Individualized Treatment in Distal and Medium Vessel Occlusion Stroke Using a Validated Explainable Counterfactual Treatment Estimation Model.

OBJECTIVE

The optimal treatment for distal medium vessel occlusion (DMVO) stroke remains uncertain, and evidence comparing endovascular therapy (EVT) with medical management (MM) is limited. We aimed to develop and validate a predictive modeling tool to assess individual treatment benefit in DMVO stroke using explainable counterfactual treatment estimation.

METHODS

Adults with isolated DMVO stroke (M3-M4, A2-A3, or P1-P2) were retrospectively identified from 7 stroke centers. To estimate individualized probabilities of favorable outcome (modified Rankin Scale [mRS] = 0-2 at 90 days), we developed a Penalized Logistic Regression (Elastic Net) model. This framework was selected a priori over other explored machine learning algorithms (Decision Tree, Support Vector Classifier, and XGBoost) for its superior interpretability and ability to handle multicollinearity among interaction terms. Inverse Probability of Treatment Weighting (IPTW) was implemented to address confounding by indication in the observational data. Internal validation used repeated K-fold cross-validation and bootstrapping; external validation was performed on an independent cohort (n = 86).

RESULTS

Of 321 eligible patients, 179 received EVT (55.8%) and 142 received MM (44.2%). Adjusted models showed no significant overall group differences in favorable outcome (adjusted OR [aOR] = 1.32, 95% confidence interval [CI] = 0.97-1.80), mortality (aOR = 1.20, 95% CI = 0.78-1.85), or symptomatic hemorrhage (aOR = 0.57, 95% CI = 0.21-1.58). However, the model identified significant treatment effect heterogeneity; EVT benefit was amplified in patients with higher National Institutes of Health Stroke Scale (NIHSS) and attenuated with increasing treatment delay. Internal validation demonstrated strong performance (area under the receiver operating characteristic curve [AUC] = 0.77, 95% CI = 0.71-0.82). External validation confirmed generalizability (AUC = 0.74, 95% CI = 0.63-0.84). Individualized treatment estimates showed high concordance with a benchmark causal T-Learner model (Pearson r = 0.97 internal and r = 0.98 external).

INTERPRETATION: Although aggregate outcomes did not differ significantly, the validated Distal and Medium Vessel Occlusion Stroke (DUSK) Tool enables individualized estimation of EVT benefit in DMVO stroke. This explainable counterfactual treatment estimation framework supports precision decision making by identifying specific patient subgroups most likely to benefit from EVT over MM. ANN NEUROL 2026;99:1198-1209.