Characteristics for Machine Learning Detection of Large Vessel Occlusion on Computed Tomography Angiography

Detection of large vessel occlusion (LVO) using machine learning on computed tomography angiography (CTA) may help stroke triage, yet applicability across varied patient and image characteristics has not been examined. The study will examine which characteristics are important when using a convolutional neural network to identify LVO on CTA. A retrospective cohort study (November 2017-May 2019) at a comprehensive stroke center evaluated 677 stroke-alerted patients with an LVO of the internal carotid artery, M1, or M2 (n=150) and a matching number without LVO were included. An Inception module-based network was trained for binary classification of LVO presence. Results were examined by LVO location, window settings, non-LVO findings, demographics, risk factors, presentation status and times, interventions, and outcomes. Three hundred patients were included (48% women; median age 65). Mean{+/-}95% CI for cross-validation test and external validation, respectively, are area under precision-recall curve 0.871{+/-}0.094 and 0.742{+/-}0.018 and area under receiver operating characteristic curve 0.920{+/-}0.051 and 0.852{+/-}0.004. 145 true positive (TP), 5 false negative (FN), 39 false positive (FP), and 111 true negative (TN) patients were identified. Significant comparisons (P<0.05) identified: lower window settings for misclassifications, smoking history for all FN versus 33% TP (P=0.005), and tissue plasminogen activator treatment for 41% FP versus 20% TN (P=0.017). Our LVO detection tool had high performance across patient characteristics with few exceptions. FP had pathology warranting detection, including distal occlusions. Lower window settings among misclassifications highlight the need for image quality when using machine learning for decision support.