CARDEA angiography AI matches cardiologists on case complexity
TL;DR
- CARDEA reached 0.90 accuracy on coronary angiography complexity assessment (95% CI 0.82–0.97), comparable to two interventional cardiologists in the study.
- Only reinforcement learning with verifiable rewards lifted zero-shot vessel-severity F1 to 0.686, from a 0.513 base and 0.312 always-normal floor.
- The model exposes bounding-box spatial evidence in its reasoning trace, but the authors say clinical deployment still needs prospective validation.
CARDEA, a vision-language model for invasive coronary angiography, reached 0.90 accuracy on case complexity assessment (95% CI 0.82 to 0.97) and was 'comparable to the cardiologists on complexity assessment' it was benchmarked against, according to a preprint on arXiv by Jia-Jen Lee and colleagues. Two interventional cardiologists served as the human comparators.
The novel piece is the training recipe. CARDEA was trained solely on public datasets in three stages: visual feature alignment, 'a self-distilled Chain-of-Box (CoB) cold start,' and reinforcement learning with verifiable rewards using a reward 'encouraging bounding-box use in the reasoning trace.' Report generation was held out of training entirely; only after the RLVR stage did zero-shot vessel-severity macro-F1 rise to 0.686, up from a 0.513 untuned base and a 0.312 always-normal floor.
On in-distribution dominance classification CARDEA trailed a dedicated classifier; it 'drew level under domain shift' at 0.91 accuracy. The abstract publishes no per-vessel severity breakdown and names no hospital partner for follow-on trials. 'Clinical use requires prospective validation against expert cardiologists,' the authors write.
Originally reported by paper
Read the original article →Original headline: CARDEA Coronary-Angiography AI Matches 10-Year Cardiologist Accuracy With Auditable Spatial Reasoning