Background: Prior studies show AI applied to chest X-rays (CXRs) can estimate cardiovascular risk. With >70 million CXRs performed annually in the United States, routine CXRs offer scalable opportunity for opportunistic cardiovascular disease (CVD) risk flagging. Because formal risk scores often require clinical and laboratory inputs unavailable in imaging centers, emergency departments, or hospital encounters, AI may identify patients warranting formal risk assessment, including selective CAC scoring.

Objectives: To externally validate an open-source AI model, here termed AICVD-CXR, originally trained to predict cardiovascular mortality in hospital-based patients without recorded CVD.

Methods: We analyzed 14,830 MIMIC patients without baseline CVD using the first posterior-anterior CXR. Outcomes were ASCVD, HF, AF, CVD-associated mortality, and all-cause mortality. AICVD-CXR was compared with PREVENT-ASCVD, PREVENT-HF, and CHARGE-AF in complete, range-valid subsets.

Results: Over median 3.9 years, ASCVD occurred in 831 patients (5.6%), HF in 1,088 (7.3%), AF in 971 (6.5%), CVD-associated mortality in 704 (4.8%), and all-cause mortality in 1,983 (13.4%). AICVD-CXR discriminated ASCVD (AUC 0.749), HF (0.784), AF (0.797), CVD-associated mortality (0.819), and all-cause mortality (0.767). The highest AICVD-CXR quartile had higher hazards for ASCVD (HR 2.77, 95% CI 2.44-3.15) and HF (HR 4.17, 95% CI 3.72-4.68). PREVENT-ASCVD outperformed AICVD-CXR for ASCVD, whereas AICVD-CXR was comparable to PREVENT-HF and CHARGE-AF. Among patients with PREVENT-ASCVD <5%, AICVD-CXR Q4 had more than double the ASCVD event rate.

Conclusions: AI may serve as an opportunistic CXR-based trigger for formal CVD risk assessment and selective CAC evaluation when clinical risk inputs are missing or conventional risk appears low.