‘Upended overnight’: Medical advances in artificial intelligence spark optimism, suspicion
Click Here to Manage Email Alerts
A chatbot may have written this sentence, or maybe it was written by an actual writer. Or maybe it was written by a doctor. It is becoming increasingly difficult to tell these days.
[Editor’s note: It was indeed written, and edited, by actual people.]
Source: Jacob M. van Laar, MD.
It is similarly difficult to divine, at this early stage, whether the impacts of artificial intelligence in health care are cause for optimism or suspicion. What is certain, however, is that what limited forms of AI are currently available — such as machine learning and predictive and/or generative text — will continue to advance, both in terms of technological complexity and within the medical arena.
“The release of ChatGPT opened up Pandora’s box of possibilities, reflecting a shocking jump in capabilities compared to similar systems only a couple years ago,” Jonathan H. Chen, MD, PhD, assistant professor at the Stanford Center for Biomedical Informatics Research, Division of Hospital Medicine, and the Clinical Excellence Research Center, told Healio Rheumatology. “This is compelling, with people across the world discovering novel uses for this technology, but also daunting, with many unexpected and unintended uses.”
In an editorial entitled “ChatGPT: When Artificial Intelligence Replaces the Rheumatologist in Medical Writing” — published in the Annals of the Rheumatic Diseases in April — Verhoeven and colleagues suggested that the “domains are infinite” for the use of AI in medicine. These could, conceivably, include helping, or replacing, rheumatologists “in the writing of scientific articles,” they wrote.
“While this notion may be unsettling for some in the medical community, it is important to acknowledge that AI has the potential to greatly enhance the field of rheumatology,” Verhoeven and colleagues wrote. “The ability of AI to process vast amounts of data, analyze complex patterns and make accurate predictions could revolutionize the way rheumatologists diagnose and treat their patients.”
In fact, this is already happening, according to Leonard H. Calabrese, DO, RJ Fasenmyer chair of clinical immunology at the Cleveland Clinic, and chief medical editor of Healio Rheumatology.
“What many people do not understand is that we have been using AI in medicine behind the scenes for quite some time, to assess skin lesions, retinal scans or pathology slides, among other applications,” Calabrese said. “Patients do not even know that their data is being sorted primarily through AI. We are going to have to confront them with this information in the very near future.”
One concern is that the technology is not yet perfect — to say nothing of the data being fed into it — nor is the human ability to understand all of its applications.
“AI and machine learning can analyze large amounts of data very quickly, but we do not always know what to do with it,” Michael LeTang, MS, RN-BC, vice president and chief nursing officer at Healthcare Risk Advisors (HRA), part of the TDC Group in New York, said in an interview.
However, a machine analyzing flawed data is just one part of the problem. ChatGPT and other content creation technologies raise significant ethical questions in the clinic, as well as in medical education and publishing. Meanwhile, regulations, policies and laws have not kept up with the pace of technological advancement.
“There is a complete lack of regulation,” LeTang said.
Despite these hurdles, many health care professionals are eager to make more regular use of AI, for one simple reason — it can make their lives easier.
For example, dictation and ambient listening technologies that can recognize human speech have the potential to significantly reduce the administrative burden for many doctors and patients. Increasing reliance on AI could also be a solution to the looming workforce shortage that threatens to hobble rheumatology as a field for years to come. Coupled with its current uses in data accrual and analysis, it is not difficult to imagine that AI technology could completely reshape the way health care is delivered in the very near future.
At the Intersection of Technology and Malpractice
In a paper entitled “How Chatbots and Large Language Model Artificial Intelligence Systems Will Reshape Modern Medicine” — published in JAMA Internal Medicine in April — Li and colleagues outlined several broad strokes regarding how the capacity of AI to retrieve and analyze large batches of complicated data could be problematic.
“The risk of course is that this could just as easily propagate false, biased, or otherwise flawed information from such sources without regard for accuracy,” they wrote.
Put another way, human error at the input stage could lead to serious consequences, according to LeTang.
