In a highly stimulating lecture, Eric Topol, M.D., a bestselling author and a practicing cardiologist at the Scripps Clinic in San Diego and editor-in-chief of Medscape, told a standing-room-only audience at the plenary session on Dec. 2 at RSNA24—this year’s conference of the Radiological Society of North America—that artificial intelligence will transform the practice of medicine in the coming years.
Speaking to a standing-room-only audience at the Arie Crown Theater in Chicago’s vast McCormick Place Convention Center, Dr. Topol, author of the 2019 bestseller Deep Medicine, walked his audience of radiologists and others involved in radiology, through the evolution to date of artificial intelligence, and then predicted based on progress so far, what will happen next.
Topol began by contexting the current moment, noting that 800,000 Americans die or are seriously disabled every year because of misdiagnosis; what’s on the horizon, he emphasized, is a new era in which AI tools will help physicians better diagnose and treat, and even predict the onset of, disease. He said that the foundational work over the past numerous years in developing algorithms and working with large language models, has set the stage for massive change. For example, the data gathered from enormous amounts of data and images, is already leading to better diagnoses, as in the case of gastroenterology, where gastroenterologists are already using AI-facilitated endoscopy to achieve detect more polyps than they could previously. And data is being gathered even from such diagnostic images as x-ray, creating massive lakes of data that are being used to support physician diagnosis processes. This phenomenon he referred to as “Machine Eyes”—the collection of data that, when analyzed and poured into clinical decision support, can improve diagnostics. Amazingly now, studies are finding that the analysis based on chest x-rays can lead to the diagnoses of a surprising range of diseases, including diabetes. He cited a September 2023 study based on the analysis of 1.6 million retinal images gathered in the U.K. that produced breakthrough predictive diagnostics.
Now, Topol told his audience, medicine is on the cusp of being able to make use of two types of multimodal AI—one based on text, speech, and images, and the other based on human data. “Where can multimodal AI take us?” he asked. “You can get into a much different level of precision and accuracy medicine going forward,” he predicted. “For example, hospital-at-home can be contemplated more in the future,” as the analytics needed to support such leading-edge care delivery forms will more and more be available.
What’s more, Topol reported, “Four foundation models in pathology have been posted in the past year,” in clinical journals. They are going to make it possible to achieve “diagnosis from a whole-slide image.”
Meanwhile, he said, what’s becoming clear is that “AI does a really good job of its text for completeness, correctness, and conciseness. AI reports are tighter, easier to understand, and more complete than reports produced by physicians.” He also made note of a couple of studies that have concluded not only that AI does a better job of diagnosis than human physicians, but two studies have found that AI alone actually does a better job of diagnosis than AI + humans. That result, though, he quickly added, is probably related to the fact that the studies were “contrived,” artificial tests, not based on actual patient care situations. It is interesting to note, though, he added, that AI appears to promote the expression of empathy among physicians.
Ambient intelligence and a new range of capabilities
Topol noted that “Generative AI, not just NLP [natural language processing], can be made into audio notes in multiple notes for the patient. In fact, it’s more accurate than normal notes in EHRs. It can set up follow-up appointments, order prescriptions, etc. And it can even coach physicians to become more empathetic and to become better communicators.”
Most of all, Topol said, AI can help to give physicians “the gift of time,” through “keyboard liberation,” the ability to synthesize the patient’s data, the capability to engage in primary screening preview of all images, and the automated diagnosis of routine conditions.
And one of the greatest types of potential, Topol said, is longitudinal data that can facilitate “individualized medicine from pre-womb to tomb.” It’s that kind of data, which was involved in recent research at the Weizmann Institute in Israel, that is uncovering the “biological clocks” inside human bodies that are aging at different rates. That same data could help to individualize diagnostics in oncology; for example, he noted, pancreatic cancer is one of the most difficult cancers to detect very early on in its progression; data analytics could suggest which patients might be most at risk. And research is continuing forward in using plasma proteins, gathered using just “a couple of milliliters of blood,” that can detect disease risk.
The chief obstacles right now to progress in this entire area around AI, Topol said, are the following: medical community resistance to change; reimbursement issues; regulatory challenges; the need for greater transparency; the need for compelling evidence; engendering trust among clinicians and the public; and implementational challenges.
