Subscribe to eTOC
Figure

Will Machine Deep Learning or Artificial Intelligence Replace Neurology Know-How?
Not Likely, Say Neurologists

Article In Brief

Two neurologists square off in a tongue-in-cheek debate about the merits of artificial intelligence and machine learning versus the personal neurologist's touch.

PHILADELPHIA—The idea that artificial intelligence and/or machine deep learning devices could someday stand-in for neurologists was rejected overwhelmingly by neurologists after listening to pro and con arguments about the future of their medical specialty here during the Controversies in Neurology Plenary session at the AAN Annual Meeting.

Clearly preaching to the choir, Joseph R. Berger, MD, FAAN, professor of neurology at the Perelman School of Medicine at the University of Pennsylvania in Philadelphia, and associate chief of the multiple sclerosis division, said that while machines may be able to perform amazing feats, nothing can replace the doctor's personal touch.

“Empathy and compassion are critical to what we do and are indispensable elements of neurological care, cannot be replicated by artificial intelligence,” Dr. Berger said. “There is something special about that human touch.”

But David E. Newman-Toker, MD, PhD, FAAN, professor of neurology at the Johns Hopkins University School of Medicine in Baltimore, and head of the division of neuro-visual & vestibular disorders and the Armstrong Institute Center for Diagnostic Excellence, said not only can machines replace neurologist, they also should and will replace neurologists.

His impassioned presentation, delivered tongue-in-cheek, did not find much of a following; the human audience of neurologists, voting on an interactive AAN app, voted 65 percent to 35 percent that, indeed, human neurologists would not be replace by machines.

“I suspect there is a little bias in our voting,” plenary co-chair Amy Brooks-Kayal, MD, FAAN, chief of child neurology and Ponzio Family Chair in Pediatric Neurology at the University of Colorado Denver, commented wryly as the votes were tabulated.

“I don't think that artificial intelligence will ever replace neurologists,” Dr. Berger said, “because unlike other disciplines, neurology is not just a matter of pattern recognition. There are many other things that we do.

“Obtaining a history is critical and requires a human interface,” he said. “The stories need to be meticulously teased out from the patient and relevant others. Language used often needs clarification: How many times have we heard a patients say they are dizzy, and we have to ask: What do you mean by that? Is it an illusory sense of movement or are you lightheaded and are about to faint?”

History taking takes time as well, to get to know the patient, he said.

Dr. Berger also suggested that development of a robot neurologist would not be financially worthwhile. “We are simply too cheap,” he said, noting that health care costs in the United States total about $3.3 trillion a year, but the cost of neurologists amounts to about $372 million—or 0.012 percent of the health care expenditure in the United States a year. This is why, Dr. Berger suggested, it is hardly worth developing a robot to replace them.

But, he didn't want to imply that artificial intelligence was worthless to the neurologist. “We will employ artificial intelligence to augment what we do as neurologists,” he said. “Artificial intelligence will provide us more time to spend with our patients. Artificial intelligence will assist in preventing diagnostic errors. Artificial intelligence's vast compendium of the medical literature will suggest treatment regimens supported by the best evidence. But artificial intelligence will never replace us!”

Figure

“Obtaining a history is critical and requires a human interface. The stories need to be meticulously teased out from the patient and relevant others.”

—DR. JOSEPH R. BERGER

Dr. Newman-Toker disagreed. “There have been real advances in the use of technology and learning networks letting machines think at a higher level, systems that are mimicking the human brain in many respects,” he said. “Artificial intelligence is already better than some of our colleagues, for instance, better than dermatologists in detecting skin cancer.”

He said that once the image data are sufficient, artificial intelligence can learn what is cancer and what is benign. He cited one study [Esteva et al., Nature 2017] that illustrated that machines were more sensitive than dermatologists in determining what was cancer, and what was melanoma.

Dr. Newman-Toker said that “just with commercial grade video technology we are now able to digitize the entire process of the examination. All of the components of neurology that rely on visual pattern recognition and interpretation are potentially accessible to machines and artificial intelligence.

“So machines can replace us, and here is why they should replace us,” he said. “It is an equity issue. Access to quality health care is unfairly distributed around the world and even in America there is very limited access to neurologists in so-called neurology deserts. Improving diagnosis is also a public health imperative.”

The public health system in the United States currently misses about 9 percent of strokes at first contact—that accounts to as many as 100,000 missed strokes or transient ischemic attacks, Dr. Newman-Toker said. Meningitis, encephalitis, and spinal abscesses are infectious diseases missed at high rates, he noted.

“How are we going to solve this problem without machines, especially in areas where there are no neurologists?” he asked.

Dr. Newman-Toker presented a future in which machines do replace neurologists.

He said that health care is in a phase which, after concentrating innovation in centralized academic centers, these innovations are now becoming decentralized. This combined with digitization “has allowed us to migrate those technologies out to the periphery more and more.”

He said he uses portable video-oculography—the Eye-ECG—to tease out cases of dizziness that may be relation to stroke, an area fraught with a high percentage of misdiagnosis. He suggested that by digitizing this test it could be applied to machine learning, and, “you don't need a neurologist to do the examination. A technician at the level of an ECG technologist can do it, and that produces results that can be shifted to experts—at this initial phase—for teleconferencing machines that would be interpreted by humans. Later, through development of decision support technology based on tens of thousands of cases, machine learning or artificial intelligence can determine if the patient has a stroke or something else.”

Dr. Newman-Toker said that the National Institutes of Health is already developing this system through the AVERT trial, and he predicted “this is the wave of the future.”

He suggested that the steamroller of machine learning in neurology is coming and current neurologists have to decide whether they want to be in front of the steamroller or driving it.

Figure

“There have been real advances in the use of technology and learning networks letting machines think at a higher level, systems that are mimicking the human brain in many respects.”

—DR. DAVID NEWMAN-TOKER

As a disclosure, Dr. Newman-Toker said, “The AAN told me to do this and told me which side I had to argue for. But I was comfortable enough with this assignment to step up.” His admission was greeted with appreciative laughter, but it didn't sway the audience.

Disclosures

Dr. Berger disclosed relationships with Biogen, Alkermes, Amgen, AstraZeneca, EMD-Serono, Genentech/Roche, Genzyme, Merck, Millennium/Takeda, Novartis, ExcisionBio and Inhibikase. Dr. Newman-Toker disclosed relationships with GN Otometrics and Autronics-Intracoustics. Dr. Brooks-Kayal had nothing to disclose.

Link Up for More Information

•. Schorr E, Brandstadter R, Jin PH, et al.AAN Abstract S39.002: Diagnostic accuracy among neurology residents: Six-year data from the Close the Loop resident clinical acumen assessment project https://n.neurology.org/content/92/15_Supplement/S39.002.