Note: This article was written by a real person.
If you ask the artificial intelligence text generator called ChatGPT how it can help in medicine, it will answer you. “ChatGPT can be a valuable tool in various medical applications,” before providing a 10-point, and fairly detailed, explanation of its practical uses in health care. (Penn’s David Asch, MD, asked it this exact question.)
But rather than taking ChatGPT’s word for it, some researchers at Penn, like Samiran Mukherjee, MBBS, chief fellow in Gastroenterology at the Perelman School of Medicine at the University of Pennsylvania, are studying it.
“I don’t think artificial intelligence technology is going away,” said Mukherjee. “It behooves us to understand both how medical professionals can use it to support their work and how patients may choose to interact with it. I’m excited by the potential of artificial intelligence, and it could ultimately mean more accurate and efficient treatment for patients. Humanity can use AI as a tool just as it’s used cutting edge machines and technology for generations.”
An Educational Tool
Mukherjee’s latest research, published in Gastro Hep Advances, and a good place to start in describing the potential of ChatGPT and other conversational or interactive AI tools, centered around education. Mukherjee and coauthors including Michael L. Kochman, MD, the Wilmott Family Professor in the Division of Gastroenterology, are gastroenterologists by training. They understand the importance of regular colonoscopies and colon-cancer screening and wished to explore if ChatGPT could tell a patient if they needed a colonoscopy taking into account their age, their latest colonoscopy, and other health data. They decided to pose colonoscopy questions to ChatGPT and found that the software was able to apply medical standards to indicate many of the instances when a colonoscopy is recommended.
“My colleagues and I did not want to test ChatGPT to see if it could replace counseling from clinical experts but rather assess if it could be a useful tool to provide information to patients regarding cancer screening and prevention,” said Mukherjee. “It’s safe to assume that patients are going to ask ChatGPT questions about their health in the same way they ask Google or look things up on sites like WebMD. If we know if there are any inaccuracies in answers or pitfalls when asking medical questions to ChatGPT, we can mitigate that by coaching patients on effectively using AI. Physicians and researchers can also team up with software developers to make design improvements.”
In the study, questions ranged from easier questions like “When should colon screening be repeated in a patient with a quality colonoscopy?” to medically complex ones like “What is the recommended interval for the surveillance endoscopy if the baseline finding revealed an adenoma with tubulovillous/villous histology?” Individual clinicians who specialize in gastroenterology then evaluated each response for accuracy.
The software specifically struggled when it came to more nuanced questions, such as a two-part question where the second part involved carrying forward the information presented in the first portion of the question. It defaulted to answering the second portion in a more generic manner. Nevertheless, no answers given by ChatGPT were entirely wrong.
Mukherjee and Kochman say that as ChatGPT learns from itself, that issue will likely become less significant, and they foresee people asking ChatGPT all sorts of questions about their health and receiving answers based on the latest-available data. It’s just not there yet.
“Since this technology is still in its infancy, those who use ChatGPT to ask medical advice should always consult with their doctor,” said Mukherjee.
As generations age and technology helps us live longer, more people continue to require medical care. In the United States, access to clinicians is not a huge issue. But long appointment wait times for specific specialty care, nurse shortages nationwide, and remote locales make it difficult for some patients to receive efficient medical care. Kochman said AI chatbots might be a solution.
“AI may help patients identify conditions where immediate medical care is called for and conversely may allow patients to try simple home and over-the-counter remedies for less concerning or dangerous conditions,” said Kochman. “All that frees medical resources for true emergencies and urgent conditions.”
There are still things that artificial intelligence can’t do…at least not yet. The issues: critical thinking and a different kind of intelligence – emotional intelligence.
“AI is not able to digest and prioritize conflicting guidelines and decide upon relevant issues such as the quality of the evidence and the strength of the recommendations,” said Kochman. “In addition, the words that a chatbot reads that patients type into a form are much different than a physician’s seeing a patient, watching their body language and hearing the words and emotion with which they are voiced. These are all critical inputs for a seasoned clinician which chatbots are not able to integrate and consider when formulating answers.”
For now, when an urgent problem arises, paging a human doctor instead of Dr. ChatGPT is the best choice.
“As much as I love AI, real people possess an innate gift,” said Mukherjee. “Humans are and always will be problem solvers.”