Abstract DP130: Individualized Treatment in Distal and Medium Vessel Occlusion Stroke Using a Validated Counterfactual Simulation Model: Beyond One Size Fits All

Abstract

Introduction: The benefit of endovascular thrombectomy (EVT) in distal and medium vessel occlusion (DMVO) stroke remains uncertain, and strategies for patient selection are limited. Hypothesis: We hypothesized that a counterfactual predictive model (DUSK Oraculum) could simulate patient-specific outcomes under EVT versus medical management (MM), thereby guiding personalized treatment decisions. Methods: We analysed data from a multicenter cohort study of adults with isolated DMVO stroke involving the middle cerebral artery (MCA) M3/M4, anterior cerebral artery (ACA) A2/A3, or posterior cerebral artery (PCA) P1/P2 segments admitted to seven comprehensive stroke centers between 2017 and 2021. A core set of features, including time to treatment, thrombolysis, age, site of occlusion, and NIHSS variables, along with their treatment interactions, was consistently included in all models. The remaining candidate features underwent automated feature selection to identify the most predictive variables for a favorable 90-day modified Rankin Scale (mRS 0–2). Four machine learning models were developed and compared, including logistic regression, decision tree classifier, linear SVC, and XGBoost to evaluate treatment effect heterogeneity. Model performance was assessed using repeated 10-fold cross-validation (500 repetitions) with bootstrap resampling (1,000 iterations) to derive 95% confidence intervals for discrimination metrics. External validation was performed in an independent U.S. healthcare network, including two thrombectomy-capable centers. Results: Among 321 patients (EVT; n=179 [55.8%], MM; n=142 [44.2%]), adjusted analyses showed no significant differences in 90-day functional outcomes or mortality between treatment groups. Logistic regression with forward feature selection outperformed alternative models (Figures 1–2). Internal validation yielded a bootstrap AUC of 0.76 (95% CI, 0.71–0.81), with consistent performance in the external cohort (n=86; AUC 0.71, 95% CI, 0.60–0.81). A web-based application was created to enable real-time clinical use, allowing clinicians to input patient characteristics and receive individualized counterfactual treatment recommendations (Figure 3). Conclusions: The DUSK Oraculum reliably estimated patient-specific outcomes with EVT versus MM in DMVO stroke, supported by strong internal and external validation. This counterfactual modeling framework enables personalized, data-driven treatment decisions and may guide future randomized trials.

Ähnliche Arbeiten