Opportunistic Screening of Carotid Atherosclerosis and Cardiovascular Mortality Risk Using Chest Radiographs: A Comparative Study of Foundation Models.
Rehman A, Kim J, et al. • Journal of the American Heart Association • 2026
Foundation models enable opportunistic detection of carotid atherosclerosis from chest radiographs and yield a biomarker that provides prognostic information for CVD death, with explainability analyses demonstrating alignment with cardiovascular anatomic regions.
Key Findings
Results
Rad-DINO achieved the highest performance among foundation models for detecting carotid atherosclerosis from chest radiographs.
Rad-DINO was compared against DINOv2, OpenCLIP, and CheXagent across three adaptation strategies: linear probing, selective fine-tuning, and low-rank adaptation (LoRA).
Selective fine-tuning yielded an AUROC of 0.74 (95% CI, 0.71–0.77).
Low-rank adaptation also yielded an AUROC of 0.74 (95% CI, 0.70–0.76).
The evaluation was conducted among 5,785 participants with carotid sonography annotations.
Results
Rad-DINO with low-rank adaptation demonstrated the highest anatomic relevance among all tested model configurations.
Explainability was assessed using gradient-weighted class activation mapping (Grad-CAM) and a clinical relevance index (CRI).
The CRI quantified model attention within cardiovascular anatomic masks.
Rad-DINO under low-rank adaptation achieved a CRI of 0.31, the highest among all configurations tested.
Higher CRI values indicate greater alignment of model attention with clinically relevant cardiovascular anatomic regions.
Results
Higher Deep-Learning Chest X-Ray Atherosclerosis Score (DL-CXR-AS) tertiles were strongly associated with cardiovascular disease death after adjustment for Framingham Risk Score.
The association was tested in 32,524 participants using Framingham Risk Score-adjusted Cox proportional hazards models.
The adjusted hazard ratio for CVD death in the high versus low DL-CXR-AS tertile was 9.36 (95% CI, 3.35–26.15).
The trend across tertiles was statistically significant (P-trend < 0.005).
The model concordance index for CVD death prediction was 0.73 (95% CI, 0.67–0.75).
Methods
The study used a multi-model comparative framework evaluating three distinct adaptation strategies for applying foundation models to chest radiograph-based atherosclerosis screening.
The three adaptation strategies evaluated were linear probing, selective fine-tuning, and low-rank adaptation (LoRA).
Four foundation models were compared: Rad-DINO, DINOv2, OpenCLIP, and CheXagent.
Performance was assessed by area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals.
Carotid sonography annotations served as the ground truth for atherosclerosis classification in the 5,785-participant detection cohort.
Results
The DL-CXR-AS biomarker derived from the best-performing foundation model provided prognostic information for cardiovascular mortality beyond traditional risk factors.
Cox models were adjusted for the Framingham Risk Score, a standard multivariable cardiovascular risk tool.
Despite adjustment, the DL-CXR-AS tertiles retained strong independent association with CVD death.
The large sample of 32,524 participants was used for the prognostic analysis, substantially larger than the detection cohort.
The concordance index of 0.73 indicates moderate-to-good discriminative ability for CVD death.
What This Means
This research suggests that artificial intelligence models originally trained on large collections of medical images can be repurposed to detect signs of carotid artery disease—a key marker of atherosclerosis, or hardening of the arteries—directly from standard chest X-rays. Chest X-rays are among the most commonly performed medical imaging tests worldwide, but they are not typically used to assess artery health. By applying several state-of-the-art AI foundation models to chest X-rays from nearly 6,000 people who also had specialized carotid artery ultrasound exams, the researchers found that one model (Rad-DINO) could detect atherosclerosis with moderate accuracy. Importantly, the AI also appeared to focus its analysis on the correct anatomical regions of the chest associated with cardiovascular structures, lending credibility to its decision-making.
The study then went further, generating an AI-derived score from chest X-rays and testing whether that score could predict who would die from cardiovascular disease in a much larger group of over 32,000 people. Those with the highest AI scores had more than nine times the risk of cardiovascular death compared to those with the lowest scores, even after accounting for traditional risk factors captured by the established Framingham Risk Score. This suggests the AI-derived score provides additional prognostic information beyond what standard clinical tools already offer.
This research matters because it points toward a future where routine chest X-rays—already obtained for many reasons such as pre-operative assessments or respiratory complaints—could be simultaneously analyzed by AI to flag individuals at elevated cardiovascular risk, without any additional tests or cost. This kind of 'opportunistic screening' could help identify high-risk individuals earlier, particularly in settings where specialized cardiovascular imaging is not readily available. Further validation in diverse populations and prospective clinical settings would be needed before such approaches could be considered for clinical implementation.
Rehman A, Kim J, Lee H, Chang J, Park S. (2026). Opportunistic Screening of Carotid Atherosclerosis and Cardiovascular Mortality Risk Using Chest Radiographs: A Comparative Study of Foundation Models.. Journal of the American Heart Association. https://doi.org/10.1161/JAHA.125.048485