Confession: I've been using Ambient AI Scribes...

Will I continue using them? And who are they good for?

In partnership with

I’m hosting a live webinar August 26, “How I Educate Physicians on Coding Basics” for members of “The CDI and Coding Village” online community. Don’t miss out and join here:

You’ll also get access to my weekly 1-minute “CDI quick tip” videos, anatomy and physiology reviews, and our very engaged community of CDI and coding professionals.

“I agree with many of your recommendations, but you’re going to be left behind if you don’t use Ambient AI scribes."

I received that DM on Twitter by… Let’s call him Dr. U.
Challenge accepted, Dr. U.
Let’s see what they’re all about.

Ambient AI scribes are one of the hottest topics in the AI/Medical space. They promise to revolutionize physician efficiency, turn our focus back to the patient, and limit our administrative burden.

But as with any new technology, will it solve some problems and create others?

While watching some demos of these products and before my trial, I was VERY excited and hopeful about Ambient AI Scribes.

What I Used and How

I used the free versions of Heidi and Freed. These are direct to consumer ambient AI scribes. The goal of this newsletter is not to compare those two products, but instead to comment on the technology (as best as I can) after my brief experience.

My initial goal was to trial as many Ambient AI scribes as possible, including some that do not have free trials as the companies offered to let me use their products for free. That goal was not accomplished, as you’ll come to see why.

First, no patients declined my use of the Ambient AI scribe. I thought some would inquire more about it or show hesitation. However, none did.

The apps ran on my phone, which was upside down in my shirt pocket to keep the microphone exposed. The voice recognition by both was impressive. I did not identify any “hallucinations” by either - a top concern of AI’s. Incorrect output seemed to be related closely enough to the conversion that I suspected was an incorrect pickup by voice recognition rather than a hallucination. In fact, there were moments of garbled, slurred or rushed speech that I figured the AI scribe would have difficulty with, so I knew to look out for them.

Prompt Adherence

Customization is key and one platform had the ability to customize your output with prompts. However, adherence to that customization was incomplete at best and reminiscent of earlier versions of ChatGPT where the AI wouldn’t quite “get it.”

Example: The AI’s output would include, “Smith is complaining of….” After it failed to include “Mr.” or “Ms.” before the name, I told it to never say the patient’s name and only include, “the patient” instead. It wouldn’t completely follow that command, forcing me to edit the output.

Another example: I wanted it to populate my “Billing timeout” but only include this section when I specifically spoke the words, “Medical Decision Making” and then dictate my words verbatim. This way, I could “do my billing documentation” while walking from the patient’s room. However, I could never get it to do this well consistently.

I imagine these issues will 100% be fixed and perhaps other platforms do it better. But these are things to consider when adopting this technology early in their development.

Usefulness by Note Section

History

Prompt and voice recognition issues aside, I see documenting the history as the best use-case for ambient AI scribes as the technology exists now. Although the HPI output certainly didn’t satisfy many of the ideals I set forth in my four newsletters about the HPI, it was good enough. I do still recommend writing the first sentence yourself, however, for reasons I discussed in this newsletter issue.

Yes, I’ll mourn the loss of synthesizing that occurs while writing the HPI. But the use-case is just too obvious.

I also found it as a great tool to remember what we discussed. Sometimes patients ask us about seemingly “small things” that we forget about once we’re back to a computer, such as ordering TUMS for their heartburn. But I often found the Ambient AI scribe reminding me of these small details.

I also specifically prompted the AI to write, “Dr. Oubre asked the patient what additional questions they had” when it heard me ask the question. It did a fairly good job of capturing this, and this experience made me realize a good benefit of ambient AI scribes:

With the use of auto-populated templates and smart phrases, a major legal concern is: “It’s documented but did it really happen?” While you can always edit the AI’s output, I found this as a possible defense for legal purposes. “No, your honor, I didn’t just auto-populate that or write it myself, The AI heard me educate the patient on those risks and therefore scribed out that discussion.”

Physical Exam

In order for the scribe to document the physical exam it must be spoken out loud. I imagined this process would improve doctor-patient communication because I had to verbalize my physical exam findings. Perhaps it did, but both AI scribes populated aspects of the history into the physical exam.

