Twists & Turns Ahead on AI’s Road to Acceptance in Cardiology

If media coverage and buzz are any indication, artificial intelligence (AI) is poised to assume a position of importance in healthcare—and cardiology is no exception. Some cardiologists perceive it as a valuable tool for helping to improve disease prevention and detection, interpret test results, personalize therapies, monitor for secondary issues and more. However, AI as a whole has, as one source puts it, “far to go” before it will be embraced by the cardiology community.

Recent clinical studies suggest that AI could indeed offer value, efficacy and safety. One study examined the application of AI (specifically, deep learning) in resting coronary CT angiography (CCTA) of the left ventricular myocardium. The researchers found that using deep learning  analysis to help evaluate the degree of coronary artery stenosis improved diagnostic performance and increased the specificity of resting CCTA, which could “potentially decrease the number of patients undergoing invasive coronary angiography” (Eur Radiol, online Nov. 12, 2018). 

Another study revealed that AI in the form of a signal-processed surface electrocardiographic algorithm can predict abnormal myocardial relaxation, allowing for earlier detection of cardiac disease than would otherwise be possible (J Am Coll Cardiol 2018;71[15]:1650-60).

And a third study found that using machine learning of data from more than 40,000 patients in the Swedish Heart Registry to assess the appropriateness of defibrillator therapy for individuals with heart failure yielded markedly better risk stratification and prognostication than ejection fraction (J Am Heart Assoc 2018;7[8]:e008081). 

Powerful proof needed 

While not insignificant, studies like these will not suffice for convincing cardiology leaders or in-the-trenches practitioners to endorse AI—either wholeheartedly or in part. Before taking the plunge, sources tell CVB, cardiologists will require proof of AI’s power to enhance diagnostic accuracy and, in turn, improve patient outcomes as well as decrease morbidity, mortality and the cost of care. 

Such proof will need to come largely from repeatable clinical trials, with results validated by independent clinical experts, says Steven Steinhubl, MD, director of digital medicine at Scripps Research Translational Institute and a cardiologist at the Scripps Clinic in La Jolla, Calif. He believes that for a majority of cardiologists the most credible clinical trial results will be gleaned from studies involving the implementation of AI in real-world scenarios.

“One challenge we’ve been seeing is that these studies are mainly being implemented in ‘artificial’ systems or under contrived circumstances, where physicians and patients act differently than they typically would,” he explains. “They are of limited value and not entirely convincing.” Close observation and reporting of real-world applications will be needed to get the cardiology community on board.

Jagmeet Singh, MD, PhD, associate chief of cardiology and professor of medicine at Harvard Medical School in Boston, agrees. He adds that clinical trials (and perhaps other assessments) of AI must be as rigorous as for any type of therapy as well as involve multiple covariants, rather than just one or two. Without these testing parameters, neither AI’s full value nor the safety of applying it in any of its guises (e.g., machine learning, deep learning) can be guaranteed, spurring hesitancy among cardiologists to adopt AI.

The potential of AI solutions to curry favor with cardiologists is equally predicated on whether algorithms integrate a variety of data. “For example,” Singh says, “when it comes to the use of AI to enhance the delivery of therapy, the learning algorithms will have to integrate demographics and clinical characteristics along with structural, procedural and outcome data.”

Cardiologists’ inclination to add AI solutions to their technology toolbox might increase if the solutions provide reliable, reproducible answers to their questions and bring to light insights that would otherwise be difficult or impossible to glean, says Joseph Hill, MD, editor-in-chief of Circulation and chief of cardiology at UT Southwestern in Dallas. One example of the latter is insight into complex, system-level responses to heart disease in different patients, he says.

Cardiologists also will be more apt to adopt and use AI solutions if developers revise their solutions over time. “Evidence changes and, with it, guidelines,” notes Andrew Freeman, MD, director of clinical cardiology and cardiovascular prevention and wellness at National Jewish Health in Denver and a member of the American College of Cardiology’s Health Care Innovation Section. “These changes must be reflected in adjustments to AI mechanisms,” he says. Otherwise, AI’s credibility with—and value for—the cardiology community will be compromised.

Scaling barriers

While exhaustive clinical trials and other elements may bode well for AI in cardiology, potential barriers to general acceptance still will need to be scaled. Some cardiologists question whether AI will be worth leveraging when, in certain cases, its benefits do not seem to outweigh its drawbacks.

James Kirkpatrick, MD, director of the echocardiography laboratory at UW Medical Center and adjunct associate professor of bioethics and humanities at the University of Washington, cites as an example AI in echocardiography. It is quite  possible, he acknowledges, that the machine learning component of AI will soon be able to generate reports containing measurements and interpretations and that robot-assisted scanning will someday allow machines to capture all echocardiographic views from standard imaging windows without human direction. 

“Even if AI can replace echocardiographers and sonographers, though, we need to look at the ethical issue of whether it should,” Kirkpatrick says. “Will the cost savings from increased efficiencies be worth the loss of the human element in echocardiographic interpretation and scanning? And will machines be able to put findings into full clinical context in discussion with referring clinicians? These are tough questions to answer. And some way to maintain patient relationships while taking advantage of AI must be found.”

That AI solutions require significant time to learn and be trained comprises yet another obstacle to acceptance. And, unlike human intelligence, deep learning is not capable of integrating prior knowledge like humans can. It also has difficulty identifying rare anomalies. 

“This may change in time as larger data sets and new ways to train systems to function more like the human brain are developed,” notes Partho Sengupta, MD, chief of cardiology and chair of cardiac innovation at the West Virginia University Heart and Vascular Institute in Morgantown.  Until some of these limitations are addressed, he proposes, the best use of AI is to relieve cardiologists of repetitive, low-level tasks such as measurements, data preparation, standardization and quality control. AI would then free up cardiologists to focus on more complex interpretations, patient care and clinical decision making.

Finally, some sources say, there will be cardiologists who hesitate to accept AI if they are not privy to what lies behind the technology’s conclusions. Not everyone, they contend, will be swayed by proof of AI’s efficacy.

“It’s a bit of a quandary,” Singh concludes. “But no matter what, with evidence of efficacy and advancements in the technology, to name a few, cardiology will buy in.”