Abstract P510: Ct Angiographic Radiomics Signature Predicts Functional Outcome Following Endovascular Thrombectomy in Large Vessel Occlusion Stroke

Abstract

Purpose: Endovascular thrombectomy (ET) is the standard of care in large vessel occlusion (LVO) stroke, with CTA and CT/MR perfusion guiding patient selection. We hypothesized that radiomics imaging features extracted from admission CTAs could predict post-ET outcome. Methods: We included patients with anterior circulation LVO who had ET at our institute, 01/2013–12/2019. We extracted 1116 radiomics features from each MCA supply territory. We applied and evaluated a framework of 6 feature selection techniques and 6 machine-learning classifiers for prediction of discharge and 3-month follow-up outcome, defined as favorable (modified Rankin score, mRS≤2) vs poor (mRS>2). Post-ET reperfusion success was determined by the modified thrombolysis in cerebral infarction (mTICI) scale. We used Bayesian optimization for hyperparameter tuning and performed 20 repetitions of 5-fold cross-validation, for which the average area under the receiver operating characteristic curve (AUC) across validation folds for each of our 36 feature-selection/machine-learning combinations was calculated. Results: 501 patients (228 male) were included, with mean age 70.3±15.5 years, median NIH stroke score 15 (interquartile range=6–24), and occlusions in ICA (n=123), M1 (n=318), and/or M2 (n=154). Functional outcome was available for 496 patients at discharge and for 375 at 3-months. Best performing models combining NIHSS, age, gender, IV thrombolytic treatment, and post-ET mTICI achieved an average AUC of 0.82±0.05, while models trained on radiomics and mTICI achieved an AUC of 0.71±0.05. Conclusion: The combination of automatically extracted CTA radiomics features and post-ET reperfusion success (mTICI) can predict LVO stroke functional outcome – even without baseline clinical variables. Such models may guide treatment decisions by predicting outcome for various degrees of post-ET reperfusion and automating assessment of baseline stroke CTA scans.