Is AI Going to Take Scribe Jobs?

[dogmomtodoc]

My Dad just went to the doctor and had to sign a waiver because his physician uses "Dragon Ambient eXperience (Nuance DAX), an ambient clinical intelligence solution that automatically documents patient encounters accurately and efficiently at the point of care." Has anyone had any experience with this? They're pretty much selling it as a better alternative to medical scribes that reduce physician burnout.

Thoughts? Feelings?

---

Members don't see this ad.

 

[SleepyInsomniac1]

The doc I scribe for uses Dragon but only when he is sitting in the office adding or fixing the notes. He does not use it during encounters with patients because it usually does not pick up what he says very well and he often has to use his keyboard to fix what it mistranslates. Not sure if the one you are specifically referencing is some upgraded version that is better or what.

 

[Med Ed]

Every time we purchase a software solution it's advertised as the best thing since sliced bread. The serious problems and shortcomings don't become apparent until later, which is why you typically have to shell out even more for the platinum support contract. And then, after 2 years of chronic disappointment, you go on the hunt for the next "solution"...

 

[baguettes]

Premed here, past career in data science. No experience using the tool, but a doctor I work with just showed me the demo video on the Nuance website via the same link you posted. My opinion: this software is still being propped up by layers of manual review, beyond the physician review step described in the demo. Its value is taking unstructured data, i.e. the patient interview, and converting it to structured data, i.e. the note. This technology will improve and will take scribe jobs, to your question. Since administrative costs contribute significantly to the cost of healthcare and administrative tasks contribute significantly to physician lifestyle, I think software like this has great potential. I can only suspect it will be years before this particular tool is reliably accurate. I think I would prefer the workflow of dictation over writing the note in the room and over scribing (in some settings), just based on my premed experiences.

 

[voxveritatisetlucis]

My PCP just records appoints (audio) rather than having a scribe. This isn’t really AI and I don’t know how common this is.

This is the way the nurse/ma described it to me. Maybe it is more than audio recording

My dermatologist always has scribes and he must have 15+ on the payroll (private practice) because whenever I see him it is somebody else

 

[jhmmd]

I remember Dragon.

Echo the sentiments above about Dragon not translating well. As with most industries, there is no substitute for work put in by humans.
e.g. See this article about how AI is changing nursing

 

[The Knife & Gun Club]

I use a dragon mic at work but that’s because I’m EM and my entire note is a series of spoken dot phrase templates except for one paragraph of MDM.

I just say “normal physicals exam” and it writes my whole PE for me. Then I quickly edit the 1-2 areas of interest.

I can not imagine a software tool that’s anywhere near being able to convert a visit to a note. When I work with a scribe I have to train them in what to say not to say for every exam because so much is open to interpretation and left unsaid in the room.

For example - a patient reports tenderness of her bilateral arms, legs, abdomen, upper back, lower back, neck, and head and lips. But the patient DEMONSTRATES pain in their lower back and does not have reproducible pain on my exam to the other areas - so I don’t chart those complaints which are likely related to distress associated with pain in her back. It’s impossible for an AI to interpret that.

 

[Mr.Smile12]

Define AI. I more appropriately envision little robots accompanying doctors and nurses or there being a permanent robot-like device in each patient room that would take notes and send info to the nurses station. If anything, trying to warn the nurses before a code occurs would be where AI would be very useful. Could scribing be part of those duties? I can imagine it to a point.

I say that as a robot in the grocery store takes inventory, rolling down the aisles as if it were an R2/R1 Star Wars unit.

 

[baguettes]

It’s impossible for an AI to interpret that.

I would say it's impossible today

Define AI.

jhmmd linked a nice article that summarizes the definition at the beginning. AI represents a suite of algorithms that take data, "learn" from it, and produce a model that makes recommendations. To get an accurate model for note production based on a recorded patient encounter alone, a high volume of data is needed, i.e. millions of patient encounters recorded representing terabytes (petabytes?) of data. The model is essentially "listening" to these conversations and learning the structure of many different types of visits, to the point where if the model encounters a visit type that it has not "learned" before, it will make its best guess based on similarity to known visit types.

My guess - Nuance probably has a fair-to-middling model for note creation that is built on thousands of recorded patient encounters, utilizes online learning, and relies heavily on the manual review step by the physician for its presently available software.

