ChatGPT: Relieving Healthcare’s Administrative Burden?

That jaw-dropping quote about artificial intelligence (AI) is from Dr. John Halamka. Previously Harvard Medical School’s chief information officer, today he’s the president of the Mayo Clinic Platform, the health system’s initiative designed to improve care delivery through new technologies. And his point is that healthcare administrators and clinicians who understand how to effectively use AI platforms like ChatGPT will enjoy substantial competitive advantages in the job market, along with more promotions and raises.

Why? Although at some point these chatbots may directly aid diagnosis, treatment, and prevention, their capabilities are first likely to help relieve healthcare’s administrative burden—a developing crisis with severe consequences throughout the industry.

Most of us have no idea just how much time clinicians and healthcare administrators actually spend on paperwork, a job function that’s exploded since the widespread adoption of the electronic health record (EHR) platform by health systems across the United States between 2011 and 2016. This is what’s meant by the healthcare industry’s term administrative burden.

Although this phrase primarily refers to advocacy supporting prior insurance authorizations, it also encompasses clinical documentation charting combined with administrative reporting required by organizational and government oversight policies. For clinicians, this burden also comprises attendance in continuing education courses required to maintain their licenses.

The amount of time and resources devoted to managing this burden is staggering because about a quarter of America’s $3.8 trillion annual healthcare spending goes to healthcare administration, more than in any other nation. For example, a landmark 2016 tracking study across four states published in the Annals of Internal Medicine concluded that physicians’ administrative burden amounted to roughly two hours of work for each hour of direct patient care.

Then in 2018, a survey of clinician compensation by Medscape described as “mind-boggling” the time actually spent each week by physicians on these tasks. The publication reported that a whopping 89 percent of doctors spend five or more hours each week on administration—including almost a third of the sample who routinely invested at least half of what amounts to an entire workweek for most Americans: 20 hours or more.

According to the Advanced Clinical Education Association (ACEA), as much as a quarter of a typical physician’s time is devoted to these administrative duties. ACEA also estimates that the opportunity cost of the total time spent managing this administrative burden amounts to about $15.5 billion annually.

Not only do nonclinical tasks like these deplete the time available for healthcare workers to interact with their patients, but these bureaucratic duties also amount to a leading cause of clinician burnout. From 35 to 50 percent of all medical professionals suffer from burnout; research shows that once they experience their first episode, the condition periodically resurfaces throughout the rest of their careers. A 2016 article in the Lancet pointed to three main elements: emotional exhaustion, depersonalization of patients, and reduced personal accomplishments.

Burnout poses risks to patient safety as well as medical quality. A 2019 study also published in the Annals reported that a single point increase on an emotional exhaustion ratings scale correlated with a five percent increase in medical errors. Moreover, the associated medical error rate climbed by 11 percent for a single-point increase on a depersonalization scale.

So how could artificial intelligence help reduce this administrative burden? One of the first ways it can help is by slashing the time healthcare administrators and clinicians spend writing letters to insurance companies requesting prior authorizations and appealing claim denials. This was one of the key initial capabilities of the ChatGPT platform that clinical administrators and doctors first recognized.

San Francisco startup OpenAI released ChatGPT on November 30, 2022. Only two weeks later, Palm Beach rheumatologist Dr. Clifford Stermer had already figured out how to program the software to write a letter appealing an insurance company’s coverage denial. He then posted a hypothetical demonstration on social media of how fast and easy this process might be.

The video shows Dr. Stermer’s entering into ChatGPT a simple natural language prompt: “Write a letter to United Healthcare asking them to approve an echocardiogram on a patient with systemic sclerosis. Make references to supporting scientific literature, and list the appropriate articles.” It then took the chatbot a mere 42 seconds to return a compelling three-paragraph insurance appeal form letter, complete with citations referencing two medical journal articles.

At the time of this writing, Dr. Stermer’s video has received almost 604,000 views between two of the platforms where it appears, TikTok and Twitter. With that kind of attention all over social media, it wasn’t long before a clinical administrator in charge of hospital service at a large New England academic medical center would tell Forbes that he saves hours every week with the bot’s help:

Every week, Eli Gelfand, chief of general cardiology at Beth Israel Deaconess Medical Center in Boston, wastes a lot of time on letters he doesn’t want to write—all of them to insurers disputing his recommendations. A new drug for a heart failure patient. A CAT scan for a patient with chest pain. A new drug for a patient with stiff heart syndrome. “We’re talking about appeal letters for things that are life-saving,” says Gelfand, who is also an assistant professor at Harvard Medical School. . .

