AI ‘Virtually Omnipresent’
Former NIH’er Horvath Explains Why Machines Won’t Replace Doctors
If it feels like computers control almost everything these days and that soon we’ll all have to visit “Doctor Bot” to cure what ails us, then the recent Clinical Center Grand Rounds talk “It’s an Artificial Intelligence (AI) World and We Are All Just Living in It” by the Association of American Medical College’s Dr. Keith Horvath, may offer some reassurance.
“AI is virtually omnipresent,” he admitted. “It’s already telling us what to watch, what to listen to, how to get from place to place and what to buy [and its reach] continues to expand.”
To illustrate AI conquering new territory, Horvath described what he called the “mood watch,” a device designed to absorb speech intonations and provide the wearer with social clues about the speaker’s disposition or temperament. The high-tech wristband could help people with conditions such as Asperger’s syndrome.
“The general AI of science fiction movies is still a ways away,” Horvath reported. It’s specific AI—exemplified by machines such as IBM’s Deep Blue chess master computer of the 1980s and 1990s—that is growing more effective and quickly having an impact on medicine in the U.S. and around the world.
“In the past, AI has had a lot of hype and promise, but in many cases that has been cooled by an ‘AI winter’ in which the promise has not been achieved,” he said. “Where we are right now is at another peak of that hype cycle. Is another winter to follow? We’ll see.”
Currently serving as AAMC’s senior director of clinical transformation, Horvath, a former clinical investigator at NIH who was director of cardiothoracic surgery research at NHLBI and chief of cardiothoracic surgery for NIH’s Heart Center at Suburban Hospital, gave at least five reasons robots won’t replace physicians.
Primarily, he said, advances in AI should never be viewed as tech versus human.
“Humans and machines together can excel in different ways that individually they cannot,” Horvath pointed out. “The two combined can accomplish what neither of them can do alone.”
Secondly, physicians have a non-linear working method that easily adapts to ever-changing conditions and quickly evolving situations. That dexterity is still difficult to teach to a computer.
Next, competent digital technologies need competent professionals. “The ability for humans to override the machine is key,” Horvath noted.
Also, he said, there are tasks that algorithms and robots can’t complete.
Most importantly, AI cannot replace empathy. “As we all know, the first thing in caring for the patient is caring for the patient,” he emphasized.
Horvath described areas of medicine and clinical research where machine learning, a component of AI, already has made significant inroads. For example, he said, it plays a major role in automated imaging disciplines—radiology, ophthalmology, dermatology and pathology. Machine learning can also be applied to tasks involving signal processing, such as electrocardiography and audiology, and in jobs where integration with other datasets helps clinical workflow.
Types of machine learning include supervised, unsupervised and reinforcement.
An example of supervised machine learning would be cancer detection from MRI results, where the machine has “learned” mapping from a tagged dataset.
In unsupervised machine learning, Horvath explained, the machine is fed “massive datasets to sort through and make diagnoses or come up with predictive analytics.” Examples would be sepsis or identifying research subjects for different trials.
Reinforcement training is a hybrid that combines the other two.
There’s no question that computers can search through more datasets faster than humans, Horvath said, but speed and efficiency may not always translate to accuracy.
“Sometimes machines still get it wrong,” he noted, “especially without being assisted or prompted by human guidance.”
To illustrate, Horvath showed two 4-by-4 grids of similar images in which the computer was unable to distinguish between the pictures.
“It couldn’t tell the difference between a parrot and guacamole, or between a Chihuahua and a blueberry muffin,” he pointed out. “Part of that has to do with needing human expertise to understand better.”
AI’s journey to this point has not been without pitfalls, Horvath said. Information overload remains a genuine concern.
“We’ve gotten through the data tsunami of 2005 to 2016 being generated by electronic health records [EHRs], smartphones and wearable devices,” he said. “Industry lined up to figure out how to fix that in the last few years. So, there’s been a fusion of that idea.”
In essence, companies collaborated to find ways to use and model the wealth of data. “This has been buoyed by the 100 percent adoption of EHRs,” Horvath pointed out. “In the next few years, we’re going to see commercialization of this.
“That will allow AI to enhance the treatment capabilities we have and make them more personalized and specific to potentially combine with other types of data such as home or genomic,” he predicted. “Then we’ll get to the point where AI applications are embedded in all of our clinical and investigative workflows.”
Horvath also addressed the double-edged sword of EHRs.
On the one hand, he said, “AI feeds on data, which is critical for delivering evidence-based health care and developing any of the AI algorithms.”
U.S. consumers use approximately 3 petabytes—3 million gigabytes, or 39 years of high-definition television video—of internet data every minute of every day.
EHRs have been called both a “wellspring of data” and a “cesspool of data,” Horvath said. It’s much easier to put data into systems than to get data out.
A lot of folks have a lot riding on perfecting EHR functionality. Horvath recalled that the HITECH Act passed more than a decade ago has twice—in 2009 and 2014—infused $36 billion into the effort to adopt EHRs into practice nationwide. And although acceptance is now nearly universal in the medical community, there have been significant side effects.
“A lot of physicians now feel that they’re data entry clerks and not patient care folks,” Horvath reported.
He said despite such issues, AI’s promise remains manifold—to increase efficiency and decrease costs by shifting human labor to the more complex tasks, to identify workflow optimization strategies, to reduce medical waste by strengthening coordination of care, eliminating over-treatment or low-value care and automating highly repetitive processes that are largely administrative and to allow the physician to focus on actual care.
Basically, Horvath concluded, AI means more time available and less infrastructure required. “Time is essential to the quality of care patients receive,” he noted. “And, better work/life balance for clinicians is critical to success for everybody…AI is not going to replace physicians, but physicians who use AI are going to replace physicians who don’t, and that may be the cautionary tale.”