“As health care professionals documenting a visit with a patient, one concern we have for data being input into AI and machine learning systems is that the information we enter is not always accurate, up to date or standardized,” he said.
In short, asking a machine to interpret flawed data will naturally yield a flawed result, which, in the clinic, could be a treatment plan ending in a fatality. Meanwhile, if the data are being input for research purposes, the result could be an incorrect conclusion regarding the safety or efficacy of a medication.
These limitations of AI technology could also have real-world implications for the physicians associated with such mistakes.
“In broad terms, we have made a lot of progress in using technology in health care, from the adoption of the electronic health record in 2009 to recent advances in AI and machine learning,” LeTang said. “But as AI and machine learning start to make its own interpretation of data, the risk of medical malpractice will increase.”
However, tech developers are working on ways to improve the use of AI in rheumatology overall, and not just to minimize the likelihood of malpractice.
In a paper entitled “Artificial Intelligence and Deep Learning for Rheumatologists” — published in July 2022 in Arthritis & Rheumatology — McMaster and colleagues looked at machine learning specifically in the rheumatology setting.
“The greatest benefits of deep learning methods are seen with unstructured data frequently found in rheumatology, such as images and text, where traditional machine learning methods have struggled to unlock the trove of information held within these data formats,” they wrote.
The earliest uses of deep learning — interpreting radiologic images or in assessing findings from a colonoscopy — could point the way forward in rheumatology, allowing clinicians to detect joint erosions or identify a halo sign on a temporal artery ultrasound, according to the McMaster paper.
However, McMaster and colleagues cautioned that it is “imperative” that rheumatologists work in concert with tech developers to ensure that any new tools keep patient care at the forefront.
“The best applications of deep learning in rheumatology must be informed by the clinical experience of rheumatologists, so that algorithms can be developed to tackle the most relevant clinical problems,” they wrote.
As experts wrestle with the data going into AI technology, the information coming out of these systems raises a whole different set of ethical questions.
‘Confidently Misleading or Erroneous’
Calabrese described ChatGPT as the “cause celebre” of the moment.
“Playing around with ChatGPT is pretty awesome but also can be pretty scary,” he said.
The “pretty scary” part is where the consequences in medicine come into focus, particularly regarding the use of ChatGPT and other generative or predictive text technologies for medical advice.
“These systems come with a thin disclaimer to not use it for medical advice, but clearly both patients and clinicians are already using these tools for such purposes, even when the systems can be confidently misleading or erroneous,” Chen said.
The implications of this represent a clear and present danger, both in the clinic and in medical education, he added.
“The entire nature of rheumatology medical education has been upended overnight, with a widely available public tool that can more than pass medical licensing and complex reasoning exams intended to assess the competency of human practitioners,” Chen said. “We need to completely revamp our approach to continuing medical education in terms of the skills trained and knowledge assessed. The invention of internet searches made memorization of minutia a largely useless skill, while defining search terms and rapidly parsing article summaries became a critical everyday ability.”
A further point regarding medical education is that the data availability in a program like ChatGPT has a cut-off point of around 2021, according to Calabrese.
“When you put in a question about the COVID vaccine, for example, you get nothing useful,” he said. “So much has changed since then.”
The danger here is that many people across industries are using this technology under the assumption that the information for the topic at hand is accurate and timely. The reality that users may be receiving outdated information could obviously have significant implications if the use is in patient care or management.
It is for this reason that most experts believe that the ideal use of AI in medical care is in concert with a live physician.
“It is important to note that AI will not be able to replace the critical thinking, expertise and experience of rheumatologists,” Verhoeven and colleagues wrote in their paper. “While AI may be able to provide a wealth of information, it is the rheumatologists who must interpret this data.”
Although this seems reasonable, a growing body of data is beginning to demonstrate how rheumatologists may be able to apply AI for the all-important purpose of combating the specialty’s looming workforce shortage.
