GPT-4 Is Here. How Can Doctors Use Generative AI Now?

In the past 3 months, ChatGPT has risen from its obscure testing grounds at OpenAI’s headquarters in San Francisco to become the fastest-growing application in history, prompting a deluge of viral medical TikTok tutorials, new research policies, and being the subject of studies itself.

Even as healthcare professionals still figure out how to use this advance in generative artificial intelligence (AI) to their benefit, OpenAI announced the limited release of the new and vastly improved GPT-4, which the company called its “latest milestone” in its work to scale the model’s deep learning capabilities.

Among the many advances revealed in this new version is the improved ability on the Medical Knowledge Self-Assessment Program (MKSAP), a self-assessment tool for physicians developed by the American College of Physicians. Previously, GPT-3.5 managed to score ~53%. GPT-4 scored ~75%, according to OpenAI.

Despite these leaps forward, Robert Pearl, MD, of Stanford University in California, said the potential of generative AI’s impact on healthcare has been significantly understated.

“If it’s 1% of the potentiality, that’s a high estimate,” Pearl told MedPage Today. “Because Moore’s law says that computing power and technology is going to double every 2 years, so that means a decade from now it’s going to be 30 times more powerful — that’s the difference between a bicycle and a racing car.”

Pearl believes ChatGPT has the potential to ease the knowledge and documentation burdens for physicians by providing the most pertinent information exactly when it is needed. He and several other experts interviewed by MedPage Today said physicians should embrace it and start preparing for its future in healthcare.

ChatGPT Bots as Office Assistant

While generative AI still needs to develop and be implemented properly into existing digital infrastructure, Pearl, a former CEO of the Permanente Medical Group, said he thinks the technology could already allow physicians to streamline some documentation processes.

For example, physicians can use generative AI to help write more concise and properly sourced notes, especially with the availability of search-enabled AI chatbots like Bing Chat. This would also allow physicians to prompt AI chatbots to write letters for patients about disease summaries, medication overviews, or even patient visit summaries without requiring the time and focus typically needed.

Another strategy, Pearl said, is to use generative AI as a medical knowledge assistant to provide lists of potential symptoms and recommendations on specific tests related to those symptoms, as well as providing ideas for differential diagnoses.

Some physicians have even documented how they use ChatGPT to write prior authorization letters and rebuttals to insurance coverage denials, as previously reported by MedPage Today.

With all of these tasks, the idea at this point is to use the technology to write a first draft, experts said. Nothing should be disseminated without close physician review and editing.

Nonetheless, “You can start to move into embedding it into your practice,” Pearl said. “You can actually start to use it as a tool to individualize the care you provide to patients. And that will be a big step forward.”

ChatGPT as Research Assistant

Vikram Savkar, a senior vice president and general manager at publisher Wolters Kluwer, said generative AI is already having a big effect on the speed and quality of medical research that can be published.

The most notable use of the technology, he said, is translating text in research papers, because the first criteria that journals use in accepting new manuscripts is the quality of the writing.

In fact, he said, at least half of all submitted research is dismissed immediately because the quality of the English writing doesn’t meet a journal’s standards.

“I think that there are ways in which ChatGPT can help scientists from non-English speaking backgrounds make sure that their paper is written better and accepted more by prestigious journals, reaches more clinicians around the world, and ultimately has positive impacts on more patients’ lives,” Savkar said.

He predicted that because generative AI has the potential to help researchers reach a global audience, the technology will eventually be accepted by journals and publishers will formulate clear policies around the proper use of tools like ChatGPT.

“What is ultimately at the heart of papers is the research that is done by the clinician, and ChatGPT doesn’t affect or create or replace that in any way,” Savkar continued. “It just helps the researcher communicate the research that she or he has done in language that is up to global standards.”

Savkar said he believes the standards already in place at medical research journals will allow for generative AI to be incorporated in ways to expand access to ideas. Medical research publishing is a “carefully mediated” industry that is capable of using this technology without diminishing the quality of the research and science being shared, he said.

AI’s Future in Medicine

Michiel Schinkel, MD, of Amsterdam University Medical Center in the Netherlands, told MedPage Today that physicians should not be concerned about the current technical barriers, because the technology is improving so quickly.

“[Healthcare professionals] don’t need to know the exact details of the inner workings of these algorithms — they just have to get familiar with it,” he said. “If you can convince people to at least try it, then you’re pretty much 80% there.”

Schinkel and his colleagues recently wrote a paper outlining the potential uses for generative AI in research writing, which included a test to show how difficult it is to differentiate human writing from AI-generated writing. He said he believes this technology has already started to have a major impact on the way medical research is being written, but that he also thinks the biggest changes will come when the technology is more widely adopted in medicine.

Pearl noted that the skepticism surrounding generative AI use in healthcare is not unexpected, and that distrust of new technology is part of the culture of medicine.

He highlighted the story of Daniel Fahrenheit, the inventor of the mercury thermometer, a device that could measure temperature to tenths of a degree, in 1714. Despite the technological advance, it wasn’t until the mid-19th century before doctors widely adopted the tool for clinical practice.

It is a story that is frequently shared when talking about the potential of generative AI models in healthcare. Much like Fahrenheit’s early version of the mercury thermometer, the technology for generative AI is still being perfected, but many healthcare researchers and physicians agree that now is the time to put it to use in medicine.

“We can’t wait 100 years,” Pearl said.