An Incidental Signal on Chest CT: Can AI Detect Cardiovascular Risk Earlier?

Many chest CT scans are originally performed for lung, tumor, or other thoracic issues, with cardiovascular risk not the main focus of the examination. But the images often also record clues such as vascular calcification, cardiac morphology, and changes in surrounding tissue. If this information can be read reliably, existing examinations may be able to take on an additional preventive medicine function: before myocardial infarction or stroke occurs, they could point clinicians toward people who need more active evaluation.

A study published in Nature Medicine on June 17, 2025, reports an AI system that estimates cardiovascular risk from routine chest CT images. According to the study abstract, the team not only built the model but also conducted external validation, with the aim of predicting the future risk of major cardiovascular events such as myocardial infarction and stroke from opportunistic imaging data. This makes it different from AI models that simply demonstrate image-recognition capability, pushing the question toward clinical prognostic interpretation: can a CT scan that has already been taken provide usable signals about a patient’s future risk?

The appeal of this kind of “opportunistic screening” is that it does not require patients to undergo an additional imaging examination solely for risk assessment. Chest CT is frequently used in hospitals. If an algorithm can automatically analyze images within existing workflows, in theory it could help identify high-risk groups that traditional risk factors may not fully capture. However, this also means the model must face the complexity of real-world imaging, including different scanning parameters, patient populations at different medical institutions, and the limitation that the images were not originally designed for cardiovascular assessment.

The study abstract specifically mentions external validation, an important step for this type of medical AI to move from technical demonstration toward clinical credibility. Internal testing often overestimates model performance because training and testing data may share similar equipment, populations, and clinical workflows; external data can more rigorously test the model’s stability in different environments. Even so, the abstract does not provide all the details needed to judge clinical usability, such as the number of events, follow-up time, performance differences across populations, and the magnitude of improvement compared with existing risk scores. Interpretation should therefore retain a sense of scale.

For clinical practice, the real question is not whether AI can calculate a risk score, but how that score should then be used. If the model indicates that someone is at higher risk, should clinicians arrange further tests, adjust lipid-lowering or blood pressure-lowering strategies, or simply place the result in the medical record as auxiliary information? Without prospective studies and clear workflows to support these actions, AI prediction may increase alerts but may not necessarily improve patient outcomes.

Regulation and responsibility will also emerge. Routine chest CT is obtained for specific clinical purposes. If AI additionally analyzes cardiovascular risk in the background, patient consent, notification of results, anxiety caused by false positives, and the burden on medical care all require institutional design. If the model is to be deployed across different hospitals, it must also prove that it can maintain performance under data drift, equipment updates, and population differences, rather than appearing accurate only in research data.

The significance of this study lies in how it moves the discussion of medical AI from “seeing lesions in images” to “predicting events that have not yet occurred.” Chest CT may not therefore become a universal entry point for cardiovascular screening, but it reminds clinical medicine that existing imaging data still contain biological signals that have not been fully used. The key next step will be to prove that these signals can not only be read by algorithms, but also be responsibly translated by medical systems into earlier and more accurate care decisions.