Artificial intelligence in coronary angiography: benchmarking the diagnostic accuracy of ChatGPT-4o against interventional cardiologists.

Abstract

Background: The integration of artificial intelligence (AI) into medical diagnostics has significantly impacted cardiology by enhancing diagnostic precision and therapeutic strategies. Coronary artery disease continues to be a leading cause of global morbidity and mortality, with coronary angiography being the diagnostic gold standard. However, the subjective nature of angiographic interpretation can lead to inconsistent assessment. AI aims to provide automated, objective assessments to mitigate these challenges.

Methods: This study evaluated ChatGPT with Generative Pre-trained Transformer (GPT)-4o (OpenAI, USA), for automated coronary angiogram interpretation. Due to its inability to process video data, we extracted maximum contrast frames from diagnostic angiogram views. These anonymised images were analysed by GPT-4o. Its diagnostic findings and stent recommendations were compared with expert cardiologist assessments.

Results: We included 100 patients who underwent coronary interventions between January and April 2024. GPT-4o accurately identified coronary vessels in 98% of images. The overall sensitivity for detecting lesions requiring intervention was 71.6%, with a specificity of 57.2% (F1 score 0.652). Performance varied by vessel with best results for left anterior descending artery (sensitivity 81.0%; specificity 69.3%) and right coronary artery (sensitivity 86.5%; specificity 61.4%). Identification of the target vessel based solely on imaging was 47%, which improved to 87% with additional clinical information.

Conclusions: GPT-4o shows potential as a supportive tool in coronary angiography interpretation. Its diagnostic performance improves significantly when contextual clinical information is included. However, its accuracy based on static images alone remains below the threshold required for reliable diagnostic and therapeutic support. The lack of cine-loop data as an essential element in real-world angiographic interpretation is a key limitation. Future developments should focus on enhancing AI capabilities for analysing complex anatomical structures and integrating dynamic imaging data to augment clinical utility.