As I See It: Doctor AI
February 26, 2024 Victor Rozek
AI has been getting a lot of press and chatter lately, its vast potential for good and ill spawning prodigious debate. It’s not surprising that a technology predicted to impact 40 million jobs engenders a wide range of possible outcomes. But whether the future tilts toward transformational or apocalyptic depends, in part, on which profession is doing the assessment.
One of the more ardent transformationalists is Eric Topol. To say Topol is over-accomplished would be an understatement. He is a cardiologist, a scientist, and an author. In his spare time he became the founder and director of the Scripps Research Translational Institute, a professor of Molecular Medicine and, the executive vice president at Scripps Research Institute. Plus, not to be outdone, he’s also a senior consultant at the Division of Cardiovascular Diseases at Scripps Clinic in La Jolla. In short, it’s safe to assume he’s highly respected and knows of what he speaks.
Topol is extremely bullish on AI’s future in diagnostic medicine. Already, he notes, people whose conditions do not improve after treatment recommended by their physician, are entering their symptoms in ChatGPT and receiving diagnoses missed by their doctor.
In fact, misdiagnosis is a huge, underreported failing of medicine with equally weighty consequences. Topol points to a Johns Hopkins study which found that an astounding 800,000 Americans either die, or are seriously disabled, each year through misdiagnosis. Every physician is predicted to have at least one misdiagnosis during their career – a powerful argument for always insisting on a second opinion.
Like doctors, AI is not infallible, but Topol sites instances where lives were saved, or health restored, based on an esoteric AI diagnosis. More remarkable, however, is AI’s medical predictive prowess.
Studies have shown that AI can predict when a patient is vulnerable to a wide variety of diseases and conditions – before any symptoms actually occur. They include: diabetes, high blood pressure, liver and kidney disease, gall bladder issues, and a heart calcium score which, in excessive levels, can interfere with heart and brain function. AI can predict heart attacks, strokes, and Alzheimer’s before any clinical symptoms manifest. Astonishingly, it can spot Parkinson’s seven years before symptoms emerge.
The obvious question, and one that was asked of Topol in a recent NPR interview, is: How is AI able to do that? Does it look at the results of blood work? X-rays? Scans? Of course it can do all those things – and in many instances already does – but its main source of patient data is garnered from . . . drumroll please . . . the retina.
The retina? So the obvious follow-on question was: What does it see that retina specialists can’t? And the wholly unsatisfactory answer was: “We don’t really know.”
Swell.
Nonetheless, the potential for revolutionizing diagnostic medicine is undeniable. Initially, the AI studied “100,000 images with supervised learning,” to determine “if it could see things that people couldn’t see.” Could it, for example, identify the sex of a person from a retinal scan, something retinal experts cannot do. Turns out it could, with 97 percent accuracy.
As a cardiologist, Topol has studied thousands of cardiograms, but admits there were things that graphing the heartbeat will not reveal to the human observer. For example: the age and sex of a patient; the ejection fraction, or amount of blood the heart pumps each time it beats; how anemic a patient is; atrial fibrillation and stroke risk; diabetes and pre-diabetes risk; the filling pressure of the heart; hyperthyroidism, and kidney disease. But AI is able to make all of these extrapolations.
AI can read CT scans, MRIs, and ultrasounds, as well as, or better than radiologists. It can even identify race from a chest Xray, a head-scratcher for physicians. In 21 randomized studies AI was better able to pick up polyps during colonoscopies than gastroenterologists working alone. Notably, AI’s diagnoses were especially useful “late in the day” when physicians are tired and tend to make more mistakes. For cancer patients, AI can identify structural genomic changes, driver mutations, tumors versus normal cell origin, and determine a patient’s prognosis. Topol predicts that, in the not-too-distant future, we will have an app on our phones able to scan our retinas and provide on-demand health screening.
The latest generation of AI from OpenAI and Microsoft, GPT-4, with over 1 trillion parameters (akin to synapses in the human brain), including language, images and speech, is responsible for this medical bounty. And one of its side benefits, which appears to excite Topol nearly as much as its diagnostic prowess, is what he calls “keyboard liberation.” Apparently, doctors spend a good deal of their time typing up notes, instructions, and whatever other medical minutia doctors are obliged to track. AI, claims Topol, can do all of this with ease, and will liberate doctors from functioning as data clerks, thus giving them more time to focus on their patients. Or play golf.
On a very human level, Topol tells the story of six-year old Andrew who, for three years, experienced “relentlessly increasing pain, arrested growth,” an impaired gait caused by dragging his left foot, and severe headaches. During that time, his parents consulted no fewer than 17 doctors, all unable to help. Desperate, his mother entered all of his symptoms into ChatGPT, which came back with a diagnosis of occulta spina bifida (a tethered spinal cord) missed by all 17 doctors. After a surgical procedure to release the cord, Andrew made a full recovery and reclaimed a normal childhood.
As a diagnostic tool, AI has little downside. At worst, AI will verify a physician’s diagnosis. At best, it will catch missed evidence in support of an alternate or additional diagnosis. Precision medicine, so promising a trend, will be even more efficacious when driven by a higher degree of accuracy.
The only possible downside is a growing dependence on machine knowledge that will eventually create a cognitive role reversal. At some point the line will blur between user and tool with uncertainty about whether the doctor is using the tool, or the tool is using the doctor.
Also, human nature being what it is, some medical students will prefer erudition without effort and expertise without knowledge. Perhaps some day they can opt for a chip implant in the brain which will give them access to all the data available to AI. Years of medical school may hardly be necessary. But attractive as a brain implant may sound, it’s good to remember that if Elon Musk has anything to do with it, there will probably be a recall.
RELATED STORIES
As I See It: The Other Eight Hours
As I See It: Elusive Connections
As I See It: Entitlement Master Class
As I See It: Greetings, Comrade IT Professional
Generative AI requires a lot of “quality raw materials” to train.
See the New York Times controversy.
So there is always expert analysis at the ground level.
Somebody somewhere discovered an anomaly, spent time, justified the time spent, money spent, wrote papers, books, and gave the concept a name called “occulta spina bifida”, a list of symptoms, which AI fed on. In this case AI is – what once was called – a very useful “expert system” to query, and should indeed be used by specialists; and to be honest, given a list of symptoms, something that can discovered even by good old fashioned search engine search and time to read and analyze the results.
Could AI have discovered in the first place the existence of a defect in humans called “spina bifida” that it is considered an anomaly? maybe not yet….