Also, one of the most important parts of a physical exam in the acute setting is the “general appearance” - are they awake, alert, and talking? Or toxic-appearing and lethargic? This is not something I ever found easy to describe out loud. I did try warning the patient that, “I’m going to say something out loud so the scribe can pick it up” but I found this awkward, cumbersome and distracting. So, I did not continue the practice.

Because I had to edit the output and add additional findings, I did not find this to be a time-saver. However, I did find it as a great way to keep track of physical exam findings (that are easy to be mixed up when you’re seeing 15+ patients and writing your notes hours later). I typically combat this with my normal workflow by dictating this section immediately after leaving the patient’s room.

I learned something about Medicare Day 1 of subscribing to today’s sponsor. It’s something we all need: 1 source of fact-based news not just about politics. (That’s also one banger of a graphic). Support me and this newsletter by clicking and checking them out.

The Daily Newsletter for Intellectually Curious Readers

  • We scour 100+ sources daily

  • Read by CEOs, scientists, business owners and more

  • 3.5 million subscribers

The Assessment and Plan

Dr. U is a specialist who is primarily outpatient based, and he uses an ambient AI scribe practically 100%. My brother is also an outpatient physician, and he also uses ambient AI scribe for his notes 100% with 0 edits.

Me?
For the A&P?
0%.

Fellow hospitalists I’ve talked to?
No one is using it.

CIO’s have told me they’re recognizing this trend as well. Ambient AI Scribes are being adopted MUCH faster in the outpatient setting than inpatient.

Why the difference?
After my experience, I understand why.

Diagnosis Capture

I may be treating a patient for heart failure. That is the patient’s main concern and the focus of our discussion. The problem?

They also have hypokalemia, hypomagnesemia, AKI, diabetes type II with neuropathy, obesity, pulmonary hypertension, paroxysmal atrial fibrillation, COPD, etc.

Did we discuss all of those things during my visit where the Ambient AI scribe could listen? Nope.

Is it still important that I consider those diagnoses in the process of treating the patient? Yup.

Will I make small changes to some of their medications after my visit and perhaps while writing my note because the process of writing my note reminds of those other conditions and interactions? Absolutely.

Do those conditions still need to be in my A&P? 100%.

This is different than the outpatient setting, which is often either focused on a singular problem, or that which was discussed during the patient visit IS all that needs to be documented (rather than other acute issues that often prop up during inpatient encounters such as electrolyte abnormalities, organ dysfunctions, etc.).

An additional difference to outpatient:

Diagnosis Consistency

Inpatient progress notes are part of a larger encounter. You often want consistency, day-to-day to comment on things that are starting/stopping/changing.

Additionally, from a coding perspective, you want consistency of a named problem. You do not want the AI scribe to document “Cardiogenic shock” one day, then “Acute on Chronic systolic heart failure” the next day, then “Stable chronic heart failure” the next day based on the conversation it hears. That’s a coding and query nightmare.

I’ve seen AI scribes that generate output into pre-selected problems/diagnoses. This seems the way to go for inpatient.

Can you use Ambient AI Scribe to generate an A&P for some problems and just copy/forward + edit others? Perhaps. But I suspect many can copy/forward + edit faster than editing the output of an AI scribe.

If it could take my format:

Acute on Chronic Systolic Heart failure

8/12/24 TTE LVEF 35%

- Continue lasix 80mg IV BID

- Continue Entresto 24-26mg BID

And hear that I discussed changing to PO lasix and generate the output in the same format such as:

Acute on Chronic Systolic Heart failure

8/12/24 TTE LVEF 35%

- Stop IV lasix and start 40mg PO lasix daily

- Continue Entresto 24-26mg BID

then that would be *Chef’s kiss*. Also, if it could hear that and have those orders ready for me to sign? 😍😍😍 
This tech isn’t there yet, but I’m hopeful it’ll be there soon.

Auto-Generated AI Plan

I never utilized the auto-generated AI plans.

A major part of my efficiency (which I teach in my courses) is formatting your A&P the same way, with every patient, and every note so you know exactly what and where you need to edit. Varying AI Scribe outputs would zap this efficiency. Because, yes, you WILL want to edit it…

My top legal advice is to NOT document something you didn’t actually do - this is typically in the context of auto-populated physical exam templates, auto-generated labs and radiology reports, and copy/pasted A&P’s - but AI Scribe A&P outputs could join that list very soon.

