Abstract P506: Machine Learning Models Improve Prediction of Large Vessel Occlusion and Mechanical Thrombectomy Candidacy in Acute Ischemic Stroke

Abstract

Background and Purpose: Early identification of large vessel occlusions (LVO) and timely recanalization are paramount in improved clinical outcomes in acute ischemic stroke. Multiple simple stroke scales have good sensitivity, but compromise specificity for predicting LVO. No scale has been shown to predict mechanical thrombectomy (MT) candidacy. Machine learning techniques are being used for predictive modeling in many aspects of stroke care and may have potential in predicting LVO presence and MT candidacy. Methods: 287 acute ischemic stroke patients from July 2018 to July 2019 at Loyola University Medical Center were included. 36 clinical and demographic variables were analyzed using machine learning and statistical algorithms, including logistic regression, extreme gradient boosting, random forest, and decision trees to build models predictive of LVO and MT. The best performing model was compared with prior stroke scales. Results: Random forest based model resulted in the highest classification performance to predict both LVO and MT outcomes with an area under the curve (AUC) of 0.90±0.07 and 0.94±0.04, respectively. When the predictors were reduced to 7, random forest maintained a high AUC for predicting LVO (0.89). When reduced to 10 predictors, the random forest model predicted MT with an AUC = 0.93. Random forest models had excellent sensitivity and specificity of 0.86 and 0.89 for LVO and 0.89 and 0.95 for MT, respectively. The negative predictive value was 0.94 for LVO and 0.98 for MT while the positive predictive value was 0.77 for LVO and 0.79 for MT. With equal sensitivity, the random forest model was favorable to all previous stroke scales. Conclusion: Machine learning utilizing clinical and demographic variables predicts LVO and patient candidacy for MT with a high degree of accuracy. Further validation of this strategy for triage of stroke patients requires prospective and external validation.