Algorithmic Potential: Leveraging Deep Learning to Improve Arrhythmia Identification & Classification
To many cardiologists, routine use of artificial intelligence (AI) in practice may seem as far off as human travel to Mars—within the realm of possibility but no more around the corner than the red planet is to earth. Such reservations didn’t stop cardiologists from attending the Heart Rhythm Society’s 2019 annual meeting session on computational approaches to incorporating AI in electrophysiology. Meanwhile, a few researchers are testing real-world applications, and AI’s proponents are intrigued by the potential of deep neural networks for arrhythmia classification and possible payoffs in improved patient care and reduced costs, among other benefits.
Going deep
Deep neural networks are currently the best prospect for identifying and classifying arrythmias, sources told CVB. Unlike some types of AI, such as machine learning, deep neural networks employ sophisticated modeling to process data in complicated ways. That ability to dig deeply into data could turn out to be key to leveraging AI for arrhythmia diagnosis and management.
At HRS.19, Paul Friedman, MD, professor of medicine and chair of cardiovascular medicine at the Mayo Clinic in Rochester, Minn., described essential differences between machine learning and deep learning. While machine learning uses algorithms to parse and learn from data and then draws conclusions based on what it has learned, deep learning structures algorithms in multiple layers, producing a deep neural network that has the capability to learn from the data and formulate intelligent decisions on its own, Friedman explained.
Mayo Clinic, Rochester, Minn.
“Deep learning is preferable to machine learning because machine learning algorithms don’t process data in the complex manner [that deep learning] does,” Friedman subsequently told CVB. The more complex the processing of the data, the lesser the likelihood of “missing something,” which in turn could compromise the arrhythmia diagnosis and classification process.
Better than cardiologists?
Researchers are making progress in demonstrating the viability of using deep learning to classify arrhythmias. Early research focused on how well deep neural networks could detect a few heartbeat types (e.g., normal, ventricular, supraventricular, ectopic, fusion) and how the technology could make a small number of rhythm diagnoses (most commonly atrial fibrillation or ventricular tachycardia). Newer studies are taking on a wider variety of rhythm classes as well as more questions pertinent to real-world use, says Mintu Turakhia, MD, MAS, associate professor of cardiovascular medicine at Stanford University School of Medicine and executive director of Stanford’s Center for Digital Health.
For example, Turakhia was part of a Stanford University team that examined the viability and efficaciousness of a deep learning approach to pinpointing a comprehensive list of cardiac rhythm classes and compared its arrhythmia identification and classification to that of cardiologists. The researchers developed a 34-layer deep neural network to detect 12 rhythm classes from raw single-lead ECG inputs. In addition to sinus rhythm and noise, the classes included atrial fibrillation/flutter, first-degree atrioventricular block, bigeminy, ectopic atrial rhythm, idioventricular rhythm, junctional rhythm, supraventricular tachycardia, trigeminy, ventricular tachycardia and Wenckebach atrioventricular block (Nat Med 2019;25[1]:65-9).
The sensitivity of the deep neural network was found to exceed that of the cardiologists for all of the rhythm classes. “Based on the study, we are at a point now where we know we can use an end-to-end deep learning approach to classify not just a few arrhythmias, but a wide range of them—with a caliber of diagnostic performance very similar to what one would get from cardiologists,” Turakhia says.
At the Mayo Clinic, investigators are undertaking randomized tests of AI in its cardiac care department, according to Friedman. The goal, he said at HRS.19, is to further assess the impact of AI on arrhythmia classification and patient care in a real-world clinical setting, thus going beyond controlled study parameters.
Better, cheaper, faster?
While AI believers acknowledge it is early days for AI in arrhythmia classification and much more study is needed, they expect that, ultimately, deep learning algorithms will enhance patient care in several ways. One improvement will be fewer misdiagnosed computerized ECG interpretations, predicts Turakhia, who in addition to his Stanford roles is the director of cardiac electrophysiology at the Palo Alto Veterans Affairs Health Care System. Another advance will follow from AI’s potential to accelerate ECG interpretation, yielding more efficient processes and more accurate triaging of patients and prioritization of the most urgent cases.
