Where the Lesion Spreads: Peri-Infarct Patient Graphs for Stroke Recovery Beyond Baseline

After a stroke, the dead tissue at the core of the lesion is only part of the damage. In mice, immune cells of the brain called microglia go on to "prune", or eat, healthy synapses in the rim of living tissue around the lesion. This activity peaks about a week after injury, and blocking it rescues motor function. We propose to test whether this rim, not the lesion core, is what makes stroke patients similar to one another in how much they will recover. We would connect patients into a similarity graph by how much their peri-infarct rims overlap in motor cortex. We would then ask whether the recovery of a patient's graph neighbours predicts the part of that patient's recovery that their initial severity does not already explain.

This matters because the standard way of building such "population graphs", in which patients are nodes and similarity defines edges, is in trouble. Careful benchmarks show that a well-tuned random forest matches or beats graph neural networks on most medical datasets. Stroke recovery is also a regime with thirty to a hundred patients per cohort, where baseline impairment alone explains most outcome variance. The opportunity is not a bigger graph model. It is a biologically motivated definition of the edges, one that encodes a pairwise spatial relation that cannot be flattened into a feature column. That definition is tested against a target and a timing window where it has a real chance to add information.

Core Claim

Our claim is that overlap between patients' peri-infarct rims in sensorimotor cortex carries prognostic information about recovery beyond baseline impairment. We further claim this information should appear only when imaging is acquired before the subacute pruning window closes.

Concretely, let $R_i$ be patient $i$'s rim mask. It is the lesion dilated by a fixed margin in standard space, minus the lesion itself, intersected with a sensorimotor cortical mask. The edge weight between patients $i$ and $j$ is

$$w_{ij} = \frac{|R_i \cap R_j|}{|R_i \cup R_j|},$$

the fraction of their combined rim territory that they share. Let $y_i$ be the residual recovery after regressing the outcome on baseline impairment within training folds. The prediction is a weighted average of neighbours' residuals:

$$\hat{y}_i = \frac{\sum_{j \in \text{train}} w_{ij}\, y_j}{\sum_{j \in \text{train}} w_{ij}}.$$

We deliberately use the simplest relational model. Are Population Graphs Really as Powerful as Believed? shows that graph architectures such as GCN, GAT and sheaf diffusion rarely beat tuned tabular models. The graph neural networks that did well often succeeded by ignoring the topology altogether. The same paper challenges the field to find edges that encode genuinely relational information. Our proposal answers that challenge on the edge side, where the paper's critique actually bites, rather than on the architecture side.

The pruning biology plays the role of motivation and constraint, not measurement. STING inhibition suppresses microglia-mediated synapses engulfment and alleviates motor functional deficits after stroke shows that microglial engulfment of synapses in the peri-infarct cortex peaks around 7 days after injury in mice, and that inhibiting it improves forelimb function. Clec7a Worsens Long‐Term Outcomes after Ischemic Stroke by Aggravating Microglia‐Mediated Synapse Elimination shows that microglia preferentially remove excitatory synapses, and that knocking out the Clec7a receptor rescues long-term behaviour. Both are rodent studies with histological readouts; neither offers a human proxy.

We cannot measure pruning in patients either. What we can test is the spatial and temporal signature that this biology implies. That signature is prognostic information concentrated in the rim rather than the core, present when patients are scanned acutely and absent once the window has passed. A positive result would be the first population-level human evidence consistent with that signature. A negative result would constrain how much of human motor outcome this mechanism can explain.

Research Program

The first component is building the edges. Lesion masks for each outcome cohort will be segmented and normalized to a common template. The rim margin, from a few millimetres up to about a centimetre, is chosen by inner cross-validation from a small grid.

ATLAS v2.0 contributes 955 T1-weighted scans with manual lesion masks, described in A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. It has no motor outcomes, so we use it only for two things. It supplies the segmentation model for the outcome cohorts. It also supplies a reference distribution of rim overlaps, so that edge weights can be calibrated against how often two random strokes share rim territory. Public segmentation methods on ATLAS reach Dice scores of only 0.5 to 0.7, a measure of spatial overlap with the manual masks. Rims derived from automated masks will therefore be noisy. Where manual masks exist, we will compare rims from manual and automated masks directly.

The second component is the target. Predicting response to motor therapy in chronic stroke patients using Machine Learning reports an Elastic-Net R² of 0.910 for post-therapy Fugl-Meyer scores, a standard clinical scale of motor impairment. That study also found that structural disconnectivity from a network-modification tool added nothing to the prediction. Because baseline Fugl-Meyer dominates that fit, any method judged on raw outcome will look good or bad mostly according to how it handles baseline.

We therefore evaluate all methods on residual recovery. This is the part left after a within-fold linear adjustment for baseline impairment, and, in the chronic cohort, for the interhemispheric motor-threshold difference. This residual target is the honest place to ask whether anything new is learned.

The third component is the timing contrast, the proposal's main scientific lever. The Cologne cohort in Dynamic connectivity predicts acute motor impairment and recovery post-stroke was scanned about 2.5 days after stroke, before the rodent pruning peak, and followed to six months on the Motricity Index, an upper-limb motor score. The chronic cohort of Tozlu and colleagues was imaged 3 to 12 months after stroke, long after the window. By then, whatever the rim underwent is already written into baseline impairment.

The hypothesis therefore predicts a dissociation. Rim edges should carry residual signal in the acute cohort and little in the chronic one, whereas core-lesion edges should behave similarly in both. The hypothesis can fail in either cohort, which is what makes the contrast informative.

