Background & Purpose: Patient selection for endovascular thrombectomy (EVT) is still a challenging task for neuroradiologists. Identifying patients at the earliest stage of presentation that might benefit the most from by EVT or vica versa is an imperative (1). Here, we investigated whether machine learning (ML) workflows can support interventionalists in patient selection based on early-phase clinico-radiological and laboratory data by predicting poor outcome (2). Methods: A single-center retrospective cohort of 172 (90 M; 52.3%) consecutive patients undergoing EVT in 2017-2018 was retrieved from local RIS/PACS. Admission ASPECTS was extracted from reports using NLP3 and re-evaluated by two blinded readers on imaging. Explanatory variables included age, sex, comorbidities and blood rheology parameters as well as neuro-interventional procedural data on time, retrieval count and final Thrombolysis in Cerebral Infarction following angiography. The primary outcome was the modified Rankin Scale (mRS) score at hospital discharge. Poor outcome was defined as mRS 5-6 (98; 56.9%). Previously described multistage 5-fold cross-validated ML-workflows using random forests (RF) were applied to subsets of the features available at pre- and post-EVT (2). Results: All pre- and post-EVT features were available for 140 cases. Eighty-five cases (60.7%) had poor outcome. The pre-EVT-RF model showed an accuracy of 65% while the post-EVT-RF model achieved slightly higher performance of 67.9%. Conclusion: ML-supported patient selection for optimized EVT outcome is feasible, however, this is a hard task at the earliest stage of diagnosis even when considering several clinico-radiological and laboratory parameters.