Expert-AI collaboration enhances diagnostic accuracy and agreement in CTA interpretation of acute aortic syndromes

Abstract

To assess the clinical impact of an artificial intelligence (AI)-based software-as-a-medical-device (SaMD) for acute aortic syndrome (AAS) detection on computed tomography angiography (CTA), and validate its effect on diagnostic accuracy in a multicenter reader study. In this retrospective study, 205 CTA scans (101 AAS-positive) were interpreted twice by nine clinicians across four specialties, without and then with AI assistance, separated by a washout interval of at least four weeks. Comparisons were made both on pooled reader-case observations and at the reader level. AI improved diagnostic accuracy (96.26% vs. 94.69%, p  = 0.0058) and specificity (98.72% vs. 93.80%, p  < 0.001), with a decrease in sensitivity (93.73% vs. 95.60%, p  = 0.0171). At the reader level, specificity improved in all nine readers ( p  = 0.004), while accuracy and sensitivity changes varied in direction. Inter-reader agreement improved from a Fleiss' kappa of 0.849–0.924. Clinicians reported greater confidence and efficiency. These findings support the integration of AI-driven clinical decision support tools into frontline diagnostic workflows for time-critical conditions such as AAS, provided that safeguards against false-negative AI output are in place.