ABSTRACT NUMBER: ESOC2026A1366 EARLY AND 24-HOUR XGBOOST MODELS PREDICT 90-DAY FUNCTIONAL OUTCOME IN POSTERIOR CIRCULATION THROMBECTOMY: DERIVATION AND EXTERNAL VALIDATION OF THE “FORECAST-PC” PREDICTION SCORES

Abstract Background and aims Outcome prediction after endovascular treatment (EVT) for posterior circulation (pc) acute ischemic stroke remains challenging; validated prognostic tools tailored to this population are lacking. We aimed to develop and externally validate “Functional Outcome pRediction in postErior Circulation Acute ischemic Stroke Thrombectomy” (FORECAST) pc-models at different stages of acute management. Methods We included patients with pc stroke selected for EVT within 24 hours from last seen well and with prestroke modified Rankin Scale (mRS) ≤2. Gradient-boosted decision tree models were developed at four stages: baseline, pre-EVT, immediately post-EVT, and 24-hours post-EVT (Figure 1) in “full” and “light” versions. The primary outcome was unfavorable functional outcome (90-day mRS=3–6). Models were derived and cross-validated in 384 patients from 3 centers, then externally tested in 247 patients from 11 unseen international centers (Figure 2). Results Among 384 derivation cohort patients (median age 73 years [IQR=62-81], 147/384 [38%] female),173/384 (45%) had unfavorable 90-day outcomes. In external validation (Figure 1), models A=baseline, B=pre-EVT, and C=immediately post-EVT showed similar discrimination (AUCA=0.72, AUCB=0.72, AUCC=0.72), whereas model D=24-hour demonstrated higher performance (AUCD=0.89, baseline-vs-24h AUC p<0.001). Light models performed comparably to full models (full-vs-light AUC: baseline p=0.09; 24h p=0.65. Age and admission NIHSS were major contributors, with 24-hour NIHSS being the strongest predictor. Conclusions In a large multicenter cohort of pc stroke patients selected for EVT, stage-specific models, particularly early and 24-hour models, reliably predicted 90-day functional outcome. Externally validated FORECAST prediction scores provide a pragmatic, stage-specific tool for prognostication in pc stroke. Conflict of interest All authors have nothing to disclose. Figure 1 - belongs to Methods Figure 2 - belongs to Results