By virtue of its ability to process large volumes of data and spot associations that the human brain can’t, AI could up the diagnostic ante, according to Emma Svennberg, MD, PhD, of the Karolinska Insitutet at Danderyd University Hospital in Stockholm. “We need help diagnosing,” she said at HRS.19. “Diagnostics is our biggest challenge.”
Friedman agrees. “Even the most skilled cardiac physicians do make errors in interpretation,” he told CVB. “It’s inevitable when you think about the voluminous amount of ECG data that has to be interpreted. AI and, again, in particular deep neural networks,” decrease the potential for mistakes.
Sanjiv M. Narayan, MD, PhD, professor of cardiovascular medicine at Stanford University School of Medicine and co-director of the Stanford Arrhythmia Center, says AI also could improve patient care by making atrial fibrillation diagnoses possible from data collected to wearable devices. In one study, a deep learning network was trained on 9,750 ambulatory smartwatch ECGs and then used to analyze data from 12-lead ECGs. The neural network performed well in 51 recumbent patients before cardioversion (c-statistic 0.97 vs. 0.91 for current ECG algorithms) and less well in a cohort group with ambulatory ECGs (c-statistic 0.72; sensitivity 67.7 percent; and specificity 67.6 percent) (JAMA Cardiol 2018:3:409-16).
In another study, researchers used a deep neural network to diagnose atrial fibrillation from electrical ECG sensors in a smartphone case or watch strap that leveraged Bluetooth to a smartphone. In 100 patients with 169 simultaneous wearable and traditional ECGs, 57 recordings could not be interpreted. However, for interpretable ECGs, the device diagnosed atrial fibrillation with a K coefficient of 0.77, sensitivity of 93 percent and specificity of 84 percent (J Am Coll Cardiol 2018;71:2381-8).
“These are not the best results yet, so there’s still a caveat we have to keep in mind when it comes to the wearables aspect of AI,” notes Narayan, who believes better sensors and “additional advancements in analytic algorithms” are on the way, “although probably not for several years.” He foresees a future where screening and monitoring for cardiac arrhythmias won’t require a trip to the hospital or other facility. (See related story: The blame game: Who takes responsibility for AI’s mistakes?)
Cost-cutting is another foreseeable benefit of leveraging AI as a tool for detecting and classifying arrhythmias, Turakhia and Narayan say. With AI in their armamentarium, electrophysiologists will spend less time duplicating efforts to analyze ECG data due to misgivings about the accuracy of their classifications and conclusions, they predict. With fewer inaccurate diagnoses to remedy, costs would drop.
“It’s a case of doubling back and looking at data again, and working in a trial-and-error mode, vs. a learned paradigm,” Turakhia says. “That’s very significant.”
Getting real
There are obstacles to remove before AI could become part of standard practice for identifying arrhythmias and before real benefits will materialize. Doubters are quick to ask about workflow and training, for example.
Workflow impediments—real or imagined—rank among them. Some electrophysiologists worry that AI will add another layer of complexity to interpreting ECGs and classifying arrhythmias. AI enthusiasts tend to dismiss this issue while conceding that skeptics will need to see more studies to be convinced that AI, as Friedman says, “does the work” by speeding up the arrhythmia classification process.
If it turns out AI does cause work-flow barriers, such roadblocks will disappear once deep learning capabilities are built into ECG technology, Friedman predicts. When this happens, AI analysis will be executed for electrophysiologists’ review as data are collected. Already he and his Mayo Clinic colleagues are seeking ways to “bake” AI into ECG data collection.
Neither should training be a stumbling block to AI use, says Narayan, as long as technology vendors deliver on their promise of easily adoptable solutions.
Bigger than both of those barriers are the intellectual property questions swirling around AI. Who will own the data processed with deep neural networks? Panelists at HRS.19 agreed that more work is needed to resolve this and other issues. But, as Friedman told attendees, “The future is now.”