Using AI to estimate cardiac risk from a chest X-ray
April 17, 2024
Comparison of AI CXR Algorithm and Traditional AlgorithmSource: Weiss, et al Ann Int Med March 24, 2024 LINK Left is the x-ray prediction, compared to the right which is using a traditional algorithm. The lines represent the portion of each population that had a cardiovascular event at each time period. The CXRs were already done - no patients were imaged specifically for this trial.
When we interact on the web, we leave a vast trail of digital exhaust. We might think that we are simply ordering a latte, purchasing a book, or getting directions. But our browsing history, where we have traveled, and even how fast we type and whether we linger over certain images or words is “digital exhaust,” which allows marketers to glean important information.
There is a clinical equivalent to digital exhaust. An EKG might give a hint about whether our electrolytes are in balance. A CT or MRI often shows information about organs that hadn’t worried us. These unexpected findings sometimes lead to a fruitless set of unnecessary tests, but at other times this digital exhaust can help us improve patient health.
For instance, each year, over 70 million chest X-rays are performed. Some people have a chest radiograph (CXR) to evaluate respiratory or cardiac function or search for an infection. Others have CXRs as part of preoperative evaluation, which is usually not medically necessary. Radiologists are excellent at reading chest X-rays, but artificial intelligence can interpret the equivalent of digital exhaust at an entirely new level.
For instance, last summer researchers reported in Nature Communications that they used machine learning to create an algorithm that uses the digital exhaust of a chest X-ray and access electronic medical records to predict a diagnosis of diabetes up to three years before a patient has clinical symptoms. The AI noted some fat deposits on the radiographs that would not likely have been commented on by a human radiologist.
Annals of Internal Medicine reported last month that researchers using machine learning developed rules that allowed AI to estimate risk of a major adverse cardiovascular event based on a single CXR. The AI was used on historical images with available 10-year follow-up, and was about as good as a risk score that required clinical elements. The authors note that 4 in 5 patients were missing at least one of the clinical elements needed for the ASCVD (atherosclerotic cardiovascular disease) risk estimator - and many of those people had a CXR. Using CXRs to better understand patient risk could improve care without raising costs.
Implications for employers:
Use of machine learning to establish rules that help better identify risk and manage patients is likely to continue.
Machine learning and artificial intelligence is trained on data from past clinical care, which unfortunately incorporates substantial racism. Therefore, clinicians should be careful to identify and eliminate biases in these rules.
It’s still a long way off before a disembodied algorithm within artificial intelligence will be able to replace highly skilled clinicians.
Thanks for reading. You can find previous posts in the Employer Coverage archive. Please subscribe, “like” and suggest this newsletter to friends and colleagues. Thanks!
Illustration by Dall-E
Next post on Friday.