Artificial intelligence assists cardiologists with workflow, diagnoses

Artificial intelligence (AI) and its forms of implementation via either supervised learning, unsupervised learning, machine learning or cognitive computing are already in use in some forms within medical settings, including cardiology. Its role is expanding rapidly, and medical professionals continue to grapple with trying to better define the role of AI, as well as to break down the barriers to adoption.

The concept of AI has been in use for more than a couple of decades in cardiology. AI, in its broadest forms, can be categorized as narrow or general forms of AI. Narrow AI is widely used in computer systems to carry out specific tasks. An example in cardiology is an implantable cardioverter defibrillator for patients who are at risk for sudden cardiac death, as the ICD has built-in intelligence to detect abnormal heart rhythms and deliver a shock to treat the patient. Many implantable devices with their sensors have begun to use sophisticated algorithms to predict and thereby prevent HF.

Future trends are geared to integrate this device-based data, from FDA-approved wearable and implantable sensors via remote continuous monitoring with electric health records. It is here that AI-based strategies will sift through large volumes of data and help deliver individualized treatment approaches. The general form of AI, which is akin to the adaptable intelligence of humans, is still several years away from finding its way into clinical practice.

Other applications of AI include remote monitoring to detect rhythm disturbances and blood glucose levels and an FDA-approved ECG monitoring device that utilizes both an Apple Watch and a smartphone application (KardiaBand, AliveCor).

“Artificial intelligence is prevalent today in a very limited form, and although there are many enthusiasts who may claim that it is around the corner, there are significant practical hurdles that we have to yet overcome before it truly becomes mainstream,” Jagmeet P. Singh, MD, DPhil, associate chief of the cardiology division at Massachusetts General Hospital and Cardiology Today Editorial Board Member, said in an interview.

Cardiology Today spoke with Dipti Itchhaporia, MD, Robert and Georgia Roth Chair of Cardiac Excellence and medical director of disease management for Hoag Heart and Vascular Institute in Newport Beach, California, assistant clinical professor at University of California, Irvine, who, along with colleagues, published a paper in the Journal of the American College of Cardiology in 1996 that focused on the potential of AI at that time and in the future. This paper highlighted artificial neural networks and the ability to predict relations within datasets. Since then, AI has developed into a powerful tool.

“Since then, we’ve really broken AI down, so now neural networks play an integral role,” Itchhaporia, a Cardiology Today Editorial Board Member, said in an interview. “They’ve come a long way to define it, people understand it more and there’s a ton of interest. We’re at the ‘almost there’ stage where we’re just standing at the forefront of something that’s really going to revolutionize [cardiology].”

Current uses and impact

The use of AI has already been shown to reduce workload while preserving patient safety. In a study published in Europace in 2016, researchers developed a pilot system that used algorithms to analyze electronic health records and the anticoagulation status of patients with pacemakers. This system classified the clinical importance of nearly 2,000 atrial fibrillation alerts in 60 patients, facilitating the decision-making process. Notably, the integrated approach resulted in an 84% reduction in notification workload for the clinician.

“This is a simple representation of where artificial intelligence currently exists in its earliest form, and with continuous learning and tweaking the algorithms along the way, we could make patient care more individualized while reducing physician burden significantly,” Singh said.

This technology may also aid in preventing burnout among health care providers, experts said.

“We’re being deluged with data everywhere,” Itchhaporia said. “We’re required to interpret these data very quickly and make clinical decisions based on these data. We also have pressures for operational efficiencies from our health care systems, from Medicare, from all the payers and really from ourselves and the patients. The patients want information faster and ultimately they want more personalized care. To be able to meet all these demands, we need a mechanism by which we can ease the burden of being able to process all this information quickly.”

The use of AI to analyze data from EHRs may also help clinicians monitor the health of entire populations to gauge health, experts said.

Predicting cardiac conditions is another crucial benefit of AI. With early detection, corrective actions can be taken sooner.

“Right now, we’re pretty good once someone gets their condition and comes into hospital ... but getting better at identifying those individuals way before they have their event and predicting who is going to have an event is a big area where AI can help down the line, as well as managing some of these complex conditions that have multiple variables involved with them like heart failure,” John P. Higgins, MD, MPhil, MBA, sports cardiologist and professor of cardiovascular medicine at The University of Texas Health Science Center McGovern Medical School in Houston, told Cardiology Today. “The great potential is with using these artificial intelligence algorithms along with these smart devices (eg, Microsoft’s Glabella BP tracking glasses), we may move toward a place where patients are remotely managed and will rarely have to come into the hospital.”

AI is also being used in the catheterization lab, especially when selecting patients who require catheterization. The technology is used to assess tests such as echocardiograms and CT scans to determine who is best suited for the procedure. Through this, it can reduce variance in practice amongst clinicians.

