Abstract
Background
Cardiovascular disease (CVD) remains a leading cause of death, but population‐level screening for atherosclerosis often depends on specialized imaging that is not routinely available. We evaluated foundation models for opportunistic detection of carotid atherosclerosis from chest radiographs, assessed the prognostic value of the resulting biomarker for CVD death, and examined model explainability.
Methods
Among 5785 participants with carotid sonography annotations, we compared Rad‐DINO, DINOv2, OpenCLIP, and CheXagent for chest radiograph–based atherosclerosis detection using linear probing, selective fine‐tuning, and low‐rank adaptation. Performance was assessed by area under the receiver operating characteristic curve with 95% CIs. Explainability was evaluated using gradient‐weighted class activation mapping and a clinical relevance index, which quantified attention within cardiovascular anatomic masks. The selected model generated a Deep‐Learning Chest X‐Ray Atherosclerosis Score, which was tested for association with CVD death in 32 524 participants using Framingham Risk Score–adjusted Cox models.
Results
Rad‐DINO achieved the highest performance with selective fine‐tuning (area under the receiver operating characteristic curve, 0.74 [95% CI, 0.71–0.77]) and low‐rank adaptation (area under the receiver operating characteristic curve, 0.74 [95% CI, 0.70–0.76]), with the highest anatomic relevance under low‐rank adaptation (clinical relevance index, 0.31). Higher Deep‐Learning Chest X‐Ray Atherosclerosis Score tertiles were strongly associated with CVD death (adjusted hazard ratio, 9.36 [95% CI, 3.35–26.15] for high versus low;
P
‐trend
Conclusions
Foundation models enable opportunistic detection of carotid atherosclerosis from chest radiographs and yield a biomarker that provides prognostic information for CVD death. Explainability analyses demonstrated alignment with CVD anatomic regions, supporting clinical relevance and scalable risk stratification.