His face was covered in blood. “Don’t look! Don’t look!” I told my wife as I carried my four year old into the kitchen (she’d pass out if she did). Again, another laceration. This time I didn’t have to call every surgeon I knew looking for Dermabond. I had my own kit with sterile gloves my size, sterile gauze, Betadine, a chuck pad and Dermabond. Instead of hours in the ER, I cleaned him up, glued the wound and was reading him a bedtime story an hour later - all thanks to the kit I got through Prepared Physician. Sign up through that link and use discount code OUBRE10 for 10% off!

Imagine this:

“Hey EMR, give me a brief history of this patient’s heart failure.”

Wouldn’t that be amazing?

When many think of Artificial Intelligence (AI) and medicine, they think of AI’s diagnostic & clinical reasoning abilities, as well as its ability to interpret radiology images, pathology slides and identify rashes. There’s also excitement in the realm of drug discovery. But, of course, I’m going to focus on documentation.

I’m no AI expert and I’ve been hesitant to comment on AI, as it’s impossible to predict how any new technology will be used and what regulations might or might not be placed on it… But many people keep asking my thoughts on it.

At this time, I believe the current Large Language Models (LLMs) (which is different than generative AI) could easily be deployed to DO the DOing, not necessarily the thinking, for us.

DO the time-consuming tasks for us, namely data entry and data finding, and allow us more time for critical thinking and face-to-face time with patients.

Data Finding

EMR LLM’s could immediately address two problems:

My opening example addresses both of these. These could be tailored to specialty as well. If you are a cardiologist, any problems related to cardiology could be automatically produced for quick and easy access without needing to ask. You should be able to toggle it for more or less information.

Two issues are often mentioned: reliability and consistency.

The “hallucinations” of ChatGPT (essentially making up stuff) are well known. One would need to be able to click on the information and be brought to the source of that information in the chart. Regardless, hallucinations need to be fixed.

For consistency: ask ChatGPT the same question 10 times and you’ll likely get 10 variable answers. This won’t fly in medicine, but this problem is already being worked on and (in my humble non-computer programmer opinion) likely not a difficult fix.

Despite those fixes, some still express concern about safety. To that, I give an uncomfortable response:

The error rate of a computer on such a basic task as finding text and presenting it face-up is far lower than a burned-out provider who doesn’t have time to search the entirety of the chart.

Lastly, some express concern about the inability to assess information in scanned documents. Take your smartphone and aim the camera at some text - you can likely copy and paste that text. That’s not a problem.

Does the potential for AI’s impact on documentation make all of my teaching on documentation integrity pointless?

No. In fact, it highlights the importance. In order for the output of an AI (such as my opening example) to be accurate, the notes and data that it searches must also be accurate. So, if our notes are inaccurate, the AI will be as well. This problem may be one of the main reasons this technology gets delayed - because of the many years of sloppy humans.

Data Entry

Yeah. Notes.

Ambient AI that listens to your conversation with the patient and generates a note is already being used.

Besides efficiency, there’s an obvious value-add: providers that aren’t writing notes on a computer during the patient encounter can instead focus on active listening and asking the right questions. If AI is going to assist in diagnosis-making (as I discuss later), this highlights the importance of accurate history taking.

But to my knowledge, most versions of this tech still require a human as an intermediary to double check the AI’s work and to remove ridiculous sounding sentences, hallucinations, etc. This causes a delay in note-generation with Nuance and Microsoft working hard to get this turnaround time to about two hours. Additionally, there is the concern that another human is listening to sensitive conversations held between the patient and provider. Despite those issues, patient acceptance of the technology has been high, and one health care system has seen a reduction of after-hours documentation of up to 75%.

I do have two concerns.

First, by capturing the entirety of the interaction, even with filtering out of “filler” conversations (such as the patient’s retelling of his fishing story), this will contribute to note-bloat with lengthy subjective sections and interval histories. But admittedly, it would be an upgrade from the copy/pasted non-updated interval histories we see so often.

Second, could these recordings be used during litigation and become the “body cam footage” of clinicians? Sure, in many instances the body cam footage of police officers has helped their case, but most are uncomfortable knowing they are constantly “under surveillance.” On the flip side, would this ensure clinicians actually do the things they are claiming they do - such as discussing risk / benefits / alternatives and medication side effects, etc?

The HPI

For the input, although I’m a big believer that every step of writing your note helps you think and stay organized, the value in entering subjective data is less than that of entering objective data as well as the assessment and plan. Hopefully the value-add will balance that loss.

I do hope developers ensure the output (HPI / interval history, etc) is organized for readability with a strong first sentence and a chronological story.

The Physical Exam

My concern for a decrease in synthesizing is the same for the physical exam. However, hopefully this will improve the accuracy of the documented physical exam because, as Dr. Cifu says, they become pieces of fiction when they’re written days later.

Obviously the physical exam findings would have to be expressed out loud for Ambient AI to pick it up, which may improve patient-physician communication.

The Assessment & Plan

Technology that searches the chart and suggests diagnoses already exists. However, from my demo of that technology, it seemed more that it was making the provider aware of previous diagnoses (such as afib, heart failure, hypokalemia) than truly “making” a diagnosis in the classical sense. (Are these queries and should they be held to compliant query standards? This is something Dr. Erica Remer discussed on the Jan 23 2024 episode of Talk Ten Tuesdays.)

Are computers better at medical decision making than humans? That’s a discussion I’ll save for Dr. Adam Rodman and his podcast (This episode is Part 1. Part 2 is yet to be published).

As described in that podcast episode, before we allow AI to completely take over medical decision making and writing of the A&P, perhaps a probability-based diagnosis could be suggested and then the provider could choose to consider that diagnosis more heavily or not. This brings up two concerns. First, would this result in MORE pop ups and alarm fatigue? Second, would the provider then have to explain why they didn’t pursue this AI-generated diagnoses for medicolegal purposes resulting in more documentation burden?

Regardless, at this time we certainly should have much more caution in deploying AI in the A&P.Not to mention we would need a zero-tolerance stance on hallucinations here.

Despite the current bastardization of the A&P by copy and paste hell, the assessment and plan remains precious real estate in helping physicians think by the process of detailing their thoughts and plan, while also communicating to both their future-selves and the many members of the care team.

Lastly, my greatest concern: will the increase in efficiency by the use of AI lead to an increased expectation of the number of patients seen, and ultimately result in similar levels of provider burnout? To be determined…

I look forward to looking back on this newsletter years from now and seeing if both my hopes and fears have come true. More than likely, there will be realities I would have never considered. These are exciting times!

That’s all for now. I hope that was helpful.

Please feel free to reach out and ask questions as they help inspire future issues!

Cheers,

Robert

When you’re ready, there are two ways I can help you:

Check out The Resident Guide to Clinical Documentation. The course that has helped 480+ people:

• Cut note-writing time by >50%

• Prevent medical errors and lawsuits by using notes to stay organized

• The peace of mind to know how to write shorter yet more effective notes.

Sponsor this newsletter to promote yourself, your business, your conference, etc to 4,645+ subscribers.

If you were forwarded this newsletter and would like to subscribe: