Diagnostic performance of artificial intelligence-based coronary artery analysis in the emergency department: A multicenter study.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, Pusan National University Hospital, Pusan National University School of Medicine and Medical Research Institute, Busan, Republic of Korea.
- Department of Radiology, Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea.
- Department of Radiology, Dongsan Medical Center, Keimyung University School of Medicine, Daegu, Republic of Korea; Division of Cardiothoracic Imaging, Department of Radiology, Emory University, Atlanta, GA, USA.
- Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul and Dankook University Hospital, Cheonan, Republic of Korea.
- Siemens Healthineers AG, Forchheim, Germany.
- Siemens Healthineers, Seoul, Republic of Korea.
- Department of Radiology, Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea. Electronic address: [email protected].
Abstract
On-premise artificial intelligence (AI) software may enable rapid interpretation of coronary CT angiography (CCTA) in the emergency department (ED), although its diagnostic performance relative to invasive coronary angiography (ICA) remains uncertain. We aimed to evaluate the diagnostic performance of an AI-based coronary artery analysis software for detecting obstructive coronary artery disease (CAD) (CAD-RADS ≥3) in ED patients with acute chest pain, using expert consensus and ICA as separate reference standards. In this retrospective multicenter study, 1193 patients with acute chest pain who underwent CCTA across four academic EDs (January 2019-August 2024) were analyzed using on-premise AI software. Diagnostic performance for detecting obstructive CAD was evaluated in two cohorts, using expert consensus (n = 885) and ICA (n = 308) as reference standards. Overall, 342 patients (28.7%) had obstructive CAD. Using expert consensus, AI achieved a sensitivity of 79.5% and a negative predictive value (NPV) of 95.4%. With ICA as the reference, sensitivity and NPV were 94.6% and 89.2%, respectively. AI and expert radiologists showed comparable sensitivity (94.6% vs 95.2%; P = 0.56), NPV (89.2% vs 91.4%; P = 0.20), and area under the curve (0.87 vs 0.87; P = 0.86), although AI demonstrated lower specificity (68.0% vs 78.7%) and positive predictive value (PPV) (81.8% vs 87.2%) (both P < 0.001). On-premise AI provides high sensitivity and NPV for excluding obstructive CAD, comparable to expert radiologists, although specificity and PPV were significantly lower. These findings support its role as a second-reader and clinical decision-support tool in emergency settings.