The fourth component is a secondary, exploratory link to dynamic connectivity. The Cologne study found an "integration paradox": less time spent in a globally integrated functional connectivity state acutely predicted better six-month recovery. We will test whether rim overlap and fraction of time in the integrated state explain overlapping or distinct parts of residual recovery. The test fits both jointly and compares each one's unique contribution. We treat any interpretation in terms of pruning as speculative.

Small-sample hygiene is built in throughout. Graphs are inductive: test patients attach to training patients through their rims only, and edges among test patients or to test labels are never used. All adjustment and margin selection happens inside nested cross-validation. Significance comes from label-permutation tests that rerun the entire pipeline. The analysis plan, including the rim-versus-core contrast and the acute-versus-chronic dissociation, is fixed before outcomes are unblinded.

The two outcome cohorts are not public, so the work depends on data-sharing agreements with the originating groups. We state this constraint openly rather than assume access.

Baselines

The naive reference is to predict zero residual recovery. On raw outcomes, the equivalent is a linear model on baseline impairment alone. Beating the naive reference means any significantly positive out-of-sample R² on the residual target, established by permutation over the whole pipeline.

The strongest clinical baseline is the Elastic-Net from Predicting response to motor therapy in chronic stroke patients using Machine Learning. It reports R² = 0.910 for post-intervention Fugl-Meyer using baseline scores and neurophysiology. We rerun it with the same nested cross-validation and hyperparameter budget as our model. We then evaluate it both on raw outcome, to reproduce the reported regime, and on residual outcome, which is the comparison that counts.

The strongest imaging baseline for the acute cohort is the random forest on dynamic connectivity features from Dynamic connectivity predicts acute motor impairment and recovery post-stroke. It reports an AUC of 0.89 ± 0.01 for six-month recovery, using fraction time in the integrated state combined with initial impairment. We rerun it on the same thirty follow-up patients, with the same folds, and report both its binary AUC and its residual R². Beating it means rim edges improve residual prediction when added to its features, or match it with fewer free parameters.

The strongest single-subject lesion model is Gaussian process regression on voxel-wise lesion patterns, from Decoding post-stroke motor function from structural brain imaging. It reports R = 0.83 against R = 0.51 for corticospinal tract lesion load, but cross-sectionally in chronic patients. We refit it on lesion voxels within each outcome cohort, giving it the same inner tuning budget.

The strongest learned-graph baseline is the pairwise association encoder of Edge-variational Graph Convolutional Networks for Uncertainty-aware Disease Prediction, which learns edges from phenotypic features. It reports 81.06% accuracy on ABIDE, but it has no reported result on stroke recovery. We train it on the same features, with the same epoch and hyperparameter budget, and use its edge-dropout uncertainty as a diagnostic.

The meta-learned graph model of ADMGCN: Graph Convolutional Network for Alzheimer's Disease Diagnosis with a Meta-learning Paradigm targets classification of dementia status. We include it only if a regression adaptation proves stable.

The ablations isolate the rim itself. The first uses core overlap, meaning edges from overlap of the lesions themselves. The second uses a shifted rim of identical volume placed in non-motor cortex. The third is a k-nearest-neighbour graph on tabular features, the construction Müller and colleagues criticize. The key ingredient counts as real only if the sensorimotor rim beats all three on the acute residual target.

Why not just…?

Why not just use the Elastic-Net on baseline scores? For raw outcomes, it is probably unbeatable in chronic patients, and we expect it to win there; its R² of 0.910 comes mostly from baseline impairment. The open question is the residual, which is precisely what the Elastic-Net does not model. Our experiments report the Elastic-Net on that residual target, so if its residual R² already absorbs what rim edges offer, that will be visible.

Why not just fit a Gaussian process with a lesion-overlap kernel instead of calling it a graph? This is a real threat, because our neighbour-averaging predictor is mathematically close to kernel regression with $w_{ij}$ as the kernel. We concede the equivalence and do not claim architectural novelty. The contribution is the choice of kernel, rim versus core versus voxel pattern. The rim-kernel Gaussian process will be run as a variant of our method, and the voxel-pattern Gaussian process as the competitor. If voxel-level lesion patterns explain the residual as well as rims do, then the rim framing adds nothing, and we will report that.

Why not just use a random forest on dynamic connectivity, which already reaches an AUC of 0.89? That result predicts recovery with initial impairment included, from thirty patients, and does not separate out the residual. Rim overlap is a different, structural source of information. The unique-contribution analysis will show whether the two sources are redundant. If rim edges add nothing beyond dynamic connectivity, the simpler functional model stands.

Why not just blame corticospinal tract damage? Tract lesion load is a strong prognostic variable, but Decoding post-stroke motor function from structural brain imaging shows it captures much less than finer lesion patterns (R = 0.51 vs 0.83). We include tract load in the residual adjustment, so rim edges are credited only with information beyond it.

Proposed Experiments

The primary experiment compares residual R² for rim-overlap edges against core-overlap and shifted-rim edges in the Cologne acute cohort, using permutation p-values from the full pipeline. The second experiment repeats this in the chronic cohort, where the hypothesis predicts little or no rim advantage. The decisive check is the interaction: a rim advantage acutely that vanishes chronically.

If a rim advantage appears in both cohorts, the effect is spatial but not tied to the window, which counts against the pruning interpretation. If it appears in neither, the rodent mechanism leaves no detectable population-level structural trace at these sample sizes. We will report the minimum effect detectable with thirty and 102 patients, so that a null result is read correctly.

A robustness check reruns everything with manual versus automated masks, calibrated against ATLAS v2.0, and across a range of rim margins. Its purpose is to show that any effect does not hinge on one arbitrary dilation.

Finally, we report every baseline on both raw and residual targets, so that the headline R² of 0.910 and our residual numbers can be read side by side.