Purpose: The purpose of this study was to evaluate the effectiveness of implementation of an FDA-cleared artificial intelligence (AI) solution (Aidoc, Tel Aviv, Israel) for detecting intracranial vessel occlusions (VO) in a real-world stroke center. Materials and Methods: This study retrospectively analyzed all stroke cases performed at a single site academic medical center between December 2020 and April 2022 where a CT angiogram of the head and neck was part of the initial workup. An AI algorithm was used to analyze these head CT angiograms, and a natural language processing (NLP) tool was employed to classify the resulting radiology reports. Cases flagged as positive or discordant by the NLP tool were then reviewed by a radiologist to establish the true diagnosis (ground truth). Finally, diagnostic performance metrics were calculated. Results: A total of 1,000 CTA exams were analyzed, of which 33 were excluded due to findings of neck vessel occlusions. The AI algorithm demonstrated an overall sensitivity of 87.2% and specificity of 97.1%. The positive predictive value was 87.6%, and the negative predictive value was 96.9% as shown in Table 1. From a sub analysis of 179 patient cases with at least one occlusion present, 233 total vessel occlusions (VO) were individually identified and reviewed: 185 (79%) were anterior circulation occlusions (ICA, MCA, ACA segments) with an AI sensitivity of 87.6% for correctly detected anterior circulation vesel occlusions; and 48 (21%) were posterior circulation vessel occlusions (basilar, vertebral, PCA segments) with an AI sensitivity of 79.2% for correctly detected posterior vessel occlusions as shown in Table 2. Conclusion: The algorithm demonstrated strong diagnostic performance, supporting the potential value of AI integration into radiology workflows for evaluating both anterior and posterior circulation stroke occlusions. Clinical Relevance Statement: Artificial intelligence solutions that provide triage and alerts for suspected vessel occlusions can meaningfully augment stroke care workflows. This study reinforces the role of AI in improving VO detection.
Vermette et al. (Thu,) studied this question.