Automated algorithms for edge detection and plaque recognition during cardiac catheterizations can also assist clinicians when deciding how best to treat a patient, either by bypass surgery, maximal medical therapy or receiving a stent, experts said.

Another aspect of AI — virtual reality — is also making its foothold in cardiology, especially for training.

“It is through virtual reality that you can actually have the experience of performing different kinds of surgeries and/or procedures, even before actually undertaking them for real,” Singh said. “AI-based interactive strategies in a simulated environment that can prepare interventionalists for intraprocedural challenges and complications are in their early stages.”

Robotics using AI-driven algorithms can be used to perform procedures from the console room. Moving joysticks or simply targeting areas allows the clinician to move catheters often automatically to ablate areas of the heart and implant stents.

When considering all the current uses of AI, the most prominent area is assisting physicians to diagnose patients with complex medical imaging, experts said.

“It’s really about making cardiologists themselves more efficient and more effective because we’re getting inundated with huge amounts of information as imaging modalities come in and new biomarker tests and other things brought onto the market,” Joel Dudley, PhD, director of the Next Generation Healthcare Institute and associate professor of medicine, population health science and policy and genetics and genomic sciences at Icahn School of Medicine at Mount Sinai, told Cardiology Today. “There’s a lot more information, so artificial intelligence can help organize information and present it in a way to clinicians so they can make faster and more effective, more consistent decisions about individual care.”

This can also be a helpful tool that goes beyond imaging and helps clinicians with information presented within EHRs.

“The great thing about machines is that they can analyze a large amount of clinical data very effectively,” Collin Stultz, MD, PhD, professor of electrical engineering and computer science at the Massachusetts Institute of Technology and a cardiologist at Massachusetts General Hospital, said in an interview. “Machine learning can analyze data in complex ways, yielding new metrics that identify patients at high risk of future adverse events. Machines have the advantage that they can analyze vast amounts of data — much more than any person could examine in his or her lifetime — to provide useful information to clinicians.”

According to a review published in JACC in June, machine learning can also be split into unsupervised or supervised learning. Each of these have different goals, Kipp W. Johnson, BS, student at the Institute for Next Generation Healthcare of the Mount Sinai Health System and the department of genetics and genomic sciences at Mount Sinai Icahn School of Medicine, and colleagues wrote. Unsupervised learning aids in discovering the underlying relationships or structure of variables within a dataset, and supervised learning involves the classification of an observation into one or more outcomes or categories.

Machine learning is the form of AI that generates the algorithms used to solve problems seen in the clinical environment, experts said.

Possible challenges

Even with its potential, AI brings about challenges. One such challenge involves a lack of datasets.

“The only way you can create and then test new sophisticated algorithms is to have large clean datasets,” Singh said. “These can’t be achieved from a few hundred patients, but in fact needs to be several hundred thousand patients with multiple data points that can reproducibly generate and validate these algorithms. Another hurdle beyond the training datasets is to help foster the mindset and appropriate environment for collaborative strategies among academic institutions and industry to all work together.”

The quality of the datasets determines the quality of the AI.

“Artificial intelligence is only going to be as good as the data that you put in,” Higgins said. “If the data that you put in are bad — garbage in, garbage out. If the data that you put in, for example, are biased — it might be sex biased or race biased, then of course that’s going to affect the way that decisions are made.”

Another way to prevent bias from affecting AI is by involving multiple health care providers.

“We all have to be involved with the way this is created so that we understand potential biases,” Itchhaporia told Cardiology Today. “You need multiple groups looking at it so you don’t have any one group’s bias being inherent in that.”

Some advances have been made to understand how AI makes decisions or how to improve on decision-making, experts said, although more research is needed to ensure that the internal state of AI is not going to change suddenly.

The security of the systems themselves may also be vulnerable to hacking, which would require data protection.

One could argue that AI may also provide additional security to sensitive patient data, experts said. According to a paper published in Scientific Reports in 2016, Dudley and colleagues assessed the performance of a novel unsupervised deep feature learning method for EHRs to predict the probability of a patient developing various diseases. Researchers found that the novel method outperformed methods that used raw EHR data and other feature learning strategies.

“The interesting thing about that is that the deep patient model actually contains lots of predictive information about our patients,” Dudley said. “If I were to share that with anybody, they wouldn’t be able to identify any of our patients from that model because it’s this black box of neural network representation, so that means you could on one hand improve security because you’re able to do machine learning data in these opaque models across health systems, but actually have them share the raw data.”

Regarding research, any results produced by AI should be confirmed with data from different regions and populations to determine whether it is a true association, experts said.

Not only do results produced by AI need to be validated, but also the technology itself.

“No artificial-intelligence-based strategy will become a clinical tool unless it is appropriately vetted and validated in patients through clinical trials,” Singh said.

Further research

Although AI has the potential to affect clinical and research environments, some aspects still require additional research.

