If you’ve been to the doctor’s office in the past few years, there’s a decent chance that artificial intelligence was used in some part of that appointment.
Increasingly, many patients are also turning to large language models for answers to their healthcare questions. About a third of patients say they’ve used artificial intelligence for health advice, according to the Kaiser Family Foundation, a nonprofit focused on national health issues.
But what is AI, and how could it help or hurt a patient’s health outcomes?
How AI works
Generative AI is newer to the public and makes predictions based on probability.
Dr. Shawn Murphy is a professor of neurology, biomedical informatics and medical education at the University of Washington.
He said generative AI uses the same kind of predictive modeling, on a larger scale, that’s been used for decades to provide message suggestions when texting. But unlike early predictive texts, generative AIs — like large language models — have the entire internet to pull from.
“We train it with literally hundreds of billions of sentences from the internet, which are all people talking,” Murphy said. “So, because it learns from people talking that way, it can actually sound like a person.”
Use in medicine
Dr. Dushyant Sahani is a professor and chair of the Department of Radiology at the University of Washington.
In his specialty, AI tools are used in CT scans and MRIs to produce higher-quality images with less data.
That means that radiologists can get usable CT scans while exposing their patients to less radiation. It also means that patients getting an MRI don’t need to spend as long in the machine.
“Demand for MRIs is very, very high,” Sahani said. “Instead of spending 30 or 40 minutes, we can scan patient (in) less than 20 minutes, sometimes within 10 minutes.”
Across specialties, one of the most common ways providers use AI is automated note-taking. They’ll also use AI tools to help automate certain repetitive and bureaucratic tasks, Sahani said.
“(Providers) can spend that time improving patients' experience,” he said. “So, that is one area I'm pretty excited about.”
Some medically focused AI tools allow clinicians to ask questions that the program answers by using citations from peer-reviewed journals. Other AI tools can integrate existing patient data such as labs, physical exams, imaging and medications to provide recommended treatments.
Doctors don’t always have the ability to fully evaluate previous lab work or imaging for clues into new health issues that arise, Sahani said. But AI could change that.
“Most physicians are not able to integrate everything we have about patients,” he said. “Most are not able to see or really understand the value of these prior imaging. And I think sometimes … we might be able to integrate a lot of that.”
Patient use
AI can help patients learn about a condition without having to spend hours researching, Murphy said. It can also be used as a tool to start conversations with their doctor.
“It really is very helpful that way,” he said. “I'm a little bit cautious, though, when folks talk about using it just straight off the internet. Because it could get something like the dose of medication you should use wrong.”
Recently, OpenAI was sued by the family of a woman in Seattle who died after tapering off a life-saving epilepsy medication based on recommendations from the AI.
Some AI programs are specifically fine-tuned with training on medical literature, while other models are not.
Still, because generative AI is trained by the internet, even the best models can produce bad information or replicate existing biases.
A lot of “fine-tuning” goes into AI programming to try and eliminate those problems, Murphy said. Most healthcare-specific AI chats have additional “agents,” which are used to try and catch those mistakes before they get to a user. Essentially, another AI double-checks the first AI.
“They kinda check and cross-check and look things up to make sure that what it said was true,” he said. “But despite all of that, things get through.”
Patients who rely too heavily on AI tools also risk causing themselves undue stress.
Because uncommon medical cases tend to be written about on the internet, the training data AI models use biases them toward rare cases, Murphy said. A large language model might tell a user, for instance, that their hip pain is due to a bone tumor when the more likely cause is sciatica.
Developers have also tried creating AIs that are only trained on medical information, but those tend to not work as well.
“It's as though you never got (into) elementary school or high school, and you went right into your Ph.D.,” he said. “Even as a human, that's not a very good idea.”
Privacy
Accuracy isn’t the only concern for patients using AI programs. Depending on the situation, talking to a chatbot can be a bit like talking to a person with no filter.
“The privacy policies of what it's gonna keep to itself, so to speak, and what it's gonna actually use to train itself (are) a little bit difficult to navigate,” Murphy said.
At medical centers, such as those managed by the University of Washington, legally binding agreements are put in place to ensure patient privacy.
“Not so true with things on the internet,” Murphy said. “Often, their whole model is that they wanna use this data to help train themselves.”
Even if companies try to protect patient privacy, Murphy said, that process isn’t always simple.
“It’s a work in progress,” he said. “You’re like, ‘Well, I should just be able to take that model and use it over here.’ But what you don't understand sometimes is how all those numbers actually contain private information.”
How it’s changing medicine
Generative AI tools have “tremendous potential” in healthcare, Sahani said, but he also said it’s important to not become overly reliant on them.
Similar to predictive text on a phone, Sahani said, AI tools still require human oversight.
“There's an autocorrect feature (on my phone), and sometime if I don't read it properly, I realize, ‘My God, how did I make that mistake?’ It can be funny. Sometimes, it can be scary. I think that can happen if we feel that AI is doing such a great job that we lower our guard,” Sahani said.
Murphy said that as students train, it’s important they’re trained in foundational skills before layering AI into their practice responsibly.
“(Students) need to get the same foundation that they had before AI,” he said.
In addition to education on responsible AI use, Sahani said, relegating more basic tasks to AI will likely change what clinicians need to focus on.
“Teaching (students) in better bedside manners, patient communication, critical thinking, how we provide consults, how we improve patient's experience — all those things will be important,” he said. “There are many tasks AI might be able to do for us. But that human element, we need to focus more on teaching that.”