 

[TragicalDrFaust]

The practice I scribed for bought dragon and hated it enough to contract us just a while after. This was 2018 or so but I doubt the technology has improved that much.

 

[The Knife & Gun Club]

I would say it's impossible today

jhmmd linked a nice article that summarizes the definition at the beginning. AI represents a suite of algorithms that take data, "learn" from it, and produce a model that makes recommendations. To get an accurate model for note production based on a recorded patient encounter alone, a high volume of data is needed, i.e. millions of patient encounters recorded representing terabytes (petabytes?) of data. The model is essentially "listening" to these conversations and learning the structure of many different types of visits, to the point where if the model encounters a visit type that it has not "learned" before, it will make its best guess based on similarity to known visit types.

My guess - Nuance probably has a fair-to-middling model for note creation that is built on thousands of recorded patient encounters, utilizes online learning, and relies heavily on the manual review step by the physician for its presently available software.

I’m not sure what level of training you’re at so maybe you know this but there’s a ton of nuance in clinical medicine beyond what can be reasonably reproduced by technology currently available. I would be quite surprised in our lifetime if a software/machine could adequately convert my physical exam and HPI into a note that is both accurate and won’t get me sued into all holy oblivion.

 

[cj_cregg]

For some of the folks above, this is a little more advanced AI stuff than simply using plain old regular dragon dictation with your dot phrases - it listens to the conversation and then constructs a note for you. There are a few versions of this floating around out there and being piloted by, for example, the AAFP as a way to reduce documentation time and burnout, but they are still fairly "experimental" and not being widely used by practicing docs by any means. I can see them becoming popular and frequently used maybe 5-10 years in the future to the point of potentially replacing scribes, but I think it will be a while before it is (1) useful, (2) user friendly, and (3) cost effective enough to be common practice.

 

[Redpancreas]

Every time we purchase a software solution it's advertised as the best thing since sliced bread. The serious problems and shortcomings don't become apparent until later, which is why you typically have to shell out even more for the platinum support contract. And then, after 2 years of chronic disappointment, you go on the hunt for the next "solution"...

You'll be surprised with the latest models. I was watching a Grand Rounds at a top program (invited remotely by a friend- wasn't interviewing) and the attending was rounding with his students and there was this mirror-like stand. The attending was talking with an SP and in turn talking to students on a stage with a fake bed, etc. Somehow the AI put it together in this format...granted it kept removing and adding things, but the end product was chillingly on point.

CC: Hematochezia

HPI: 79F with history of HF s/p ICD, AF on Coumadin, Prostate Cancer s/p radiation to pelvis coming to the hospital for blood in his stool ongoing for the past two weeks. He first noticed it when he was having a BM and it was accompanied by a painful sensation...

ROS:
CV: + for palpitations
GI: + constipation

Physical Exam: (the attending basically only voiced what he was doing the the patient:
General: Healthy appearing male in NAD.
CV: S1, S2, irregular

Labs: Autopopulated but attending mentioned to the students a few he would highlight and they were highlighted.

Plan:

#Hematochezia:
Differential includes multiple sources of LGIB including hemorrhoids, radiation proctitis, ischemic colitis. \

The program even allowed for auto-filling and drop downs. If you said LGIB, then multiple options for LGIB showed up. If you said painful, certain one came to the forefront.

 

[jhmmd]

You'll be surprised with the latest models. I was watching a Grand Rounds at a top program (invited remotely by a friend- wasn't interviewing) and the attending was rounding with his students and there was this mirror-like stand. The attending was talking with an SP and in turn talking to students on a stage with a fake bed, etc. Somehow the AI put it together in this format...granted it kept removing and adding things, but the end product was chillingly on point.

CC: Hematochezia

HPI: 79F with history of HF s/p ICD, AF on Coumadin, Prostate Cancer s/p radiation to pelvis coming to the hospital for blood in his stool ongoing for the past two weeks. He first noticed it when he was having a BM and it was accompanied by a painful sensation...

ROS:
CV: + for palpitations
GI: + constipation

Physical Exam: (the attending basically only voiced what he was doing the the patient:
General: Healthy appearing male in NAD.
CV: S1, S2, irregular

Labs: Autopopulated but attending mentioned to the students a few he would highlight and they were highlighted.

