AI-assisted hemorrhage detection following endovascular stroke treatment: a retrospective diagnostic accuracy study

Antithrombotic therapy is essential for preventing strokes, but its use after reperfusion therapy requires careful monitoring due to the risk of hemorrhagic transformation. Non-contrast-enhanced computed tomography (NCCT) is the standard for detecting intracranial hemorrhages post-stroke. Artificial intelligence may enhance hemorrhage detection and improve patient safety. This study evaluates AI’s sensitivity and specificity in detecting hemorrhagic events in NCCT scans within 48 h after endovascular stroke treatment, compared to standard radiological assessment. A retrospective, single-center study was conducted at a European stroke center, including 495 NCCT scans from 425 patients who underwent endovascular stroke treatment between 08/2021 and 06/2024. A CE-marked AI software based on convolutional neural networks (CNN) analyzed the scans independently. The reference standard was assessments of two board-certified neuroradiologists, and AI results were compared with routine radiological reports. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated, and inter-rater reliability was assessed using Cohen’s kappa. The reference standard identified hemorrhages in 197 NCCT scans. The AI system showed sensitivity of 95.9%, specificity of 84.6%, PPV of 80.4%, and NPV of 96.9%. Radiological reports had sensitivity of 91.9%, specificity of 96.3%, PPV of 94.3%, and NPV of 94.7%. Cohen’s kappa was higher for radiological reports (0.886) than AI (0.780), indicating stronger agreement with the reference standard. AI had a higher false-positive rate (15.4%) than radiological reports (3.7%). AI demonstrated high sensitivity for detecting intracranial hemorrhages but had a higher false-positive rate compared to routine radiological assessment. While AI can aid clinical decision-making, radiologists show superior overall diagnostic accuracy. Further research is needed to explore the impact of AI-assisted decision-making on stroke management and secondary prevention.