The Benefits of a ‘Software-driven’ System
In a study published in July 2022 in Arthritis & Rheumatology, Seppen and colleagues randomly assigned patients with rheumatoid arthritis in low disease activity to receive either usual care or a smartphone automated intervention. Results at 12 months demonstrated that the patients in the smartphone group experienced non-inferior outcomes vs. the usual care group in terms of change in DAS28-ESR.
In addition — and perhaps more importantly — the app yielded a 38% reduction in live consultations with a rheumatologist for this cohort of patients.
“They were treated online, rather than having face-to-face consultations,” Jacob M. van Laar, MD, of the department of rheumatology and clinical immunology at the University Medical Center Utrecht, in the Netherlands, who was not among the researchers in the Seppen study, told Healio Rheumatology. “To my knowledge, this was the first study that demonstrated the potential benefits of a software-driven health care system in this setting.”
To say that a machine predicting flare risk in patients with RA could save time for rheumatologists is an understatement. However, that is not the issue, according to van Laar. The issue is that this technology needs to be “optimized and refined,” he said.
“We need tools to help doctors focus on patients who really need treatment immediately,” he added.
For van Laar, the “beauty” of machine learning systems is that they can learn from experience. This means that as more information is fed into the system, it can, in effect, improve and refine itself.
As these processes continue, the goal for physicians and patients alike will not necessarily be to eliminate or minimize the use of this technology. The aim will be to benefit both doctors and patients with as few compromises and disadvantages to patient care as possible.
LeTang, meanwhile, suggested thinking beyond clinical factors for tech applications that may benefit workforce issues.
“AI can help with some of the lower hanging fruit, like scheduling of appointments,” he said.
However, using it to triage patients before appointments can also be useful, LeTang added.
“There is so much basic, initial information we need to get from patients before a visit,” he said. “As a nurse walking into a room, I can take a quick glance and see that the patient has already answered a number of basic questions about their general health. It gets the process started.”
In fact, several health systems are already using AI for these purposes, which LeTang believes is an area ripe for “next-level thinking.”
“What hospitals need to do is use these technologies as recruitment and retention tools,” he said.
By showing potential applicants and new or even long-standing hires that there is a commitment to reducing the administrative burden for their providers, health systems that use tech in this way could be attractive workplaces.
Ambient listening — an application that can listen to a conversation between a provider and patient and transcribe what is said — is another tech advancement that could attract providers to a clinic or health system.
“As doctors and nurses, we spend so much time in our visit staring at the computer screen rather than focusing on the patient,” LeTang said. “Nurses, in particular, have never really had access to dictation or scribe services.”
If the provider knows that the ambient listening technology is silently documenting the conversation, they can then provide patient-centered care by minimizing the amount of time they spend typing.
However, the current and eventual application of all these technologies raises yet another question surrounding AI in medicine, which pertains to the optimal balance between digital and human interaction in the patient experience.
Patients in the ‘Driver’s Seat’
“As long as patients are well-informed and instructed how to use it, many of them may like using AI or smartphone applications to guide their care,” van Laar said. “It can make them feel like they are in the driver’s seat.”
However, many patients currently remain in the dark about the extent to which their personal information is processed by machines. This is just one of the many ethical considerations surrounding tech in medicine that must be sorted out by physicians, regulators, lawmakers and patient advocacy groups.
That said, van Laar acknowledged that for patients to feel truly comfortable receiving their health care from AI technology, they may desire to have a doctor “always there in the background” in case of emergency.
“Just as companies use AI chatbots for customer service, health systems may begin to use language models to facilitate patient communication,” Li and colleagues wrote in their paper.
Convenience is also a factor for patients, according to Li and colleagues, who noted that many patients may prefer to communicate 24/7 with an AI chatbot rather than wait months for an appointment.
“With a massive imbalance between the supply of trained medical specialists like rheumatologists and the patients who need their expertise, it is entirely likely that patients will reach for imperfect medical advice from automated systems with 24/7 availability, rather than waiting months for an appointment with a human expert,” Chen said.