Some AI scribes pride themselves on generating differentials, potential workups, etc. Can those be helpful as “nudges?” Absolutely, but again, the outputs must be edited to show what you actually did. (Also, can these outputs be tracked? If the AI posed a differential that the physician did NOT act upon and that ultimately was the diagnosis, could that be used against the physician? Are physicians going to have to start documenting WHY they didn’t work-up something the AI suggested? Food for thought.)

The A&P is also precious time to think. I do not want to outsource that to AI by signing the note without reviewing. This thinking-while-we-write is likely more beneficial to generalists (like me) who see a wide range of conditions versus specialists who mostly deal with the same subset of problems and have those differentials + managements down to mental algorithms (Dr. U agreed to that point).

EMR Integration

Lastly, both platforms required me to copy / paste from their software into my EMR. Especially with EPIC’s problem-oriented charting which creates distinct dialogue boxes for each and every problem, this was too cumbersome for the A&P.

Each also required that I create a new episode for each recording with a patient. This was an additional step that was time consuming. So, integration with EMR’s is going to be key.

A Surprise Use

A use-case I hadn’t considered beforehand is telephone calls. Details of phone calls are frequent considerations during lawsuits. They can also be lengthy and difficult to remember. With permission, I used the AI scribe to listen to my conversation on speaker phone and write the telephone note for me. This was 100% a time saver.

I didn’t, but you could also use this for provider-to-provider conversations.

Phone Battery life

This isn’t an inconsequential consideration. As hospitalists, we’re running all over the hospital. Most don’t round with computers on wheels nor laptops. So, we’d have to use our phones to run the apps. These apps drained my battery life quickly, so this also may be a reason for slower inpatient adoption. We’re going to patients. They’re not coming to us.

Who Ambient AI Scribes are Good for:

  • As evident by early adoption, outpatient providers seem perfectly positioned to use Ambient AI Scribes for their notes “top to bottom.”

  • People who under-document their conversations with patients.

    • I include myself in this group. I spend a considerable amount of time with patients, but likely under document our conversations including side effects of medicines, outpatient follow up education, incidental findings, etc. As I mentioned, this could be helpful from a legal perspective.

    • I also imagine those in the mental health field, who often have extremely lengthy visits, will benefit greatly from the automatically documented history.

  • Those who primarily copy/paste and sign their progress notes without editing.

Who Ambient AI Scribes are Not Good For:

  • Those who ghost round.

    • This includes people who round extremely early, or in the middle of the night, to intentionally avoid family members and limit conversations with sleeping patients.

    • This also includes those who walk in, say hi, talk about nothing, and then leave.

Will I Continue Using Them?

In their current iteration, not integrated with EPIC’s problem oriented charting? No.

I didn’t trial as many AI scribes as I originally planned because relearning a new platform and customizing templates every week of service was too cumbersome compared to the value they were adding to my workflow. Plus, a rumor of our health system adopting a system-wide ambient AI scribe with EPIC integration didn’t help either.

But as I discussed, they probably make sense for a lot of other people. This technology will get better, and I suspect I will be using them in some form in the future. I’m excited about the possibilities.

Don’t miss out on my first live webinar by joining our online community “The CDI and Coding Village” here.

That’s all for now. Don’t hesitate to ask questions as they help inspire future issues!

Cheers,

Robert

Thanks to Laura Samson, RN BSN CCDS, and my wife Kara (for grammar) for editing this newsletter!

When you’re ready, there are 3 ways I can help you:

  1. Join our CDI and Coding Village online community! It’s impossible to know it all, it takes a village! What you’ll get:

  • Learn from me and other experienced professionals (perfect for new coders / CDI’s and those pivoting their career)

  • Participate in monthly interactive webinars hosted by me, weekly 1 minute CDI quick tip videos, and anatomy and physiology reviews not found anywhere else

  • Share ideas & solutions to new problems with peers in real-time discussions

  • Stop underbilling your hard work & MASTER the new billing rules

  • Get 3.5 hours of CME credit! (Use your CME funds!)

  • Stop feeling guilty about writing shorter notes

  • Use notes to PREVENT getting sued

  1. Check out The Resident Guide to Clinical Documentation video course. The course that helps you:

  • Impress your attendings and improve your evaluations.

  • Prepare for real-world productivity pressures

  • Gain the confidence to write shorter yet more effective notes.

  • Use notes to make you a better, more prepared physician

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

Reply

or to participate.