AI Tech for the ER

[thegenius]

---

Members don't see this ad.

I don't think we have a dedicated threat about this.

How are people's experiences using AI apps for generating ER notes. I generally think they are kind of crappy, in that they are extraordinarily verbose. Even in the "concise" mode. Anyone happy with a particular app?

I've tried Abridge and Speke

Here is a Speke MDM, compared to mine

Speke:
The patient, a female with a history of hysterectomy, presented with hematuria, fatigue, nausea, and persistent abdominal pain. She reported episodes of red urine and significant fatigue, nausea, and abdominal pain that have persisted. The initial differential diagnosis included urinary tract infection, nephrolithiasis, and hepatobiliary pathology. The patient denied any history of bladder infections or UTIs and reported no pain during urination. She is on medications including insulin, dapagliflozin, and metformin for diabetes management.

An abdominal ultrasound was performed, revealing acute cholecystitis with a normal common bile duct. Laboratory tests showed elevated liver enzymes (ALT 804 IU/L, AST 156 IU/L, Alk Phos 209 IU/L), elevated bilirubin levels (Total 1.8 mg/dL, Direct 0.90 mg/dL), and a low albumin level (3.3 Gm/dl). The patient's sodium level was slightly low at 135 mEq/L, and CO2 was low at 19.0 mEq/L. The patient's hemoglobin and hematocrit were elevated (Hgb 16.4 Gm/dl, Hct 48.6%). The patient's glucose level was normal at 94 mg/dL.

The treatment plan included initiating IV fluids and administering antiemetic medication for nausea. The patient was advised to follow up with a gastroenterologist for further evaluation and management of acute cholecystitis. Given the elevated liver enzymes and bilirubin, further hepatobiliary imaging and possible surgical consultation may be warranted. The patient was instructed to monitor her symptoms and return to the ED if her condition worsens.

Mine (that I put in the same note):

referred for RUQ abd pain, workup shows cholecystitis. d/w surgery and will admit..



So not only is the AI generated MDM incredibly verbose and 90% useless, it didn't even get it correct. I didn't discharge her. I admitted the patient. Who in the world wants to read these AI notes?

The problem is, as I see it, are the software engineers who write these apps are not doctors. Even if they have doctors consulting, the end product is just way too lengthy.

 

[RustedFox]

Only Chuds use these things.

 

[thegenius]

Only Chuds use these things.

Modern Slang Usage:

• "Chud" is primarily a derogatory slang term used online, especially in U.S. internet culture, to describe people perceived as unpleasant, gross, or socially abnormal235.
• More specifically, it is often used to insult individuals—usually men—who hold reactionary, regressive, or bigoted sociopolitical views, particularly those associated with the far right or "manosphere" communities1357.
• The term is frequently deployed by critics on the political left to mock or dismiss individuals they see as embodying toxic or extremist attitudes, often combining accusations of racism, sexism, conspiracy theorizing, and a sense of superiority

you must be particularly spicy today, Dr. RF.

 

[TooMuchResearch]

I don't think we have a dedicated threat about this.

How are people's experiences using AI apps for generating ER notes. I generally think they are kind of crappy, in that they are extraordinarily verbose. Even in the "concise" mode. Anyone happy with a particular app?

I've tried Abridge and Speke

Here is a Speke MDM, compared to mine

Speke:
The patient, a female with a history of hysterectomy, presented with hematuria, fatigue, nausea, and persistent abdominal pain. She reported episodes of red urine and significant fatigue, nausea, and abdominal pain that have persisted. The initial differential diagnosis included urinary tract infection, nephrolithiasis, and hepatobiliary pathology. The patient denied any history of bladder infections or UTIs and reported no pain during urination. She is on medications including insulin, dapagliflozin, and metformin for diabetes management.

An abdominal ultrasound was performed, revealing acute cholecystitis with a normal common bile duct. Laboratory tests showed elevated liver enzymes (ALT 804 IU/L, AST 156 IU/L, Alk Phos 209 IU/L), elevated bilirubin levels (Total 1.8 mg/dL, Direct 0.90 mg/dL), and a low albumin level (3.3 Gm/dl). The patient's sodium level was slightly low at 135 mEq/L, and CO2 was low at 19.0 mEq/L. The patient's hemoglobin and hematocrit were elevated (Hgb 16.4 Gm/dl, Hct 48.6%). The patient's glucose level was normal at 94 mg/dL.

