retrospective observational MRI study using a deep-learning regional brain-age model.
Sample Size: 501 chronic unilateral stroke survivors from 34 cohorts in 8 countries; 17,791 UK Biobank scans for model training.
Key Statistic: Higher corticospinal tract lesion load predicted poorer motor outcomes (β = −0.355; 95% CI −0.446 to −0.267; P<0.0001), and worse motor scores tracked younger contralesional brain age (β = −0.3747; 95% CI −0.6961 to −0.0534).
