GThe feisty AI is having its moment. Peter Lee, who oversees Microsoft’s approach to healthcare and co-author of the 2023 book The AI Revolution in Medicine: GPT-4 and Beyond, calls this ” We call it the most innovative tool ever developed.
Microsoft has been working with an organization called OpenAI for several years, so when OpenAI released ChatGPT in November 2022, Lee immediately realized the need to educate the medical community about the technology.
“I started receiving emails from doctor friends all over the world, all saying more or less the same thing: ‘Wow, Peter, this is a great product and I’m using it in my clinic for this purpose. ‘We’re doing it,'” Lee said. Status “That was scary. …This has led us to take a deep dive into the benefits and limitations of these models in the medical field, and to provide education, including an entire book written expressly for doctors and nurses.” I started a long journey of writing material.”
Lee, one of 50 influential people named to the 2024 STATUS list, spoke with STAT to find out how the technology can help with tasks such as post-doctor visit summaries. He also spoke out about his concerns about its potential to perpetuate prejudice.
What do you think is beautiful about computer science?
A big part of my mind, and a lot of computer scientists’ minds, is about generative AI. It’s a collaboration with OpenAI. This big AI model is trained to do only one thing: predict the next word in a conversation. That’s all that is optimized.that [has] It’s literally not optimized for anything else.
Is it something like ChatGPT in this case?
That’s it! But to do it really well, machine learning systems needed to discover how to do arithmetic operations. If you want to choose the next perfect word in a conversation, and someone says 2+2 = ‘blank’, the best way to answer that question is to actually find out how to calculate addition. Or you can also predict the next word if the last sentence of the conversation is “…and the murderer is blank”. To be able to solve such problems, you must be able to read the entire murder mystery and follow all the steps. Deductive logic that does that simple thing in predicting the next word.
What’s so beautiful is that we think of next word prediction as just a trivial thing, like autocomplete on an iPhone. But when you do it on this huge, unimaginable scale, machine learning systems have to self-discover how to do all this thinking, and there’s something about that that I find incredibly beautiful. And that further maps to medical and healthcare.
Why?
Let’s have a conversation: You are my patient. I’m a doctor. We talk back and forth. I signed you up for the lab and received the lab report. Then the last sentence of the conversation is “…and the diagnosis is “blank.” ” To best predict the next word in a conversation, this means machine learning needs to: Study medicine. Isn’t that wild? I think that’s the most amazing, amazing, and beautiful thing in the world today.
What are the big concerns regarding generative AI used in the medical field?
The main thing I’ve been trying to teach doctors and nurses is that if a computer’s mental model is a machine that can recall perfect memories and perform perfect calculations, then the most important things to understand about generative AI are: That’s not true. Computer. It is a reasoning engine or thinking machine, but it also has the same limitations as the human brain.
If you memorize something and ask them to regurgitate it, they may hallucinate because they won’t remember it completely. When you ask them to solve large, complex math problems, they can make mistakes just like humans. At the same time, you can make statements about connections between concepts in an incredibly sophisticated way.
Sticking with this theme of finding beauty in seemingly complex things, what is so beautiful about GPT-4 and its applications in medicine?
There are probably two things going on. Most surprising is his GPT-4’s ability to understand what psychologists call “theory of mind.”
One use of GPT-4 by Epic is to help doctors write “post-visit summaries” for patients. So you’re my patient, you come see me, and then you’re sent home. You will then need to send an email with instructions on how to care for yourself after your treatment. It is difficult for doctors to write post-consultation summaries because they must be accurate. You need to access four to five parts of your electronic medical record to see if you can pick up your prescriptions or do anything else you need to do at home. If you do this incorrectly, there is a high possibility that you will be sued for medical malpractice.
Epic integrates GPT-4 to capture all your information and create notes for your doctor to review before sending. In early tests, these are controlled clinical studies, and patients rate emails written in GPT-4 as more human-like than emails written by doctors. It’s not that AI is any more human-like than a doctor. Obviously the opposite is true. But AI has an indefatigable ability to find it. [personal information from] All health records and all conversation records [and then] Try adding a little line like, “Congratulations on becoming a grandma and grandpa,” or “Good luck with your daughter’s wedding next month in Maine.” These special touches, where AI adds a personal touch, can actually make a meaningful difference to the patient experience.
The second is the general intelligence of GPT-4. It turns out that you can triangulate multiple disciplines when trying to make a medical diagnosis. So if you come to me feeling anemic, that could be a problem for an endocrinologist, a cardiologist, a nephrologist, or a psychologist. Depending on which doctor you see, or which specialist your primary care physician refers you to, you will receive a different diagnosis. GPT-4 allows you to see your condition, test results, and initial presentation all at once. What we consistently see is that doing so allows for a more comprehensive assessment.
That’s fascinating. So can generative AI help make human doctors more human?
That’s a very good question. We talk about humans nudging AI, but there are also times when AI nudges humans to take a step back and rethink a potentially difficult situation just a little bit. Let’s put ourselves in the doctor’s shoes. [Epic using GPT-4] Finally, suggest a note that says, “Congratulations to your son’s high school basketball team for winning the championship!” In fact, the doctor will read the draft for perhaps three seconds longer, reflecting on the patient’s life. This is what I call a “reverse prompt.”
It sounds beautiful, eerie, wonderful, and creepy at the same time.
As we move toward this future of generative AI in clinics and operating rooms, how do we address privacy and bias concerns?
I don’t think privacy is a problem. Microsoft’s OpenAI service on Azure cloud provides HIPAA compliance. We’re very proud of it, but our cloud is nothing special. AWS and Google Cloud offer similar compliance to enterprise customers of these clouds. It’s different from the consumer space. Privacy guarantees are less stringent when using consumer products such as Google Search or ChatGPT. However, any healthcare organization that subscribes to Microsoft Azure can achieve HIPAA compliance.
Bias, this is a serious and potentially devastating problem. Because, in my view, these models are hopelessly biased. They are hopelessly prejudiced because they have learned from us, and just as humans are hopelessly prejudiced, so too are they hopelessly prejudiced. They are prejudiced. Some of my computer scientist colleagues believe they can solve these problems, but I don’t believe that. What you need to understand is that while these AI systems are hopelessly biased, they also understand the concept of bias and why it’s bad.
One of the things we found in our research is that if you describe a situation involving decision-making about a patient and ask it to pass it to GPT-4 to check for bias, humans are much better at identifying biased decisions. This means that you can demonstrate your ability to exceed. . In the New England Journal of Medicine experiment, submitted manuscripts were read by her GPT-4. Consistently, GPT-4 was able to find non-inclusive language and bias in manuscripts that escaped the attention of human reviewers. Generative AI could become one of the most powerful tools for combating bias, even if the tools themselves are just as prone to biased decision-making as humans. [are prone to].
I don’t trust an AI system to make the decision on its own whether my health insurance claim should be reimbursed. I think humans should do that. But I want a generative AI system to act as a second set of eyes to check whether the person making my insurance decisions is biased against me. Because I think AI systems are very good at that.
