Post-thrombectomy models for outcome prediction in ischemic stroke: systematic review and external validation

Background Outcome prediction models for patients with ischemic stroke after endovascular thrombectomy (EVT) demonstrated the value of including post-procedural predictors. We systematically reviewed and externally validated models that incorporated post-procedural predictors. Methods We searched Medline, Embase, Cochrane, Web of Science, and Google Scholar for studies developing models predicting functional outcome after EVT, measured with the modified Rankin Scale score, that included post-procedural predictors. Model quality was assessed with a shortened PROBAST+AI checklist. External validation was performed in patients treated with EVT (n=1417) from three combined randomized controlled trials (MR CLEAN MED, MR CLEAN NOIV, and MR CLEAN LATE). Predictive performance was evaluated using discrimination (C statistic) and calibration (intercept and slope). Results Of 7665 screened studies, 54 were included; 30 were regression based models and 24 machine learning (ML) models. The number of predictors ranged from 2 to 50. Most frequently included post-procedural predictors were Thrombolysis in Cerebral Infarction score and post-procedural National Institutes of Health Stroke Scale (NIHSS) score. 50 models were rated as high risk of bias, mainly due to limited sample size and improper statistical methods. None of the ML based models could be validated due to non-routinely collected predictors or unavailable code. 12 regression based models were validated, with C statistics ranging from 0.67 (Lai et al ) to 0.90 (MR-PREDICTS @24H model). Four models including post-procedural NIHSS showed superior discrimination than models excluding this variable or using a dichotomized version, although two models were rated as low quality. Conclusions Regression based models showed moderate to excellent performance in predicting functional outcome after EVT. ML models were often difficult to validate due to non-routinely accessible predictors or lack of transparent reporting. Models incorporating strong clinical predictors, especially post-procedural NIHSS in continuous form, consistently showed better predictive performance.