Abstract Number: Esoc2026ys75 Subgroup-Specific Performance of Outcome Prediction Models After Endovascular Therapy for Acute Ischemic Stroke: An Eva-Trisp Registry Study

Abstract Background and aims Numerous prognostic models predict functional outcome after acute ischemic stroke, yet their performance in clinically relevant patient subgroups remains poorly characterized. Recent TRIPOD and PROBAST guidance emphasizes subgroup validation to identify bias and inequities. We externally validated established stroke outcome prediction models and assessed their subgroup performance after endovascular therapy (EVT). Methods Six multivariable models (S-SMART, NAC, THRIVE-EVT with and without ASPECTS, PRE, and MT-DRAGON) were evaluated using prospectively collected data from the multicenter EVA-TRISP collaboration of EVT-treated patients. The primary outcome was good functional outcome at 3 months (modified Rankin Scale mRS 0–2); a sensitivity analysis used an mRS-adjusted outcome accounting for prestroke disability. Discrimination (area under the receiver operating characteristic curve AUC), calibration (intercept and slope), and classification performance were assessed overall and across prespecified subgroups defined by age, prestroke mRS, ASPECTS, stroke territory, and sex. Bootstrapping was used to derive confidence intervals and compare subgroups. Results Among 12,734 patients, discrimination was similar across models (AUC 0.71–0.74) and calibration slopes were largely stable. In contrast, calibration intercepts varied substantially across subgroups. All models overestimated the probability of good outcome in very elderly patients, those with prestroke disability, low ASPECTS, combined anterior and posterior circulation strokes, and—across several models—in women, resulting in clinically relevant subgroup differences in sensitivity and specificity. Stroke outcome models with acceptable overall performance may therefore systematically misestimate outcomes in specific patient groups. Subgroup-aware validation and targeted recalibration are necessary to ensure equitable and reliable outcome prediction after EVT. Conflict of interest Lukas S. Enz, Marcel Arnold, Katarina Jood, Susanne Wegener, Patrik Michel, Sami Curtze, Gian Marco de Marchis, Paul Nederkoorn and Henrik Gensicke report no disclosures. Christian H. Nolte received honoraria for lectures/advisory boards from AstraZeneca, Bayer and Pfizer.