Regarding communication, AI may also have use in eliminating language barriers in the clinic, a problem voiced by many patients. Ideally, a program that can translate the spoken word immediately and on-site could indeed make linguistic miscommunication in the clinic a thing of the past. However, the information currently being fed into machine learning systems is largely in English, which may inhibit the ability to achieve this goal.
A companion issue is that there are complex cultural and linguistic differences even between countries that speak the same language. A machine that can understand and interpret information in English in the United States may arrive at a different — and potentially incorrect — conclusion in another country, even a majority English speaking one, based on nuanced differences in language and expression.
Regardless of the potential pitfalls, though, Chen argued that recent technological advances could be a way to address a singularly critical issue facing health care today — the overall lack of access to specialty care throughout the United States and the world.
“Tens of millions in the United States alone — let alone billions worldwide — have deficient access to medical specialists like rheumatologists,” he said. “Increasing automation and scalability through AI systems is one of the few ways to reach people who need medical care and advice.”
Providing ‘Far More Than Words’
It is important to understand that technology, ideally, is not replacing an actual rheumatologist, according to Chen.
“Ideally, every patient would have a rheumatologist to personally interpret and counsel them through their medical conditions,” he said. “A human clinician backed by the knowledge base and processing power of AI systems will only be better.”
That said, a growing body of research is beginning to demonstrate that AI might be catching up to the services provided by actual doctors, at least in terms of answers to patient questions.
In a paper entitled “Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum” — published in April in JAMA Internal Medicine — Ayers and colleagues randomly drew 195 patient questions from a social media forum. Both physicians and chatbots responded to the inquiries. A team of licensed health care professionals reviewed both sets of responses and determined that the chatbot responses were superior to the human responses in terms of both quality and empathy.
Chen suggested that a “common gut reaction” from physicians about such findings is to ask whether AI is going to take their jobs.
“The short answer is no,” he said.
However, for his slightly longer answer, Chen quoted clinical informatics pioneer Warner V. Slack, MD: “Any doctor who can be replaced by a computer deserves to be replaced by a computer.”
That said, the issue is never so simple, he added.
“Given that patients are stuck on multi-month wait lists to be seen in clinic and the overwhelming amount of work we are already trying to do to keep up with patient care, I wish technology would do my job for me,” Chen said. “With unlimited demand for medical care, there will always be a need for qualified professionals.”
For Calabrese, the meaning of “qualified professional” centers largely on the concept of “empathic communication” in the clinic.
“It is important to understand that communication is far more than words,” he said, noting cues such as tears in a patient’s eyes, or a hint of despair in their voice.
“Everyone wants a hug or someone to hold their hand at some point in time,” Calabrese added. “AI cannot duplicate that and should not supplant it.”
- References:
- Ayers JW, et al. JAMA Intern Med. 2023; doi:10.1001/jamainternmed.2023.1838.
- Li R, et al. JAMA Intern Med. 2023;doi:10.1001/jamainternmed.2023.1835.
- McMaster C, et al. Arthritis Rheumatol. 2022;doi:10.1002/art.42296.
- Seppen B, et al. Arthritis Rheumatol. 2022;doi:10.1002/art.42292.
- Verhoeven F, et al. Ann Rheum Dis. 2023;doi:10.1136/ard-2023-223936.
- For more information:
- Leonard H. Calabrese, DO, can be reached at 9500 Euclid Ave., Columbus, OH, 44195; email: calabrl@ccf.org.
- Jonathan H. Chen, MD, PhD, can be reached at 300 Pasteur Dr., Stanford, CA 94305; email: jonc101@stanford.edu.
- Michael LeTang, MS, RN-BC, can be reached at 111 W. 33rd St. 8th FL, New York, NY 10120; email: ali.moy@zenogroup.com.
- Jacob M. van Laar, MD, can be reached at Heidelberglaan 100, 3584 CX Utrecht, Netherlands; email: j.m.vanLaar@umcutrecht.nl.
Collapse
Read more about
Click Here to Manage Email Alerts