He fed the bot some basic information about a diagnosis and the medications he’d prescribed (leaving out the patient’s name) and asked it to write an appeal letter with references to scientific papers. ChatGPT gave him a viable letter—the first of many. And while the references may sometimes be wrong, Gelfand told Forbes the letters require “minimal editing.” Crucially, they have cut the time he spends writing them down to a minute on average. And they work.

Then on February 14, the LinkedIn-style professional social network for clinicians known as Doximity introduced a new beta app based on the ChatGPT platform, but which the company’s programmers had also trained on additional healthcare-specific language. This free experimental web tool, DocsGPT, contains a menu-driven curated library of prompts almost ready for clinical use, including descriptions of accomplishments and “wins” shared by the social network’s members.

For example, providers can choose a prompt like one titled “WellCare Denial Appeal” and add specific data to the suggested text to customize the prompt with relevant details from the patient’s history. In seconds, the platform drafts and displays a reconsideration request form letter. The practitioner can then complete the letter by adding the patient’s information before sending it to the payer. We counted 61 prompts like these in Doximity’s database so far, with titles such as:

• Prior Authorization Letter for Continuous Glucose Monitor
• Exploratory Laparotomy Note
• Letter of Medical Necessity (Brain MRI)
• Rx Denial Appeal (Dupixent)
• Reimbursement Redetermination Request

Along with DocsGPT, ChatGPT has won coverage from several influential healthcare industry analysts. For example, a detailed equity research report published for clients of the venerable Chicago investment bank William Blair explains that “our conversations with providers that have used the solution believe it could be a ‘game changer’ in healthcare in myriad areas.”

The William Blair analyst team then cites several examples of the platform’s potential capabilities besides those we’ve discussed. The panel anticipates that ChatGPT could soon also:

• Function as a digital assistant for clinicians; the platform could sort and copy relevant data out of patient records, then categorize the data so that patient visits could be better organized and more efficient
• Compose clinical notes like progress notes and discharge summaries for the patient’s electronic health record, and eventually integrate those notes directly with EHR platforms such as Cerner, Epic, and Veradigm (Allscripts)
• Draft pre- and post-operative care instructions
• Propose replies to email inquiries from patients about diagnosis, treatment, and prevention
• Write medical necessity certification letters
• Extract abstracts from medical journal articles and integrate the abstracts into clinician newsfeed applications
• Order prescriptions and referrals to specialists.

So why aren’t diagnosis and treatment—which appear nowhere in the William Blair analysis—among the very first applications for ChatGPT in healthcare? During a March 2023 edition of the American Medical Association’s AMA Update video and podcast, Dr. Halamka addressed this question:

So the FDA looks at software as a “medical device” in terms of risk. Now, I think we would all agree that if I use ChatGPT to, say, appeal a claim denial, OK, if it got some of the facts wrong—I mean, what’s the risk? It’s unlikely that somebody is going to be harmed.

But if I use ChatGPT for the diagnosis of a complex medical condition, there’s a high potential for harm. So I think in the short term, you’ll see it used for administrative purposes, for generating text that humans then edit to correct the facts. The result is reduction of human burden. And if we look at some of the—what I’ll call the “crisis of staffing,” and the “great resignation” and retirement of our clinicians—burden reduction is actually a huge win.

Besides this risk differential, it turns out that large language models (LLMs) like ChatGPT frequently return wrong, out-of-date, and at times nonsensical results when responding to user prompts asking for facts and details.

Why? When the model is not aware of relevant details, it frequently will just make them up. This effect is called “AI hallucination,” and this tendency is a known weakness not only of ChatGPT’s free and newer subscription versions but of all such platforms including Microsoft’s Bing AI and Google’s late entry in this market space, Bard.

This hallucination problem is so severe that a 2022 Stanford University artificial intelligence report found that many LLMs are truthful only around 25 percent of the time, and the best-measured performance at the time of that study was only about 41 percent.

Things seem to be improving because with the latest ChatGPT 4.0 subscription version released in March 2023, OpenAI claims accuracy of about 60 percent on factuality evaluations for some disciplines. Nevertheless, in Dr. Stermer’s remarkable TikTok video demonstration, one of the two medical journal articles cited in the reimbursement denial appeal letter written by ChatGPT actually does not exist. Says Dr. Halamka:

So I think of AI as not “artificial” intelligence but “augmented” intelligence. . .I did use ChatGPT and said, “Write a press release about the Mayo Clinic’s association with a new technology company. . .”

It generated a perfect, eloquent and compelling press release—that was totally wrong! So then I went in and edited all the material facts. And the end result was a perfectly-formatted document that I could send off, done in five minutes, not in one hour.

So think of it as “augmenting” your capacity—and not as replacing our clinicians.