AI Picks Up on Info Operators Miss on Angiograms: AI-ENCODE

LONG BEACH, CA—Artificial intelligence (AI) may have a role to play in detecting noncoronary data on angiograms not spotted by operators that could provide clues to modify or enhance treatment plans, the AI-ENCODE study suggests.

“It’s easy to miss what we’re not looking for,” said Mohamad Alkhouli, MD (Mayo Clinic School of Medicine, Rochester, MN), presenting the study in a featured clinical research session here at SCAI 2024.

AI has been a hotbed of research in the cardiology space, with machine-based learning being directed at sniffing out pertinent clinical information that might otherwise be overlooked. Recent examples include using it to identify low LVEF, LV systolic dysfunction, long QT syndrome, and coronary calcium.

For the AI model, Alkhouli and colleagues curated more than 20,000 angiograms from coronary cases done at their center between 2015 and 2022. The model learned to look beyond just detecting blockages and to extract data on left ventricular ejection fraction, diastolic dysfunction, right ventricular dysfunction, and cardiac index from just one or two angiographic videos. Alkhouli said the researchers chose those four elements for the model to mine because the they thought they would be helpful to an operator in the setting of an acute cardiac case.

It’s easy to miss what we’re not looking for. Mohamad Alkhouli

For prediction of an ejection fraction of ≤ 40%, as well as for prediction of high filling pressures, the AI model yielded an area under the curve (AUC) of 0.87 when compared with echocardiography performed within 30 days of the angiogram. For right ventricular function the AUC was 0.78 when compared with echocardiography, while for cardiac index it was 0.74 when compared with simultaneous right heart catheterizations.

Although further validation is needed, Alkhouli said his center has now incorporated the model into a local cloud in one of the cath labs, where it can serve as a kind of dashboard to “spit the data back to the operators in the room, so they can actually use it in an actionable fashion.”

AI as an Assistant

Panelist Yiannis Chatzizisis, MD, PhD (University of Miami Miller School of Medicine, FL), said the results point toward another way in which AI potentially can act as an assistant for busy operators.

He suggested that the model essentially represents an effort to “democratize knowledge” so that everyone in a given center has access to the information gleaned from cases done by colleagues, equating it to squeezing a lemon to get every last drop of juice out of it.

“With so much ad hoc procedures being done nowadays, this immediate feedback would be invaluable,” added panelist J. Dawn Abbott, MD (Brown University, Providence, RI). But the real test, she said, is whether using the AI translates to improved patient outcomes.

Alkhouli agreed validation will be an important next step in this research, and said that a pragmatic trial is planned to understand how using the AI drives resource utilization.

He gave the example of a STEMI patient with no prior echocardiogram in whom an intervention is not going well. “The model is telling you, well, this patient may be in cardiogenic shock, this patient has elevated filling pressure. So, it would it either increase the resource utilization by you having to verify that information and treat the patient accordingly, or it would save you from doing additional procedures,” Alkhouli told TCTMD.