The treatment plan included initiating IV fluids and administering antiemetic medication for nausea. The patient was advised to follow up with a gastroenterologist for further evaluation and management of acute cholecystitis. Given the elevated liver enzymes and bilirubin, further hepatobiliary imaging and possible surgical consultation may be warranted. The patient was instructed to monitor her symptoms and return to the ED if her condition worsens.

Mine (that I put in the same note):

referred for RUQ abd pain, workup shows cholecystitis. d/w surgery and will admit..



So not only is the AI generated MDM incredibly verbose and 90% useless, it didn't even get it correct. I didn't discharge her. I admitted the patient. Who in the world wants to read these AI notes?

The problem is, as I see it, are the software engineers who write these apps are not doctors. Even if they have doctors consulting, the end product is just way too lengthy.

I just want the bare minimum documentation to bill for the work I did. Until a robot does that, I'm not touching it.

 

[thegenius]

I kind of feel that way. EIther the AI bot should
- significantly reduce time creating AND editing a note;
- increase average RVU/note
- reduce medmal liability risk

or it's not worth using.

Another problem is these AI bots are generally made for linear encounters. You see one pt, leave the room and that's it. ED is the exact opposite.

 

[RustedFox]

Modern Slang Usage:

• "Chud" is primarily a derogatory slang term used online, especially in U.S. internet culture, to describe people perceived as unpleasant, gross, or socially abnormal235.
• More specifically, it is often used to insult individuals—usually men—who hold reactionary, regressive, or bigoted sociopolitical views, particularly those associated with the far right or "manosphere" communities1357.
• The term is frequently deployed by critics on the political left to mock or dismiss individuals they see as embodying toxic or extremist attitudes, often combining accusations of racism, sexism, conspiracy theorizing, and a sense of superiority

you must be particularly spicy today, Dr. RF.

What I'm really going for here are the chronically online types (Chuds) regardless of politics (because it's all theater now anyways, but that's a different rant) who want a bot or an app to do everything for them.

Taking pride in masculine self reliance and a DIY-ethos with accountability needs to come back.

 

[Apollyon]

Cannibalistic Humanoid Underground Dwellers

 

[EctopicFetus]

The 2 I have tried and liked are Sayvant which has an MDM / med mal piece which I will admit I havent used but did see demoed and Time by Dr.H who is an EM doc. Both can be customized to your specific needs

They can all be trained on EM billing standards. I think the issue with off the shelf stuff is that it is really all made for clinics and not ED encounters. AFAIK both of these are made for EM and are trying to now make more generalizable versions for the masses.

 

[Janders]

I’ve played with Heidi and I think it has a role as a scribe replacement. It will listen to the HPI and output a reasonable text version, and can be coached into variations in form, brevity, etc

My main issue is 1/3 of patients have such quick HPI that I don’t want to futz with an AI scribe app (i twisted my ankle on the stairs), and 1/3 of patients are so altered, demented, or involve complex interpreter scenarios that I don’t think the AI app will have anything to do.

It worked great when I went in a room with a complex undifferentiated intelligent older woman with her involved daughter who brought a copy of her out of state recent hospitalization records and could rapidly answer every question I had about her complex recent health issues.

I don’t see a lot of those patients per shift.

 

[gutonc]

I’ve played with Heidi and I think it has a role as a scribe replacement. It will listen to the HPI and output a reasonable text version, and can be coached into variations in form, brevity, etc

My main issue is 1/3 of patients have such quick HPI that I don’t want to futz with an AI scribe app (i twisted my ankle on the stairs), and 1/3 of patients are so altered, demented, or involve complex interpreter scenarios that I don’t think the AI app will have anything to do.

It worked great when I went in a room with a complex undifferentiated intelligent older woman with her involved daughter who brought a copy of her out of state recent hospitalization records and could rapidly answer every question I had about her complex recent health issues.

I don’t see a lot of those patients per shift.

I use an AI scribe in a clinic setting and this is the exact scenario where I find it helpful. Most of the rest of the time it just lets me not sit there typing in the room.

 

[JacobMcCandles]

“Patient with acute cholecystitis. Recommend outpatient management with Keppra and follow up with neurosurgery in 12 months.”

Thanks, AI.

