Validation and simplification of the MR PREDICTS @24H model for outcome prediction after endovascular thrombectomy for acute ischemic stroke

BACKGROUND: Outcome prediction after endovascular thrombectomy (EVT) for ischemic stroke is important for patient counseling and rehabilitation planning. MR PREDICTS @24H, a nine-predictor model, showed excellent performance in predicting functional outcome at 90 days of patients with acute ischemic stroke. With the expanding treatment indications, we validated the model for patients receiving EVT within 24 hours after stroke onset and simplified it for easier clinical implementation. METHODS: We used individual patient data from the Dutch MR CLEAN Registry (2014-2018), a prospective observational cohort enrolling patients treated with EVT, and three randomized controlled trials MR CLEAN-MED, MR CLEAN-NOIV, and MR CLEAN-LATE (2018-2022). We included patients with an intracranial large-vessel occlusion in the anterior circulation treated with EVT within 24 hours of symptom onset or last seen well. We assessed the effect of predictors on functional outcome (modified Rankin Scale [mRS]) at 90 days with ordinal logistic regression. Predicted probabilities of functional independence (mRS 0-2) and survival (mRS 0-5) were derived from the model formula. We evaluated predictive performance with discrimination (C statistic) and calibration (intercept, slope). The model was simplified by excluding predictors based on the Akaike information criterion (AIC). We applied leave-one-study-out cross-validation to evaluate heterogeneity in model performance between the cohorts. RESULTS: The validation cohort included 6154 patients: 4737 from the Registry and 1417 from the trials. External validation of the original model showed excellent discrimination in predicting functional independence (C statistic 0.91, 95% CI 0.90-0.92) and survival (C statistic 0.90, 95% CI 0.89-0.91). The simplified model, comprising four predictors-NIHSS at 24 hours after EVT, age, pre-stroke mRS, and symptomatic intracranial hemorrhage-performed comparably (functional independence C statistic 0.91, 95% CI 0.90-0.92; and survival 0.89, 95% CI 0.88-0.90). Cross-validation revealed heterogeneity between LATE and the other cohorts, with the model overestimating the probability of functional independence in LATE (observed 39.1% vs predicted 44.2%), whereas the observed and predicted probability of survival was similar (75.5% vs 75.7%). CONCLUSIONS: A simplified version of MR PREDICTS @24H including only four predictors performed as good as the full model, providing a practical tool that can be applied one day after EVT for reliable outcome estimation. Further validation and updating of the model are needed for patients treated in the late time window (6-24 h).