Studies are needed to compare AI with standard procedures regarding detection, screening, diagnosis and management of complex conditions, experts said. These studies should focus on the imaging cardiologist, interventional cardiologist and electrophysiologist who work with complex conditions such as HF, CAD, hypertension and arrhythmias.

Predicting and preventing these conditions is the true benefit of AI, experts said.

“The true benefit of artificial intelligence will be when it makes medicine more proactive, not reactive,” Singh said. “Clinical care, which is patient-centric and preventative, is the direction towards which artificial intelligence needs to evolve.”

Research is also needed on the potential for wearable technology to interact with implantable technology, which can lead to improved clinical care through its integration with EHRs, experts said.

Procedures, including interventions and electrophysiological procedures, may also benefit from AI, as it can help collect complex anatomical and physiological information leading to the delivery of more individualized care, experts said.

Research is also needed to see whether AI can be incorporated into systems that can categorize tests such as echocardiograms for clinicians to determine which patients need to be treated sooner than others.

“We have limited clinic space and there’s always waiting times to getting in to see doctors,” Higgins said. “Which patients — based on the algorithms — should be booked to be seen by a doctor next week, which ones can wait to be seen by them for several months, as well as all the pattern recognition work that’s being done on everything from the ECGs to integrating the X-rays, integrating basically all the medical record of the patient that they have available electronically.”

Collecting millions of datapoints is also critical to develop algorithms to be used in AI. Dudley and colleagues are currently working on a prototype of a data-driven health clinic that focuses on collecting digital information that would then be fed into predictive models.

“Until we start reimagining the cardiology clinic in a way that feeds better information and collects that information outside of images, we’re really going to be struggling to apply artificial intelligence in health care,” Dudley said.

Even though more datapoints are needed, progress has been made not only in the collection of this information, but also the computing power to process them.

“Now, we not only have the ability to generate large datasets from hundreds of thousands of patients with millions of datapoints, but also can boast of the required computing powers for analytics, machine learning and generating algorithms that can significantly impact clinical care,” Singh said.

AI with human interaction

Regardless of the benefits that AI has on cardiology, some clinicians are skeptical about using it, especially because of the rhetoric that it may one day replace them for some tasks, experts said.

“They all should really see it as it should make everyone’s lives a lot easier if we’re able to embrace it because it’s really going to be about artificial-intelligence-assisted physicians rather than artificial intelligence replacing physicians,” Dudley said.

In addition, AI on its own cannot offer what a clinician can provide to patients.

“A machine can’t sit down with the patient and ask him or her what their wishes are, how aggressive they want to be with their treatment or about the quality of life issues that are important to them. Clinicians can,” Stultz said. “As long as machine learning and artificial intelligence are viewed as methods that provide complementary information to the clinician, the insights arising from these sophisticated analyses can be integrated into a clinician’s existing workflow. These methods can therefore be used to obtain a holistic view of the patient.”

AI is missing on integral characteristic that cannot replace a health care provider, experts said.

“Physicians in general have to remember we have an inborn ability to be empathetic,” Itchhaporia said. “Ultimately, computers are never going to be as empathetic as a physician or any health care provider for that matter. We have to keep those things in mind as we go along.” – by Darlene Dobkowski

• References:
• Itchhaporia D, et al. J Am Coll Cardiol. 1996;doi:10.1016/0735-1097(96)00174-X.
• Johnson KW, et al. J Am Coll Cardiol. 2018;doi:10.1016/
  j.jacc.2018.03.521.
• Miotto R, et al. Sci Rep. 2016;doi:10.1038/srep26094.
• Rosier A, et al. Europace. 2016;doi:10.1093/europace/euv234.

• For more information:
• Joel Dudley, PhD, can be reached at Genetics and Genomics, Floor 15, Room 15-09, 770 Lexington Ave., New York, NY 10065; email: joel.dudley@mssm.edu.
• John P. Higgins, MD, MPhil, MBA, can be reached at 6431 Fannin St., MSB 4.262, Houston, TX 77030; email: john.p.higgins@uth.tmc.edu.
• Dipti Itchhaporia, MD, can be reached at 520 Superior Ave., Suite 325, Newport Beach, CA 92663; email: drdipti@yahoo.com.
• Jagmeet P. Singh, MD, DPhil, can be reached at Cardiac Unit Associates, 55 Fruit St., Boston, MA 02114; email: jsingh@mgh.harvard.edu.
• Collin Stultz, MD, PhD, can be reached at 77 Massachusetts Ave., Cambridge, MA 02139; email: cmstultz@mit.edu.

Disclosures: Dudley, Higgins and Itchhaporia report no relevant financial disclosures. Singh reports he consults for Abbott, Biotronik. Boston Scientific, Impulse Dynamics, LivaNova, Medtronic and Toray. Stultz reports he is on the scientific advisory board for Unlearn.AI and is a consultant for Peach IntelliHealth.