Reality & Rumor: Will ‘Smart’ Solutions Really Transform Cardiology?

Smart technologies are often touted as the answer to some of cardiology’s greatest challenges. But where does hyperbole end and reality begin?

Even the most skeptical cardiologists can’t help but feel the digital waves lapping at their feet. A “smart stent” with a special micro-sensor that acts as a miniature antenna to continuously monitor hemodynamic changes in the artery and transmit those data to an external reader. An artificial intelligence (AI)-driven system that recently took 1.2 seconds to accurately process and interpret acute disease in brain CT scans, 150 times faster than humans. An algorithm that can forecast how long individual patients will remain in the hospital, their odds of being readmitted and the chances they will soon die.

Suddenly, the notion of a digital phenotype of each patient that records, monitors and analyzes through biosensors heart, lung and electrical activity and knows well in advance of a catastrophic event—while there’s still time to act—is on the not-so-distant horizon. So is the idea that electronic health records (EHRs) could be transformed through machine learning and natural language processing from an amorphous information dump into a robust diagnostic and prognostic tool not only to improve healthcare delivery and outcomes but to free up clinicians to spend more face time with their patients. 

“The unique changes that are occurring as information technology collides with healthcare technology are probably the most exciting thing I’ve seen in my 35 years in medicine,” declares Peter Fitzgerald, MD, PhD, who is deeply involved in that change as professor of medicine and engineering as well as director of the Center for Research in Cardiovascular Innovation at Stanford University Medical Center. “The system today is archaic and needs to be disrupted, and it will be by the new players like Google and Apple and incredible IT folks who are coming on board.”

Smart solutions to some of cardiology’s most vexing and long-standing challenges have become grist for countless blogs, papers, conference talks and cocktail circuit chatter. At a time when physicians feel more besieged than ever by paperwork, regulations, reimbursement and burnout, the promise of technology and advanced analytics to create a more efficient and effective healthcare system is catnip to many. But there’s a caveat here: where do we draw the line between potential and hype? 

“A lot of the talk about digital health and artificial intelligence solutions is completely lacking any evidence they can actually deliver on the promise or provide a good road map of how it’s going to be clinically validated and integrated into our care,” emphasizes John Rumsfeld, MD, PhD, chief innovation officer for the American College of Cardiology (ACC).

But Rumsfeld is also among the first to cite the smorgasbord of opportunities that has the cardiovascular community excited. Wearables and smart phones have already become mainstream devices for monitoring patients’ heart rate, blood pressure, breathing patterns, glucose levels, asthma and much more, then uploading the data to the cloud for viewing by their physician. Before too long, echocardiography may be compressed into a smartphone application; wearables may measure not just heart rate but heart rate variability; wristwatches will reliably predict and know when a patient has atrial fibrillation; micro-radar sensors will detect heart and lung activity without the need for electrodes; and entire metabolic panels will be collected noninvasively (and remotely) through devices strapped to patients’ arms or foreheads. 

Still, for any of these applications to succeed, a much larger issue must be addressed: what to do with the torrent of data generated so that it can be analyzed and interpreted in a way that benefits patients and physicians.


That’s where AI and machine learning could be transformative. As Alfred Bove, MD, PhD, professor emeritus of medicine at Temple University School of Medicine in Philadelphia and a former ACC president, points out, physicians are already overwhelmed with terabyte-size data flows from their patients. “One way to handle it is to build logic systems that will collect the data, filter it and advise the doctor based on the best and least likely diagnoses,” he says. “The next step would be to recommend a plan of action that could include medicines to be taken and images that are needed.”

For Eric Topol, MD, director of the Scripps Research Translational Institute and author of a forthcoming book, Deep Medicine, the inherent strength and promise of AI reside in “deep phenotyping so you can understand each individual at an unprecedented level,” which means biologic and genomic along with the anatome and physiome. “But first you need to be able to assimilate and accurately i nterpret all that data, which is where deep learning and AI fit in so well.” 

At the heart of AI are computational algorithms that train on large, carefully annotated datasets to look for predictive patient patterns as a way to turn information into knowledge and action. Deep learning allows these algorithms to constantly train and get better at their tasks over time. “Until now, large, labeled datasets which are needed to train have not been plentiful in healthcare,” Topol says. He points to “gains with genomes and high-resolution images and electronic records” while emphasizing the need for rigorous evaluation after training to prove the algorithms can work in a clinical environment without a glitch. “Unlike a doctor who treats a single patient, an algorithm that’s misleading could potentially harm large numbers of people before it’s detected,” he warns.


If many applications for advanced analytics loom further down the road, imaging could be one of its earliest success stories. The reason, says Rumsfeld, is that the underlying data quality of images—against which machine learning algorithms are executed—is very high, allowing for complex pattern recognition and iterative learning. “I believe in the not-too-distant future AI will pre-read cardiac CTs, MRs, echoes and probably electrocardiographic tracings of all types,” he predicts. “It won’t replace the role of the cardiologist, who would still do an over-read. But it may be that while you can read one or two dozen studies in a day, that may double or triple with AI-supportive pre-reading.” 

