When your patient has puzzling symptoms, you might dig through the literature or turn to a colleague for ideas. But these days, you have another option: You can ask ChatGPT for help.
Artificial intelligence (AI) is inevitable, but also useful in the right circumstances, like for generating ideas about diagnoses and treatment. Generative AI tools — specifically, large language models like ChatGPT — can respond to your prompt with a detailed response within seconds.
But what’s the best way to use these bots in clinical practice? And how do you avoid falling prey to bad information that could harm your patient?
“GPT has been excellent at brainstorming, at giving a slew of ideas,” says Paul Testa, MD, chief medical information officer at NYU Langone Health in New York City. “It’s really up to the physician to critically review those ideas and see which ones are the best.”
How Many Clinicians Are Using ChatGPT?
Some 11% of clinical decisions are now assisted by generative AI tools, according to an Elsevier Health survey conducted in April and May of this year. That includes 16% of nurses’ decisions and 7% of physicians’ decisions. The practice also varies by region — it is more common in China and the Asia-Pacific area than it is in North America and Europe.
Overall, about 31% of physicians across the globe use AI in their practices, and 68% are excited about the future of AI in healthcare, according to market research firm Ipsos. Among AI’s most alluring prospects: the potential to automate repetitive tasks and to increase the efficiency and accuracy of diagnoses.
Recent research suggests that these tools are reasonably accurate, too. A study from Mass General Brigham showed that ChatGPT was 72% accurate in clinical decision-making across a range of medical specialties. In a pilot study from Jeroen Bosch Hospital in the Netherlands, ChatGPT performed as well as a trained doctor in suggesting likely diagnoses for emergency medicine patients.
However, that doesn’t mean AI is always accurate. AI is prone to “hallucinations” — false answers stated in an authoritative way.
Privacy is also a consideration. Many chatbots do not comply with the Health Insurance Portability and Accountability Act (HIPAA), except for proprietary versions that meet privacy and security standards. (Google is piloting a large language model — Med-PaLM 2 — tailored specifically for healthcare applications.)
Here’s how to navigate these concerns and get the best results with ChatGPT or other generative AI in your practice.
1. Ask your institution for help.
This is step one. As generative AI gains popularity, more healthcare systems are creating guidelines and other resources to help clinicians use it. For example, NYU Langone started preparing for the AI revolution more than 5 years ago, and now they have a secure, HIPAA-compliant system.
AI tools constantly process and learn from the information they’re fed. So never put protected patient information into a public version. “This was the first message that we sent out to our clinicians as GPT became popular,” says Jonathan Austrian, MD, associate chief medical information officer for inpatient informatics at NYU Langone. “Come to us; we have our safe, HIPAA-compliant GPT for that exact purpose.”
2. Use AI to broaden your perspective.
Generative AI can be especially helpful for odd cases where you want a broad differential diagnosis, says Steve Lee, MD, PhD, vice chair and associate professor of otolaryngology/head and neck surgery at Loma Linda University in Loma Linda, California. “As physicians, I think we’re quite good at figuring out the things that we see regularly,” says Lee. “But then there are these obscure things that we read about once in medical school 20 years ago and forgot about it. AI doesn’t forget.”
The chatbot might spit out a diagnosis you hadn’t thought of. You can also ask the bot to list the differential diagnoses in order of likelihood, or recommend diagnostic tests you may have overlooked.
3. Think like a prompt engineer.
There is an art and science to using generative AI, and tech mavens now aspire to be “prompt engineers,” or experts in prompting AI to give better answers.
“You have to condition the answer by asking the right question with certain assumptions,” says Samuel Cho, MD, chief of spine surgery at Mount Sinai West in New York City.
Be specific about your credentials when you ask ChatGPT a question, so that it generates an answer from medical texts and other sources with an appropriate level of complexity. For Cho, that means prefacing questions with, “Assume that I am a board-certified spine surgeon.”
4. Cross-check references.
Don’t assume that the answer you get is the final answer. Experts recommend asking ChatGPT to cite sources. When Lee did that, he noticed a concerning trend. “Sometimes it just made up papers that don’t exist,” he said.
The AI knows what academic references look like and can fabricate them. “It generated one that sounded like a real paper from a real journal, with volume numbers and all that stuff,” he said. “But when you actually go search for it, the publication did not exist.”
Other times, ChatGPT cites references that are spot on. The lesson: Check the source material to make sure it matches what the AI told you.
5. Gather solid information before you start prompting.
The quality of your patient exam will affect the quality of the AI’s response. ChatGPT’s “output is only as accurate as the input,” says Prathit Arun Kulkarni, MD, assistant professor of medicine/infectious disease at Baylor College of Medicine in Houston.
Let’s say you describe what you heard in your stethoscope and share some lab work and chest radiography results. If those observations or results are inaccurate, ChatGPT’s response will reflect those errors.
“That’s not necessarily a failure of GPT. It’s an inherent limitation,” says Kulkarni. “The accuracy of what is put in there is still determined by us.”
6. Beware of confirmation bias.
One downside of large language models like ChatGPT: They can be sycophantic. “It tends to agree with almost everything you say,” says Steef Kurstjens, PhD, a clinical chemist in training with Jeroen Bosch Hospital in the Netherlands. “If you push slightly in a certain direction, [it] will immediately agree with you.”
To minimize the risk for this error, avoid asking yes-or-no questions like “Does this patient have vasculitis?” Instead, ask the question in an open-ended, neutral way: “What do these symptoms suggest?”
7. Think of AI as just another tool.
The hype around AI can engender both interest and fear. But this technology is not meant to replace healthcare professionals but augment their work.
“In healthcare, we use all kinds of software scoring systems, all kinds of devices which support physicians in their decision making,” says Kurstjens. Does the patient need to be admitted, or transported to intensive care? Do more tests need to be ordered? You already use tools to aid these decisions, and ChatGPT could be another to add to the mix.
For more news, follow Medscape on Facebook, X (formerly known as Twitter), Instagram, and YouTube
Source: Read Full Article