Artificial intelligence is already in wide use in health care: medical workers use it to record patient interactions and add notes to medical records; some hospitals use it to read radiology images, or to predict how long a patient may need to be in intensive care.
But some hospitals have begun to contemplate using a new phase of AI that is much more advanced and could have a profound effect on their operations, and possibly even clinical care.
Indeed, never one for modesty, ChatGPT, one form of the new AI technology that can render answers to queries in astonishing depth (if dubious accuracy), called its own role in the future of medicine a “groundbreaking development poised to reshape the medical landscape.”
What that medical landscape will look like is just starting to come into focus. Hospitals, doctors, and medical companies are experimenting with different facets of its application, and even hiring staff dedicated to using state-of-the art AI language technology in what could become a pivotal moment in medicine.
“There’s been many times that we have a potentially disruptive technology that hasn’t made it,” said Dr. Isaac Kohane, professor of biomedical informatics at Harvard Medical School, and editor-in-chief of a new journal the publisher of the New England Journal of Medicine started to cover just AI. Large language models like ChatGPT may be different, he said, because the rapid adoption of the technology is pressuring the health care system to follow along.
At its foundation, “large language model” AI uses powerful analytics to convincingly generate human-like text. The latest versions of the technology behind platforms such as ChatGPT, GPT-4, and Google’s Bard have produced some astonishing results because they have been trained on enormous data sets and are run by more powerful computers.
While there have been some dire warnings issued about the threat the technology poses, hospitals see it as a useful tool to streamline administrative tasks and, perhaps one day, improve patient care. But they are moving slowly and methodically, testing the platform mostly with administrative tasks.
Boston Children’s Hospital recently posted an opening for an AI prompt engineer, to help integrate large language technology into the hospital’s operations.
Like other hospitals, Boston Children’s already makes broad use of older types of artificial intelligence, such as to read radiology and pathology scans and to make predictions about patients’ length of stay in the ICU and emergency departments.
John Brownstein, the hospital’s chief innovation officer, said the new role will build on that foundation. In the short term, the hospital hopes to implement such tools on the administrative and financial sides and is soliciting feedback from staff on where AI tools might help people. Brownstein envisions using the technology to manage licensing requirements for providers, or to cull resumes of job candidates. On the research side, such technology can be used to read and summarize scientific literature.
“From my perspective, it will be as fundamental a tool as a search engine and mobile phone, and we have to prepare people for that world,” Brownstein said. “It’s going to be hard in the future to be effective at your job without taking advantage of these technologies.”
However, he said, medical professionals are being thoughtful and deliberate about bringing the tools to patient care. There are concerns about patient privacy, consent, and bias. He said the technology also has a tendency to “hallucinate,” or say things that are not true.
Given those concerns, experts are considering guardrails for large language models. But one day, the technology could be used to navigate a person’s electronic medical record, summarize a medical history, and even potentially generate clinical and discharge notes. Writing e-mails, which clinicians have to do on a daily basis, may become more efficient with the use of AI.
“The holy grail is using these tools to optimize and provide the best possible clinical support to our patients,” Brownstein said. “But it’s also the place we have to take the most time because we have to be the most thoughtful.”
Elsewhere, Microsoft and electronic medical record software maker Epic are working to integrate large language model technology into electronic medical records. For example, UC San Diego Health, UW Health in Madison, Wisconsin, and Stanford Health Care have begun using it to draft messages to patients.
Dr. Sahil Mehta, a physician in Boston specializing in vascular and interventional radiology, uses ChatGPT with the administrative functions of his job. He just used it to help write an appeal of an insurance denial.
“While it needed some minor editing, this type of document easily could take 20-30 minutes to write, yet ChatGPT wrote it in a few seconds and required just minor changes,” Mehta said. “As this technology evolves, and is better integrated into workflows, it has the ability to significantly improve physician morale, documentation, and patient care because we will spend more time with patients, less with paperwork.”
According to a survey of 500 health care professionals conducted in April by digital health platform Tebra, more than 10 percent were using AI. Another 50 percent expressed an interest in using the technology in the future.
Perhaps the most immediate implementation of the technology has been in medical school education. Mehta, who founded MedSchoolCoach, an educational technology company focused on pre-med and medical students, said students can use the technology as a brainstorming tool for medical school essays. MedSchoolCoach is also creating study aids using large language model technology that has been trained on medical textbooks and other expert-written texts.
Meanwhile, patients have already begun turning to AI. The Tebra survey, which also asked 1,000 Americans about the use of the technology, found that more than 5 percent of those surveyed had used ChatGPT to help diagnose a problem and had followed its advice. Additionally, a quarter of those surveyed would not visit a health care provider who refused to embrace AI technology.
Early evidence shows ChatGPT can accurately answer some medical questions. A recent study by physicians at Massachusetts General Hospital and Taipei Medical University Shuang Ho Hospital, in Taiwan showed that the free version of the technology successfully answered common questions about colonoscopy prep. Most trained physicians couldn’t tell which responses were from the bot and which were taken from hospital websites.
Whether the technology can accurately answer more complicated questions is unclear, and will be the focus of future research, said Dr. Braden Kuo, a neurogastroenterologist and director of the Center for Neurointestinal Health at MGH, who worked on the study. Some research has already shown the chatbot is more empathetic than physicians when answering medical questions.
For his part, Kohane said he will be disappointed if hospitals don’t figure out how to use the technology to improve care, a vision that includes alleviating administrative burdens and allowing physicians to spend more time with patients.
While the larger existential risks are worth considering, he said, the more immediate concern is making sure health systems use AI to improve patient care and not simply to boost margins.
With the wrong goals, the technology could be problematic, squeezing ever more patient visits with every provider. Patients might be confronted with an empathetic robot when reaching out with questions or concerns, further fracturing the relationship patients have with their physicians.
The fact that the technology has been found to makes things up means it is concerning for use in patient clinical data.
“We need a public discussion about what we are going to do with this technology,” Kohane said. “It can be used to improve not just the doctor experience but the patient one, and reestablish the relationship that has been incredibly frayed over the last couple of decades. The technology is giving us the opportunity.”