Other experts in the field firmly believe that within the next decade virtually all imaging studies will be pre-analyzed through artificially intelligent machines before they ever get to the physician. AI will further enhance the review and diagnostic processes by data mining the patient’s electronic health history for salient information, allowing for a more integrated clinical-imaging approach to patient care than is now possible.

Unlike high-resolution images, EHRs suffer from poor data quality because much of their clinically relevant information is in the form of unstructured text notes. Add in clinicians’ observational biases, and you’re left with a system that needs massive   research and re-engineering before it’s ready, in the eyes of many experts in the field, to take on such complex AI challenges as risk prediction and therapeutic decision making. Still, the notion that the EHR could someday become a pillar for healthcare by making it infinitely more responsive to the needs of patients and clinicians is alive and well. 

“At the confluence of electronic medical records and artificial intelligence is the potential for benefiting the healthcare we deliver and improving outcomes,” observes Andrew Einstein, MD, PhD, associate professor of medicine in radiology and director of cardiac CT research at Columbia University Medical Center in New York. “It’s not just a case of having a clinician review [the patient’s]  previous records, but getting a second opinion through tools such as natural language processing, machine learning and a big-data-based approach to prognostication, diagnosis and recommendation of potential  therapeutic plans.” Among the enduring advantages of an imaging platform driven by advanced analytics, he adds, would be the ability to pick up errors before they get miscommunicated and propagated through the system, possibly for years, and to correlate genetic data and clinical data to provide a more personalized approach to patient management.

An example of how a technologically astute and agile EHR system might work is on display at Lakeland Health, a not-for-profit community health system in southwest Michigan. Described in Fortune (March 19, 2018), the platform, installed in 2016, has dramatically updated the slow and error-prone practice of nurses recording patients’ vital signs on charts and manually re-entering the data into the hospital’s electronic system once they returned to their workstations. Now, data are automatically uploaded from patient wristbands or entered by nurses at bedside on handheld devices. One welcome change is that nurses spend less time on data entry and more time with patients. Another huge—and unexpected—improvement was a 56 percent drop in “code blue” warnings for patients in cardiac or respiratory arrest. The reason? An AI-driven warning system built into the monitoring technology that picks up even subtle changes in vital signs and assigns patients risk scores that enable nurses to prioritize patients.


Indeed, risk prediction is one “smart” application that continues to intrigue and tantalize healthcare professionals with its siren-like call. More to the point, could an intelligent system that ingests and learns from millions of cases accurately trace a patient’s trajectory, actually know when cardiomyopathy, heart failure or arrhythmia will turn deadly?

“We’re starting to learn that we can unravel risk of heart disease even at birth like we never could before,” Topol says. “In the future, this phenotyping is going to indicate the risk levels at very young ages so that people can take preventive steps. And we’ll become much better at predicting acute events, like heart attacks, heart failure, seizures and asthma attacks.” 

As Einstein points out, a variety of tools—imaging, serology, clinical and physical exams—already provide data with prognostic importance. But the use of AI, he says, “has the potential to integrate this data with data from large datasets to more accurately identify patients at higher risk for events and change their course by initiating preventive therapy.” Diagnoses supported by machine learning could be expected to be as good as, if not better than, those from humans alone, in his estimation, since they are more likely to pick up on subtle clues from images that could be missed by an expert.

While not downplaying the potential, Rumsfeld is quick to cite the barriers to a dependable risk prediction model, the most formidable being the shortcomings of the current system of electronic health data. Calling it a work in progress, the ACC’s innovation chief stresses the need for more research to validate that input from a risk prediction tool would be better than what’s currently available, and to lay out how those data would be used in clinical decision making. To those ends, this summer, the ACC announced a partnership with Yale New Haven Hospital’s Center for Outcomes Research and Evaluation and the formation of the Institute for Computational Health (ICH). What’s unique about this effort, Rumsfeld says, is that it involves practicing cardiologists and hospitals working collaboratively with the technology sector—including computer scientists, biostatisticians and engineers—to co-create, validate and bring to market innovative solutions to healthcare delivery and outcomes.


If technology is on the cusp of reconfiguring the healthcare landscape, where does that leave cardiologists? Will they become accessories to machines that can work around the clock without judgmental bias, fatigue of fear of burnout?

That’s highly unlikely. Fitzgerald, who’s leading the technology charge at Stanford University, doesn’t even use the term “artificial intelligence” in the same breath as medicine. Instead, he has coined “intelligence augmentation,” or IA. “You’re never going to take the physician out of the loop,” he contends. “You’re just going to augment that doctor with information learned by population statistics and by other algorithms so that a 30-year-old who enters the cardiology field should be as smart a 60-year-old.”

That thought is echoed by Bove. “Even with AI making diagnoses and perhaps suggestions for treatment, you’re still going to need doctors to look at the whole picture and at how the patient’s behavior integrates with everything else,” he says. “You’ll need them to be the ultimate arbiter of all the data.” Bove is part of a sizable group that believes there will always be a role in healthcare for “low tech,” that even the fanciest computational algorithm can’t make a patient take their medicine. “I think what will never become obsolete is the face-to-face personal encounter with the patient,” he says. “That may ultimately be the most important thing we do as physicians to improve the outcomes of healthcare.”