 

[southerndoc]

Check out Cleo. It's phenomenal. cleohealth.io

 

[wook]

Check out Cleo. It's phenomenal. cleohealth.io

We were just talking about this at my shop, and several had mentioned cleo...

Now if only my shop will pay for it....

 

[thegenius]

Does cleo integrate directly into your EMR?

 

[southerndoc]

Does cleo integrate directly into your EMR?

Unfortunately, no. Have to copy and paste it. DAX integrates with our EMR, but I'm not a fan of it.

 

[RustedFox]

DAX.
Cleo.
Heidi.

"My AI bot won't work... now I can't chart... can't talk to women... can't self-regulate my emotions... can't..."

Chuds.

 

[Mount Asclepius]

From what I’ve heard from others it doesn’t seem like most AI products are quite good enough yet for charting, but I think it will get there. AI needs to integrate with the EMR. I think Epic and Cerner will do that. We waste so much time typing and dictating. I think AI is the answer to this and will completely replace dictating sometime in the next 5-10 years. Not sure how I feel about a significant change mid career, but it can’t be worse. At least I hope, or I might be end career.

 

[RustedFox]

From what I’ve heard from others it doesn’t seem like most AI products are quite good enough yet for charting, but I think it will get there. AI needs to integrate with the EMR. I think Epic and Cerner will do that. We waste so much time typing and dictating. I think AI is the answer to this and will completely replace dictating sometime in the next 5-10 years. Not sure how I feel about a significant change mid career, but it can’t be worse. At least I hope, or I might be end career.

Politely disagree; we can dissect it out if you want. Later, though.

 

[Porfirio]

If AI wrote down everything that the patient and I discussed, every patient would be CBC/CMP/UA/PT/INR/Mg/Lactic/Culturesx2/Lipase/TSH/Trop/BNP/UDS/Tylenol level/Aspirin level/ETOH/Osmol/ABG/VBG/EKG/Chest X-ray/CT head/MRI brain/CT c-spine/CTA chest/CT abdomen-pelvis/US gallbladder/US DVT both extremities/MRI lumbar spine/Vitamin B level/Iron panel/PTH/FSH/Strep/Covid/Flu/RSV (all ages)/Ddimer

+/- preg test (even if age 85)

The patients have no filter and AI certainly cannot filter the patient. It takes me 3-5 mins to see the average patient, and 2-3 mins to document. Why do I need AI?

 

[TooMuchResearch]

If AI wrote down everything that the patient and I discussed, every patient would be CBC/CMP/UA/PT/INR/Mg/Lactic/Culturesx2/Lipase/TSH/Trop/BNP/UDS/Tylenol level/Aspirin level/ETOH/Osmol/ABG/VBG/EKG/Chest X-ray/CT head/MRI brain/CT c-spine/CTA chest/CT abdomen-pelvis/US gallbladder/US DVT both extremities/MRI lumbar spine/Vitamin B level/Iron panel/PTH/FSH/Strep/Covid/Flu/RSV (all ages)/Ddimer

+/- preg test (even if age 85)

The patients have no filter and AI certainly cannot filter the patient. It takes me 3-5 mins to see the average patient, and 2-3 mins to document. Why do I need AI?

Think of it this way: once you see a full department, the AI driven workups will take 12+ hours each, you can write your notes and go to lunch.

 

[Mount Asclepius]

The error is thinking that AI can’t encounter a million patients and equate those with the end results filtering out the noise.

 

[old_boy]

The error is thinking that AI can’t encounter a million patients and equate those with the end results filtering out the noise.

Will the lawyers let the AI filter out the noise?

Honest question.

Separate but related thought: We are flirting with every patient encounter with full audio and video backup forever.

Will AI be called to testify?

“The patient said she had chest pain. Why didn’t you put it in the note?”

 

[JacobMcCandles]

If AI wrote down everything that the patient and I discussed, every patient would be CBC/CMP/UA/PT/INR/Mg/Lactic/Culturesx2/Lipase/TSH/Trop/BNP/UDS/Tylenol level/Aspirin level/ETOH/Osmol/ABG/VBG/EKG/Chest X-ray/CT head/MRI brain/CT c-spine/CTA chest/CT abdomen-pelvis/US gallbladder/US DVT both extremities/MRI lumbar spine/Vitamin B level/Iron panel/PTH/FSH/Strep/Covid/Flu/RSV (all ages)/Ddimer

+/- preg test (even if age 85)

The patients have no filter and AI certainly cannot filter the patient. It takes me 3-5 mins to see the average patient, and 2-3 mins to document. Why do I need AI?

