Chest X-ray artificial intelligence for cardiovascular risk flagging in younger patients: An AI-CVD study within the Medical Information Mart for Intensive Care (MIMIC)
Year: 2026
Type: article
Abstract: Aims Chest X-rays (CXRs) may provide opportunistic cardiovascular risk information. We implemented the deep learning CXR CVD-Risk model, developed to predict cardiovascular mortality, in MIMIC adults without recorded baseline cardiovascular disease (CVD), focusing on patients aged <60 years. Methods and results Among 10,132 patients (mean age 40.9 ± 12.0 years; 39.3% male), 756 (7.5%) developed Total CVD (atherosclerotic cardiovascular disease [ASCVD], heart failure [HF], or atrial fibrillation [AF]) over a median 3.8 years. CXR CVD-Risk had higher AUPRC than age for Total CVD (0.170 vs 0.132) and HF (0.119 vs 0.064). At the prespecified top-decile threshold, positive predictive value (PPV) was 20.2% versus 15.5% for age for Total CVD (p<0.001) and 12.6% versus 7.4% for HF (p<0.001). After age/sex adjustment, Q4 remained associated with Total CVD (HR 1.43, 95% CI 1.12–1.84) and HF (HR 2.57, 95% CI 2.04–3.23). Among patients with PREVENT-HF <5%, the CXR CVD-Risk top decile had a 9.3% HF event rate versus 2.8% in the bottom 90% (RR 3.35, 95% CI 2.46–4.58), capturing 27.2% of HF events while flagging 10%. In complete, range-valid subsets, HF PPV was 10.7% for CXR CVD-Risk versus 12.1% for PREVENT-HF (p=0.35). Conclusions In younger patients, CXR CVD-Risk modestly outperformed age in discriminating a composite cardiovascular outcome and identified a subgroup with higher PPV than age-based selection — most clearly for heart failure. The model should be interpreted as an opportunistic enrichment signal to prompt further cardiovascular risk assessment, not as a calibrated absolute-risk estimate or a substitute for validated clinical risk equations.
Authors: Hamed Zarei, Anthony P. Reeves, Mohammad Mozafary, Seyed Reza Mirjalili, Kyle Atlas, Chenyu Zhang, Amir Azimi, Amir Ghaffari Jolfayi, Ali Hashemi, Thomas Atlas, Jakob Wasserthal, Claudia I. Henschke, David F. Yankelevitz, Andrea D. Branch, Sion K. Roy, Jamal S. Rana, Zahi A. Fayad, Oren Mechanic, Jagat Narula, Roxana Mehran, Prediman K. Shah, John A. Rumberger, Kim A. Williams Sr, Paolo Raggi, David J. Maron, Michael V. McConnell, Robert A. Kloner, Matthew J. Budoff, Nathan D. Wong, Arthur S. Agatston, Koen Nieman, Robert Headrick, Rozemarijn Vliegenthart, Morteza Naghavi
Institutions: Cornell University, Icahn School of Medicine at Mount Sinai, The Lundquist Institute, Kaiser Permanente Oakland Medical Center, University of Houston, Cedars-Sinai Medical Center, Princeton University, University of Alberta, Stanford Medicine, Stanford University, University of Southern California, Huntington Medical Research Institutes, University of California, Irvine, University Medical Center Groningen
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $1,900
OpenAlex ID: W7214095346