AI for Writing, for Improving Practice and Patient Care
To manage his “task overload,” Dr. Cush uses ChatGPT for jobs such as first drafts of articles and making PowerPoint slides. It must be used cautiously for medical writing, however, as inaccuracies and false data/fabricated information — some of which has been coined AI “hallucinations” — are not uncommon.
“It’s very good at manuscript drafts, at generating bibliographies … it can do systematic reviews, it can do network meta-analyses, and it can find trends and patterns that can very helpful when it comes to writing. But you have to know how it’s a tool, and how it can hurt you,” he said.
Researchers recently reported asking ChatGPT to write an editorial about “how AI may replace the rheumatologist in editorial writing,” Dr. Cush noted. ChatGPT was “very politically correct,” he quipped, because it wrote that AI is “a tool to help the rheumatologist, but not replace him.”
Publishers want to preserve human intelligence — critical thinking and the ability to interpret, for instance — and most of the top medical journals (those most often cited) have issued guidance on the use of generative AI. “One said AI can’t be attributed as an author because being an author carries with it accountability of the work, and AI can’t take responsibility,” Dr. Cush said. Journals also “are saying you can use AI but you have to be totally transparent about it … [how it’s used] has to be very well spelled out.”
In practice, chatbots can be used for summarizing medical records, drafting post-visit summaries, collecting patient feedback, reminding about vaccinations, and performing administrative functions. “It’s really limitless as to what chatbots can do,” Dr. Cush said. “The question is, [what is] really going to help you?”
Much of the research submitted for presentation at major rheumatology meetings over the years has had questionable real-world utility and value, he said. But in the future this will likely change. “Take the PsA [psoriatic arthritis] patient who hasn’t responded to methotrexate or apremilast [Otezla]. There are [so many] choices, and there really isn’t a clear one. Shouldn’t data guide us on whether an IL-23 is better than a JAK, or maybe a JAK preferred over a TNF for some reason?” Dr. Cush said. “That’s what we’re hoping will happen down the line.”
More realistic AI-guided clinical scenarios for now include the following: AI screens the chart of a 68-year-old with RA on methotrexate and etanercept who is following up, and retrieves pieces of history — an elevated C-reactive protein 3 months ago, for instance, and diverticulosis 5 years ago. “AI tells you, based on this, he may have active disease, and here are three medications covered by his insurance,” Dr. Wells said.
Or, in the case of a 58-year-old patient with RA who has scheduled a virtual follow-up visit after having been on methotrexate and hydroxychloroquine for 12 weeks, AI detects a low platelet count in her previsit labs and also sees that she received an MMR booster 5 weeks ago at a local CVS Minute Clinic. AI retrieves for the rheumatologist a review article about thrombocytopenic purpura after MMR vaccination.