Clinical relevance of intracranial stenosis as false-positive findings of a deep learning algorithm trained to detect large vessel occlusions: a retrospective cohort study of a supraregional stroke centre.
Authors
Affiliations (6)
Affiliations (6)
- Anaesthesiology, BG Klinikum Unfallkrankenhaus Berlin gGmbH, Berlin, Germany.
- Institute of Radiology and Neuroradiology, BG Klinikum Unfallkrankenhaus Berlin gGmbH, Berlin, Germany.
- Center for Clinical Research, BG Klinikum Unfallkrankenhaus Berlin gGmbH, Berlin, Germany.
- Neurology, BG Klinikum Unfallkrankenhaus Berlin, Berlin, Germany.
- Institute for Diagnostic Radiology and Neuroradiology, Universitätsmedizin Greifswald, Greifswald, Germany.
- Institute of Radiology and Neuroradiology, BG Klinikum Unfallkrankenhaus Berlin gGmbH, Berlin, Germany [email protected].
Abstract
This study aims to explore the ability to identify high-grade intracranial arterial stenosis (ICAS) by an artificial intelligence (AI) designed to detect large vessel occlusions (LVO) and the clinical relevance of these 'false positive' findings. We are presenting a retrospective cohort study. The study was conducted at a supraregional stroke centre of an urban tertiary care provider. Consecutive stroke cases treated between January 2023 and December 2023 of patients >18 years of both sexes and any ethnicity were eligible for inclusion. 934 patients (52.7% male) with a mean age of 71.7±13.6 years (25-101 years) were included. CT angiographies were analysed by a deep learning algorithm for LVO detection of the anterior circulation. AI results were compared with radiology reports and secondary focused evaluation. Diagnostic accuracies for ICAS detection by the AI were calculated. Primary reports identified 30 ICAS and nine additional ICAS were detected during secondary evaluation (incidence 4.2%). The sensitivity of radiology reports was 77% (95% CI 0.61 to 0.89), the specificity 99% (95% CI 0.98 to 1.00), negative predictive value (NPV) 99% (95% CI 0.98 to 0.99) and positive predictive value (PPV) 79% (95% CI 0.65 to 0.88). The AI identified 13 of 39 ICAS correctly. 18 false positive cases (neither LVO nor ICAS) were flagged by the AI. The sensitivity of the algorithm was 33% (95% CI 0.19 to 0.50), the specificity 98% (95% CI 0.97 to 0.99), the NPV 97% (95% CI 0.96 to 0.98) and PPV 42% (95% CI 0.28 to 0.58). Detection of high-grade ICAS by an algorithm trained to identify LVO is per se a false positive finding but occurred in 13 of 39 cases. Dedicated training for ICAS might lead to a beneficial tool during the diagnostic work-up for ischaemic stroke. German Register for Clinical Trials (DRKS: DRKS00034019 https://drks.de/search/de/trial/DRKS00034019).