Abstract MP32: Machine Learning Models Identify Predictors of Poor Outcome in Patients Undergoing Mechanical Thrombectomy for Acute Ischemic Stroke

Abstract

Introduction: Predicting outcome after mechanical thrombectomy (MT) for ischemic stroke due to LVO can inform prognosis and guide early management. Prior studies report heterogeneity in risk factors for poor outcome. Machine learning may identify patterns of poor outcome from diverse variables that are difficult to discern with conventional statistical methods. Methods: Using a retrospective database of 233 stroke patients (2015-20) who had MT for LVO, we created machine learning predictive models with clinical and imaging variables for the following 4 outcomes: decompressive craniectomy, discharge mRS ≥4, development of post-stroke cerebral edema with mass effect, and in-hospital mortality. We compared 10 learner models: AdaBoost, Tree, Random Forest, Neural Network, CN2 Rule Induction, Logistic Regression, Naïve Bayes, kNN, Stochastic Gradient Descent, and Support Vector Machine. Variables were ranked by 5 scoring methods: information gain, information gain ratio, gini decrease, chi-square, ReliefF, and fast correlation-based filter. A prediction model was created using the top 5 variables to maximize the area under the receiver operating characteristic curve and classification accuracy. Models were 5-fold cross validated. Analyses were conducted via Stata and Orange Data Mining. Results: Prediction model sets of 5 variables were generated for the 4 outcomes of interest. Infarction volume was most important for predicting decompressive craniectomy, discharge mRS ≥4, and in-hospital mortality. Cerebral edema was important for decompressive craniectomy, discharge mRS ≥4, and in-hospital mortality. Initial NIHSS was important for decompressive craniectomy, discharge mRS ≥4, and in-hospital mortality. Contrast staining on post-procedural CT was important for cerebral edema (χ 2 11.9) and in-hospital mortality (χ 2 21.8). Patient age was important for discharge mRS ≥4 and decompressive craniectomy. Conclusion: We identified prediction models consistent with established prognostic variables. Post-MT contrast staining is a novel and important predictor of poor outcome, which merits further research. In conclusion, machine learning can be used to create accurate prediction models for outcome after MT for ischemic stroke with LVO.