Automated detection of large vessel occlusion: a multicenter study validating efficacy and proving clinical implications

Abstract

AbstractObjectiveWe aimed to validate a software, JLK-LVO, that automatically detects large vessel occlusion (LVO) on computed tomography angiography (CTA) using deep learning, within a prospective multicenter dataset. In addition, we calibrated the predicted probability of LVO against observed frequency and assessed the clinical implications of LVO probability in terms of follow-up infarct volume and functional outcome.MethodFrom 2021 to 2023, we prospectively collected data from patients who underwent CTA within 24 hours of symptom onset at six university hospitals in Korea. The diagnostic performance of the software was evaluated using the area under the curve (AUC), sensitivity, and specificity across the entire study population and specifically in patients with isolated middle cerebral artery (MCA)-M2 occlusion. In addition, we compared LVO probabilities after stratifying patient into acute LVO, chronic LVO, isolated MCA-M2 occlusion, relevant MCA stenosis, and without steno-occlusion of MCA groups. We calibrated LVO probabilities in two ways: through mathematical calibration using logistic regression, and by refining LVO probabilities based on the observed frequency of LVO. We then assessed the association of LVO probability categories with infarct volume on follow-up diffusion-weighted imaging (DWI) and modified Rankin Scale (mRS) scores three months post-stroke, using ANOVA and the Cochran–Armitage test.ResultsAfter excluding 168 patients, 796 remained; the mean (SD) age was 68.9 (13.7) years, and 57.7% were men. LVO was present in 193 (24.3%) of these patients, and the median interval from last known well to CTA was 5.7 hours (IQR 2.5 to 12.1 hours). At default threshold of 0.5, the software achieved an AUC of 0.944 (95% CI 0.926–0.960), with a sensitivity of 0.896 (0.845–0.936) and a specificity of 0.904 (0.877–0.926). In isolated MCA-M2 occlusion, the AUROC was 0.880 (95% CI 0.824–0.921). Compared to the without steno-occlusion of MCA groups (median LVO probability 0.5, interquartile range 0.1 – 6.5), relevant stenosis (median 15.3, 2.4 –77.4) and isolated MCA-M2 occlusion (82.1, 40.9 – 98.2) groups had significantly higher LVO probability. Due to sparse data between 20-60% of LVO probabilities, recategorization into unlikely (0-20% LVO scores), less likely (20-60%), possible (60-90%), and suggestive (90-100%) provided a reliable estimation of LVO compared with mathematical calibration. The category of LVO probabilities was significantly associated with follow-up infarct volumes on DWI and 3-month mRS scores.ConclusionIn this multicenter validation study, we proved the clinical efficacy of the software in detecting LVO on CTA. Additionally, using large-scale real-world data, we calibrated the LVO probabilities, which may provide a more confident estimation of LVO for practicing physicians.