Abstract WP295: Post-procedural Outcome Prediction Models for Ischemic Stroke Patients after Endovascular Treatment: Systematic Review and External Validation

Background: Prediction models to inform ischemic stroke patients about functional outcome after endovascular thrombectomy (EVT) have highlighted the significance of including post-procedural predictors. We aim to systematically review and externally validate existing models, focusing on those incorporating post-EVT predictors. Methods: We searched within Medline, Embase, Cochrane, Web of Science, Google Scholar to include studies that developed models which included post-EVT predictors to predict functional outcome measured with the modified Rankin Scale (mRS) at 3 months. Methodological quality was assessed using a shorter version of PROBAST+AI and models were validated in the combined data of three RCTs (MR CLEAN-MED, MR CLEAN-NOIV, and MR CLEAN-LATE). Predictive performance was evaluated based on discrimination (C-statistic) and calibration (intercept, slope). Results: From 7662 screened studies, 54 were included: 28 using traditional regression methods, and 26 employing machine learning (ML) to predict outcome. Number of predictors ranged from 2 to 50. Most frequently used post-EVT predictors were TICI score and post-EVT NIHSS. 51 models were assessed with low quality using a shorter version of PROBAST+AI, primarily due to limited sample size and improper statistical methods. 12 models were validated in EVT patient data (n=1417), with C-statistics ranging from 0.67 (Lai et.al) to 0.90 (MR-PREDICTS @24H). Best calibrated model was MR PREDICTS @24H (intercept: 0.11, slope: 1.07). Models that included post-EVT NIHSS performed significantly better than models that did not. ML-based models could not be validated due to unavailability of data on predictors or model code. Conclusion: Regression-based models showed moderate to excellent performance, while ML models remain challenging to validate due to numerous required predictors or limited code sharing. The inclusion of strong clinical predictors limits the impact of poor methodological quality on predictive performance. For clinical implementation, models should prioritize methodological rigor, data availability, and clinical usability.