Plan:

#Hematochezia:
Differential includes multiple sources of LGIB including hemorrhoids, radiation proctitis, ischemic colitis. \

The program even allowed for auto-filling and drop downs. If you said LGIB, then multiple options for LGIB showed up. If you said painful, certain one came to the forefront.

This is the basis of most EMRs these days (Epic comes to mind) but you still need a human being to 1) Select the appropriate drop-down menu 2) Generate the things that then become auto-populated 3) Make sure that the chart makes sense.

 

[Med Ed]

You'll be surprised with the latest models. I was watching a Grand Rounds at a top program (invited remotely by a friend- wasn't interviewing) and the attending was rounding with his students and there was this mirror-like stand. The attending was talking with an SP and in turn talking to students on a stage with a fake bed, etc. Somehow the AI put it together in this format...granted it kept removing and adding things, but the end product was chillingly on point.

So the latest model assembled a draft note from a contrived encounter that took place in a controlled environment. I think I'll stick to my strategy of being a late adopter.

 

[Redpancreas]

So the latest model assembled a draft note from a contrived encounter that took place in a controlled environment. I think I'll stick to my strategy of being a late adopter.

I was still pretty impressed! You will be too.

 

[Hyein]

.

 

[cj_cregg]

I don't think so. All these things do is copy down what you say. A good scribe can concoct the story based off of what the patient said/circumstance/what the doctor says/wants. Using the dragon the doc still has to figure out how to craft the story which can sometimes be the most annoying part, taking the nonsensical situation and making it make sense and relate to what testing was done and what orders were made.

As I mentioned above this is not just a simple dictation software. The more advanced products that are still being piloted do exactly what you just said, including the one linked above.

 

[thegecko]

Many naysayers here. These kinds of NLP tasks have simply only gotten better over time and with compute. We are developing more powerful hardware, assembling larger datasets, and creating interesting new models and training paradigms. Just because it doesn't exist today doesn't mean it won't in 5-10 years. After all, it took 20 years of AI winter to discover backpropagation.

The state that some of these products are in is fairly impressive today. I think it will be exciting to see more usable and developed products years down the line.

 

[baguettes]

Many naysayers here. These kinds of NLP tasks have simply only gotten better over time and with compute. We are developing more powerful hardware, assembling larger datasets, and creating interesting new models and training paradigms. Just because it doesn't exist today doesn't mean it won't in 5-10 years. After all, it took 20 years of AI winter to discover backpropagation.

The state that some of these products are in is fairly impressive today. I think it will be exciting to see more usable and developed products years down the line.

So true and I am with you. Unless you have studied topics in machine learning + artificial intelligence you might not understand just how far computer science has gotten even in the last decade. I still know people in Silicon Valley who are research computer scientists working on NLP problems like producing generative models for human speech -- I don't just mean audio, but actual conversation. I'm excited (and a touch nervous) for the years ahead!

 

[SpecificSchooler]

So I think I offer a unique perspective on this as I worked for Nuance on the DAX team specifically on both the client facing side (physician facing) and research side.

Premed here, past career in data science. No experience using the tool, but a doctor I work with just showed me the demo video on the Nuance website via the same link you posted. My opinion: this software is still being propped up by layers of manual review, beyond the physician review step described in the demo. Its value is taking unstructured data, i.e. the patient interview, and converting it to structured data, i.e. the note. This technology will improve and will take scribe jobs, to your question. Since administrative costs contribute significantly to the cost of healthcare and administrative tasks contribute significantly to physician lifestyle, I think software like this has great potential. I can only suspect it will be years before this particular tool is reliably accurate. I think I would prefer the workflow of dictation over writing the note in the room and over scribing (in some settings), just based on my premed experiences.

@baguettes is pretty spot on with his review. Currently, it does require a good amount of manual review behind the scenes. With that said, from when I started to when I left, the AI had made substantial progress. One of the key aspects of Nuance, and likely what played a large role in their acquisition, was the amount of data that was collected and used. We worked in many different specialties and went through so much data while I was there and I know that, for the DAX team, it was only a small chunk compared to what greater Nuance had. I think the scribe's greatest concern is still a little ways off, but unfortunately/fortunately (depends on how you interpret it) it is coming and will take scribe jobs.