If you were to ask 10 healthcare providers for their opinion of the role of generative artificial intelligence in medicine, you would get 10 different answers. Some view the rapidly evolving technology as too error-prone to be of any use in diagnosing and treating patients, or as a contributor to harmful biases in medicine. Others see great potential for the technology to benefit doctors, patients and the healthcare system as a whole.
Robert Pearl, MD, considers himself a member of the optimists’ club. In fact, he’s so enthusiastic about generative AI that he co-wrote his most recent book with ChatGPT—and even gave the chatbot credit on its cover.
That book, ChatGPT, MD, makes a strong case that the limitations of today’s generative AI tools will soon be overcome, and that the technology will become a powerful ally to health providers and their patients. Pearl, former CEO of The Permanente Medical Group and current clinical professor of plastic surgery at Stanford University School of Medicine, spoke recently with Cure about the risks and benefits of AI in medicine.
This conversation has been edited for length and clarity.
Cure: Where does AI have the most potential in healthcare in your opinion?
Pearl: In the short run, it will be used mainly for administrative tasks like populating electronic health records. But I believe that within three to five years it’s going to be empowering patients to better manage chronic diseases. About 60 percent of Americans have one or more chronic diseases such as diabetes, heart disease or asthma, and more than two-thirds of people on Medicare have two or more chronic diseases. They account for 75 percent of healthcare costs because of complications such as heart attacks and strokes.
We know that when we control hypertension, we prevent strokes. But it takes a lot of effort and patient oversight. Generative AI will make controlling chronic diseases easier. Dedicated clinicians—working with patients who are empowered by AI and using it effectively—will accomplish far more than either could alone.
How can AI support more productive doctor-patient relationships?
Take wearables, for example. They measure blood pressure, pulse, blood oxygen, blood glucose and heart rhythm exceedingly well. But patients can't figure out what the measurements mean, and their doctors don’t have time to sift through all of the information. A generative AI tool would be able to follow the patient’s progression and provide continuous feedback to them and their doctor.
Say I’m a physician who prescribes an antihypertensive drug to a patient in January, with a plan for follow-up visit three months later. I could tell the tool that I expect to see an improvement in four weeks. Then the wearable can track the patient’s blood pressure and, using Bluetooth, provide feedback to me that will tell me if the data are within my expectations. If in February I see the patient’s blood pressure is no better, then I know it’s time to either change the dose or the medicine. There’s no reason to wait three months.
You write about doctors being prone to bias in their diagnoses and treatments. How can AI help solve this?
We know that black patients don't get the same care as white patients. They get complex cardiac procedures far less often than white patients do and less pain medication after surgical procedures. AI could say something like, “Of the last 50 patients you treated, you would have normally prescribed more pain medication after a surgical procedure than you did.” There may be a good reason why you didn't do that, but at least it brings the discrepancy to your consciousness.
I also write in the book about how AI can help prevent confirmation bias, the tendency of physicians to make assumptions based on past experience. I write about a physician who did a complex neck surgery. At the end the patient’s vocal cords were not moving. Normally that patient would be put in the ICU or given a tracheostomy. Using ChatGPT, the doctor found several identical case reports, and in each case, it was determined that the local anesthetic used for the procedure blocked the nerves in the larynx. Within 30 minutes the anesthetic wore off and the problem was solved.
What are the biggest risks of using AI in healthcare?
The biggest challenge right now is the technology is known to hallucinate, so we need clinician oversight every time we use ChatGPT. There is also a risk that AI will spread misinformation. But I'm optimistic that because AI tools can access the totality of information that’s fed to them, they will gain the ability to recognize misinformation.
How can AI help bring down healthcare costs?
People often talk about value-based medical care, which is the goal of lowering the cost of healthcare by raising quality and making care more convenient and accessible. That is the holy grail. I think the way to get there is to keep people healthier. We can make healthcare more affordable by lowering the frequency of medical errors. We can lower the cost of healthcare by helping people manage their lifestyles so they don’t get chronic diseases. And if they do get those diseases, we'll be able to control them 90 percent of the time.