Discover more from Till the Cavalry Arrive
The Time-Saving Magic of Chat GPT for Doctors
Leveraging AI to improve efficiency is just the beginning of a healthcare revolution
Like many of you, I’ve been in awe as Chat-GPT broke the internet this week.
The frenzy began after this tweet
I’ve been playing around with it, and scouring Twitter for the best examples. In this week’s post, I’d like to categorize and show you some of the best early use cases. Then, postulate where this all might be headed.
Thanks for reading Till the Cavalry Arrive! Subscribe for free to receive new posts and support the series.
The best part about this is recapturing time. Time is our most valuable asset (besides our health). The crux of the burnout problem, in my view, is that doctors have lost control over their time.
They face a packed clinic schedule, and as they grab the exam room doorknob, they think “how long will this take?” When it goes as planned, great. When you get derailed for any reason, the train falls off the track.
If any task, patient, phone call, or complication unexpectedly steals time from you, you’re forced to borrow it from somewhere else. The next patient’s time? Your lunchtime? Your precious personal time?
What is Chat GPT?
Let’s get this out of the way first. You may have seen the hype, but still, you have no idea what anyone is talking about.
It’s a free artificial intelligence chat tool. Type a prompt, and it spits out what seems like magic. I prompted: “What is Chat GTP in 4 sentences” and it gave me this:
Chat GTP is a technology that enables natural language conversations between humans and computers. It uses artificial intelligence and natural language processing to understand user input and respond accordingly. 3. Chat GTP can be used in a variety of industries, from customer service to healthcare, to automate tasks and help speed up workflow. Chat GTP is becoming increasingly popular as it provides a more natural and intuitive way for people to interact with machines.
Here’s how Chat GPT can potentially help in your day-to-day job right now.
Answering patient questions, providing reassurance
Your inbasket is necessary, but it’s a time sinkhole. In a perfect world, you would answer every patient question thoroughly, and with empathy.
So far, it seems like the GPT model has empathy, at least on the surface. An early recovery patient after prostatectomy was concerned (of course!) about incontinence. It showed up as a patient message.
Pay attention to the prompt.
Is it perfect? No. But we’d all rather be editors than authors. I’m happy to copy that, make some tweaks, and send it.
Here’s another great example from Twitter, where a physician was trying to reassure a patient reluctant to take statins because of side effects.
The magic is really in the prompt, as all of these examples will slowly reveal. Her prompt was expertly crafted. You can click on the tweet above to read the actual letter that was generated. It’s far better than what I could come up with quickly (and not just because I’m a urologist).
As the doctor points out, the time saved is a gift.
Creating Blurbs for a plan, or a Smartphrase
Blurbs, blurbs, and more blurbs. The end of a clinic note - the assessment and plan, really is a blurb, isn’t it? I mean, it’s not supposed to be a master’s dissertation. I experimented with creating these on the fly during clinic, and it was magical.
For instance, a patient with GG1 prostate cancer on surveillance, stable PSA, with a distant biopsy plan. I fed it a prompt, but it spits out a letter. Keep in mind this is an interactive engine, so you can reply with alterations and it knows what you mean.
You can see I objected “not as a letter” but “as a summary statement in a doctor's progress note.” Boom. That’s what I wanted. Done. I’m still experimenting here, so you might argue not much time was saved there. Work in progress.
Ultimately, I got the hang of it for a bladder cancer patient. My prompt became more specific. I augmented this by asking it (really testing it) to see if it knew that cytology wasn’t necessary. It did figure that out correctly.
In this case, I might grab something I like here and make it immortal by turning it into a SmartPhrase (dot phrase) for future me.
Refreshing your stale SmartPhrases
Refreshing stale smartphrases in electronic health records is important to ensure accuracy and up-to-date information. This helps improve the quality of patient care and reduces the risk of medical errors.
But who has time to refresh stale SmartPhrases? Not me. I put a lot of work into them for the rollout back in 2016. Since then, not as much. Could Chat GPT help me refresh old Smartphrases?
I did an experiment with a dot phrase that I use at the end of my visits for prostate cancer on active surveillance.
Not earth-shattering, but again the quality of the output depends on the prompt, and I’m still learning. If I put in a better prompt, the output gets better. Hey, this is close to usable!
Even if you keep your old templates - sometimes you want an on-the-spot refresh. Just copy your completed plan, and write “Rewrite: <your copied text>”, which will give a fresh take on your tired writing (I could use this for sure).
R.I.P. Dot phrases
The next logical question is this: do we need dot phrases anymore? Why doesn’t EPIC generate these on the fly? Todd Morgan
A great point. One of the key tenets of this substack is that we won’t wait until the cavalry arrives. In the meantime, you’re stuck toggling back and forth between Chat GPT in one window, and your ancient EHR in the other. But I don’t imagine that a Chat GPT API is too far off, where EHR integration will become table stakes.
Operative notes - a starting point
I’m not advocating using Ai to write your notes (yet). But let’s say you just did a new procedure, and you’re not sure about common parlance to describe the surgical technique. Or, you’re a trainee, and this applies to many operations you first encounter. Here’s an example for Urolift.
I began with a simple prompt.
What if you want more detail? Just ask. “Add more detail in the description of procedure section.” It’s like a personal assistant with limitless energy.
Writing letters to insurance companies
Ooooh. Now we’re talking. The ultimate time suck. Here’s one for a biopsy naive patient who required a prostate MRI. I immediately showed my administrative assistant, whose jaw is still on the floor since the reveal.
