Invasive coronary angiography (CAG) remains the gold standard for diagnosing coronary artery disease, yet its interpretation varies widely between observers. Existing AI tools can improve consistency, but they typically output only a final prediction, leaving the reasoning process opaque and hard for clinicians to verify.
Researchers introduced CARDEA, a unified large vision-language model designed as the inference core of an end-to-end CAG pipeline. It runs from raw multi-view videos through keyframe selection to study-level diagnosis, and it exposes auditable spatial evidence behind its conclusions using a Chain-of-Box (CoB) approach that indicates the image regions the model attends to.
The model was trained solely on public datasets and closed-ended tasks in three stages: visual feature alignment, a self-distilled CoB cold start, and reinforcement learning with verifiable rewards (RLVR) that included a CoB reward encouraging bounding-box use in the reasoning trace. Report generation was excluded from training and evaluated zero-shot on an external cohort.
CARDEA trailed a dedicated classifier on in-distribution dominance classification but drew level under domain shift, reaching an accuracy of 0.91 (95% CI, 0.86 to 0.95). On complexity assessment it was comparable to two interventional cardiologists, with an accuracy of 0.90 (95% CI, 0.82 to 0.97).
Only the RLVR stage improved zero-shot report generation, raising vessel-severity macro-F_1 to 0.686 (95% CI, 0.664 to 0.707) from 0.513 for the untuned base model, and more than double the always-normal floor of 0.312. The authors note that RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not, while cautioning that clinical use requires prospective validation against expert cardiologists. Model weights and inference code are publicly available.