"The ultimate goal is to have as much of our time spent on actual patient care and complex decision-making [as possible]," says David Canes, MD.
In this video, David Canes, MD, discusses his recent Journal of Urology paper, “Harnessing generative artificial intelligence to improve efficiency among urologists: Welcome ChatGPT.” Canes is a practicing urologist at Lahey Hospital and Medical Center in Burlington, Massachusetts.
So, email is a good one. Let's talk about that for a second. One of the nice things about ChatGPT is it remembers what it just gave you, and you can ask it to alter it. So, you can respond 'take out the part about blah blah blah' or 'change the tone; make it more friendly and upbeat.’ You do have to be careful though, because as we discussed in our article, right now, generative AI is a very clever guesser and can give wrong answers and even hallucinate. So, because ChatGPT right now is a clever guesser, you have to proofread and double check the output. That's why what we're proposing is that this is good for low-stakes communications and outputs.
For example, you could, say, write a letter to an insurance company asking for Myrbetriq [mirabegron] to be covered for overactive bladder in a patient for whom Detrol has failed, and include references to the scientific literature. It'll do that, but then the references to the scientific literature may be completely fabricated. Like, totally made up. I'm not an expert on how the large language models are constructed, but from what I understand, the neural network is very good at guessing what the very next word should be, so it's a constant statistical likelihood of what the next best word is in a string of words. It will not say 'sorry, I don't know that.’ It knows what a reference looks like and can completely make it up. Incidentally, GPT4 just came out, and I haven't used it yet, but apparently the hallucination aspect has gone down substantially, from what I hear.
Apart from emails and letters, this is really good for brainstorming. For example, you're giving a talk. You put in a prompt, 'I'm giving a talk on blah, blah, blah,' can you give me 10 proposed engaging titles for this talk and learning objectives for CME? I've done this, and the output is fantastic. I've used titles for talks that directly came out of ChatGPT. If you're facing a blank page, and you have to come up with something from scratch, this is something that that you ought to consider.
I have also used it in the context of EHR documentation. For example, I spend time in the exam room with the patient explaining—let's just use a basic example—explaining TURP [transurethral resection of the prostate] for BPH [benign prostatic hyperplasia]. That's the part of doctoring that we all really enjoy, the part where we're sitting face to-face with another human being, not the documenting of that experience. We don't like that. I've experimented with saying, 'Give me a blurb for medical record, reviewing a discussion of TURP for BPH.’ What's really interesting is now the skill becomes the prompt. In other words, composing the prompt in a way that gets you the output that you desire. For example, if you give the prompt that I just told you, sometimes you have to say 'no, have it be in the third person.’ It'll put out a letter, and you say, 'no, not a letter.’ You get good at crafting the prompt that you're looking for.
Now, Doximity has come out with, I will call it a ChatGPT wrapped up in an outer wrapping that doctors might find interesting. They have created some prompts in the background that can help you generate these types of things that I'm telling you about. I think we're going to see more of that moving forward.
Another example is answering In Basket messages from patients. Patient asks a question, you can put that into ChatGPT, without putting in HIPAA sensitive [information], you don't want to put any patient information in. [Let's] say the patient has a [urinary tract infection] and is still having irritative symptoms 1 day after starting antibiotics. You could say 'explain to a patient, at an 8th grade reading level, why having irritative voiding symptoms 24 hours after starting antibiotics for a urinary tract infection is expected.’ You'd be amazed when it comes up with. A nice, long, friendly, easy-to-understand few paragraphs that, if appropriate, you can lightly edit and copy and paste right into your EHR. If you're really looking out for your future self, you create one of those and it's so good, you turn it into a template or a dot phrase so that you can do that again the next time a patient asks a similar question.
What fascinates me about all this is that this is very early days. We're going to look back on this 5 years from now and think 'that was really cute, what we were doing.’ As it becomes baked into our software that we're already using, the tools that we're already using, I think the result, I hope, it will be lightening the load for doctors. The ultimate goal, after all, is to have as much of our time spent on actual patient care and complex decision-making. I think for most of us, we would really enjoy our jobs fully if that's predominantly how we spent our time. The more we can use AI to streamline all the less attractive aspects of our job, the better.
This transcription has been edited for clarity.