Doctor time is often wasted on administrative tasks, which takes away from their ability to focus on providing patient care. This burden needs to be relieved in order to allow doctors to focus on what they do best: treating patients.
Laura Bukuvina gave the best example of this on twitter. I’ve been using this to compose emails for the last 2 days. I’ve even pasted the text into Chat GPT and said “rewrite this,” or, “make the tone more polite and upbeat.” Rewriting for tone is an Ai superpower it seems.
Assisting clinical and translational research
To give credit where credit is due, check out Anobel’s twitter feed over the last few days. I originally found out that this capability existed from Anobel, and he keeps coming up with incredibly creative use cases.
I’ll leave a few here, and you can click to read them in detail. What seems clear to me is that if clinical research is your passion, you can immediately use AI to augment your workflow. Whether it will (nearly) replace your workflow someday soon is aspirational and somewhat frightening and exciting to ponder.
Improving lit searches:
Creating a prostate cancer survival nomogram:
Creating a full-on manuscript in seconds:
Mind blown. 😱
Developing patient education material
John Gore nicely pointed this out on twitter today.
Experiment number 1, Urolift:
What if a patient was going to have a cardiac ablation for atrial fibrillation, a topic about which I have zero knowledge.
Incredible. Will libraries of such canned text disappear in favor of malleable explanations that are autogenerated at teh point of care? (e.g. I want two paragraphs, 8th grade level, and include a discussion of risks).
My brain hurts
Discharge instructions, or any patient-facing instructions
I have toyed with the idea of creating discharge instructions using Chat GPT. The results are a great place to start.
Summarizing parts of the chart at patient-friendly reading levels.
Get ready for this… I took a sample MRI report for John Doe off the internet, and gave the prompt, “Summarize this in 5 sentences at a 7th-grade reading level.”
You can see what a beautiful job it did summarizing the key findings out of the entire report.
You can take this to its next logical conclusion - shouldn’t EHRs automatically do this? Maybe chart transparency has nothing to do with making patients wade through the actual doctor-facing records. Perhaps patient summaries should be auto-generated?
For now, you can do the same thing for the patient calling or messaging that they don’t understand their report, by giving them a readable summary.
Staying up to date
Doctors are committed to staying up to date on the latest medical advances and best practices, but they are often limited by time constraints. With the rapid pace of medical research and the vast amount of information available, it can be challenging for doctors to keep up. Additionally, competing demands abound: from their patients, administrative tasks, and other responsibilities, which can make it difficult to find time for ongoing learning and professional development.
Chat GPT to the rescue somewhat? I can see the power here, especially for medical students and trainees.
Let’s take PSA screening. You can dig up the latest review article, or do this:
I have to say this is fickle. I tried the same thing a few minutes later, and it didn’t want to cooperate.
What if you want to check the latest (the test data may not be completely up to date, but close) information on a topic, and make it patient-friendly for a progress note. I did this with BCG refractory bladder cancer.
Nursing triage? Ask it a medical question. Go ahead.
Telephone triage is hard. Ask any nurse, MA, or administrative assistant. Sometimes they know what to say, not how to say it. Other times they’re not quite sure. This requires some experimentation but is a potential launching point to augment or facilitate the challenging tasks that triage demands.
How should I explain stent symptoms after ureteroscopy?
Who knows, maybe patients will be using this technology instead of resorting to google.
Not bad making the connection to cellulitis or DVT. I’m sure we’ll be seeing more of this in the future. Right now, Ai powers chatbots on many contemporary websites. Why not have an Ai powered chatbot as the first line of triage before it gets to the doctor and their staff? That would be a significant relief if the lawyers approve, and patients with series problems don’t fall through the cracks.
Figuring out the call schedule
Anbobel Odisho strikes again. Recognizing that the arduous task of creating a call schedule is as painful as it is time consuming, he decides to put Chat GPT to the task.
The fact that it created the call schedule instantly is one thing (well, truth be told there were some mistakes if you check the results carefully, and the model required some feedback). I’m more impressed with Anobel’s prompt. He’s very clear and logical with the set of rules (he has a coder’s logical mindset). Just the fact that he even thought to do this is impressive.
The lesson here is that anytime you see manual processes and spreadsheets, there’s a solid chance software can take over the job.
What does the future hold?
Right now I’m still reeling from trying this thing out over the last 48 hours. I’ve been laughing out loud every time someone turns any of the above into a poem or a rap lyric.
Several things are clear to me. First, I was originally skeptical of Ai-powered “Scribes” that could effectively summarize what goes on during a patient encounter. However, after seeing this kind of output, I’m more bullish on this tangible possibility.
Second, the concept of virtual assistants, powered by Ai, to track symptoms and answer patient questions as the first line of support seems imminent.
Third, when the technology becomes embedded within the EHR, translating care events to patients will become automatic. Asking the EHR for running pithy summaries of disease evolution over time will be a welcome feature. I don’t imagine interns 5 years from now will be doing discharge summaries, for instance. I hope I’m not wrong about that.
Lastly, if the technology excels at pattern recognition, will it be able to scan medical records, augmented by remote monitoring, to predict who is at risk for adverse events, or the development of disease?
I’m sure all of these predictions will be wrong. I will say the world feels different today. This is an inflection point.
More than ever, I’m interested in what you’ve found by poking around with this incredible tool. What do you envision for the future? Leave a comment!