ABSTRACT NUMBER: ESOC2026A1793 WHEN EXPERTISE MEETS ALGORITHMS: EVALUATING MODEL-ASSISTED PROGNOSTICATION IN ACUTE ISCHEMIC STROKE

Abstract Background and aims Accurate outcome prediction after acute ischemic stroke due to large vessel occlusion is critical for treatment decisions but remains challenging. Prognostic tools like MR PREDICTS may support clinicians, though their benefit depends on how accurately neurologists estimate model inputs and incorporate predictions. This study quantified the added value of MR PREDICTS and examined challenges that could be addressed with multi-modal deep learning (DL) approaches that automatically extract imaging features. Methods Using MR CLEAN trial data, six neurologists predicted 24-hour NIHSS and 90-day mRS in 40 patients based on clinical and CT imaging data, with and without MR PREDICTS. Performance was assessed using multiple metrics. Predictions were compared with a DL model trained in a cross validation on the same data. Results MR PREDICTS modestly improved ordinal mRS (QWK 0.27 vs. 0.41) and 24-hour NIHSS prediction (QWK 0.42 vs. 0.53). Estimating model inputs was challenging for neurologists (occlusion location accuracy 17–70%; ASPECTS deviation 3.36 points, collateral score accuracy 44.5%), resulting in an average 5.3% deviation from the estimated MR PREDICTS benefit. QWK of MR PREDICTS alone was 0.51 for ordinal mRS, with AUC 0.77 [0.69, 0.83] and 70% accuracy for binary mRS prediction. The DL model performed similarly for the 40 patients (AUC 0.77 [0.37, 0.9]; 73.7% accuracy) without requiring input estimation. Conclusions MR PREDICTS can modestly improve neurologist predictions, but challenges in estimating inputs limit its utility. Automated DL-based assessment offers a scalable alternative by directly learning imaging features, potentially improving outcome prediction in acute stroke care. Conflict of interest None