I tend to agree. I think AI for physician charting is currently a solution looking for a problem. I can crank out a note on a patient in a few minutes. It's a note I feel comfortable with and defending in the event of a bad outcome on my decision making and thought process. I think I'd spend more time reading and editing an AI note other than me just doing it from the start. Now, you want AI to do some of the ridiculous nurse charting and questions they have to ask? Go for it.

 

[RustedFox]

The error is thinking that AI can’t encounter a million patients and equate those with the end results filtering out the noise.

This is one of the points I wanted to discuss. I don't think AI could filter thru the nonsense that patients say, especially if we "teach" it via the honest examples of HPIs written by a GOOD emergency physician.

 

[xaelia]

I tend to agree. I think AI for physician charting is currently a solution looking for a problem.

I think we're basically looking at this problem very narrowly through our EM lens. The pajama-time primary care docs would rather quit that lose their scribes.

With good templates/smartphrases/smarttext/Dragon shortcuts/whathaveyou, EM docs are probably #1 in the world at maximizing per-patient efficiency of evaluation and documentation. An basic AI scribe doesn't have that much to add.

But, then, there are some shops that use human scribe services – yeah they do your note, but they also queue up orders, set up discharge paperwork, etc. The next generation of Epic/Cerner integrated scribe products will do a lot of that for you, just based on your conversation content, as well. Now we're getting somewhere – even if the note-writing speed/quality is a wash.

And, then, yeah, further versions will obviously allow docs more control over the style, format, and content of different sections – scribe companies need to do this to allow for all the various permutations of outpatient settings, so that'll trickle down back to the ED.

Eventually it'll be second-nature to us all, even if that day is not yet today.

 

[TooMuchResearch]

I think we're basically looking at this problem very narrowly through our EM lens. The pajama-time primary care docs would rather quit that lose their scribes.

With good templates/smartphrases/smarttext/Dragon shortcuts/whathaveyou, EM docs are probably #1 in the world at maximizing per-patient efficiency of evaluation and documentation. An basic AI scribe doesn't have that much to add.

But, then, there are some shops that use human scribe services – yeah they do your note, but they also queue up orders, set up discharge paperwork, etc. The next generation of Epic/Cerner integrated scribe products will do a lot of that for you, just based on your conversation content, as well. Now we're getting somewhere – even if the note-writing speed/quality is a wash.

And, then, yeah, further versions will obviously allow docs more control over the style, format, and content of different sections – scribe companies need to do this to allow for all the various permutations of outpatient settings, so that'll trickle down back to the ED.

Eventually it'll be second-nature to us all, even if that day is not yet today.

Our scribes can barely write sentences, would rather misspell a consultants name than look it up, and forget to do most of the things I ask them to do. They also do not do any of those other things you mentioned. Some of my partners think we need to pay them more. I think we need to let them all go and move on with our lives.

 

[Mount Asclepius]

I think AI will replace scribes. The bad ones are never worth it. The good ones leave for medical school or another better endeavor. They do help with other tasks (maybe their better use), but I also think there is room for some automation. I’ve never been a fan of using scribes.

Even only a few minutes charting per patient (say 2-3 minutes) adds up to around 30-60 minutes total of charting per shift. I’m skeptical and think most exaggerate how little time they spend charting, or it shows with a bare bones or completely templated note that doesn’t always help with billing, peer review/medmal or downstream patient care. On top of that, time spent clicking through the EMR repeatedly during a patient’s ED course adds even more uncounted time. We are highly efficient at multi tasking and optimizing time, yet we still spend a lot of time sitting at a computer documenting what someone else tells us and what we think along with other small tasks. This is an area ripe for improvement and a change will significantly increase job satisfaction for many. Again, I don’t think it’s fully developed yet, but I think it will be the future even if others don’t agree that it will be or see its potential benefit yet.

 

[RustedFox]

Our scribes can barely write sentences, would rather misspell a consultants name than look it up, and forget to do most of the things I ask them to do. They also do not do any of those other things you mentioned. Some of my partners think we need to pay them more. I think we need to let them all go and move on with our lives.

Who TF still has scribes?

 

[thegenius]

Me, motherfocka!