Vulnerability of EM to machine learning?

[Brahnold Bloodaxe]

If you've been keeping your ear to the ground you may have noticed that we're possibly crossing a "threshold" at which weak AI starts being functionally useful. There has been a tremendous spike in both public interest, research papers, and real world achievements (image recognition, Go, Poker etc) in machine learning. Ironically, despite the assurances of people in the medical field that doctors will be among the last occupations to be automated away, the current push into machine learning is primarily directed into two fields:

1) Autonomous vehicles
2) Healthcare

That's right, Silicon Valley has identified the two lowest hanging fruits for disruption by AI and they seem to think they are truck drivers and doctors! It is easy to imagine how AI might do away with a significant chunk of radiologists, for example, by increasing throughput so much that only a small fraction of the current radiology workforce is needed to handle the same workload.

How about EM? Based on the nature of the work you guys do, do you consider a typical ED shift to be vulnerable to...er..augmentation by AI?

How much of an EM doc's workload consists of being handed information gathered by say a midlevel, and then generating an output based on that information such as labs to order, treatments to give, etc? And how much of the information you guys use to generate your "output" cannot be gathered by lower trained, lower paid providers?

I think you know where I'm going with this. If a large part of what an EM doc does is data analysis and decision making based on information gathered by third parties, then a sufficiently advanced machine learning algorithm could theoretically do that part of the job. Instead of gathering data and presenting it to a EM doc, an ED provider could feed the data into an algorithm and the algorithm could generate a list of further actions to be taken based on that data.

So I wrote all that just to set the stage. Obviously, a large part of EM consists of making decisions and physically intervening in a time critical manner that is not conducive to the model I described above. If you're running a code or performing a critical intubation then obviously that is no time for a midlevel-AI-midlevel loop.

The question I have is, based on the nature of your average shift, do you think a "competent" AI in combination with a midlevel to provide the "human" capabilities could be a potential game-changer in the field?

(In case anyone is interested here is a longish recent article on the topic:
A.I. Versus M.D.)

---

Members don't see this ad.

 

[Petypet]

If its any thing like the mumbojumbo the EKG spits out on top of the EKGs, I wouldn't personally invest much stock in AI...

 

[Janders]

I'm not too concerned, the ED is so chaotic and so much is learned/focused history taking by a skilled provider, and gestalt-formation that I have zero fear of being replaced in the next 1-2 decades.

It would be way easier to AI-replace well child visits, annual PCP visits, screening colonscopies, etc etc. Much harder to replace a skilled ED doc.

 

[Psai]

So they can't even make a decent computer system where you can write down all the information and look things up efficiently to act on it but suddenly they're going to replace doctors with their fancy algorithm learning?

Yeah okay.

 

[TBoneKoko]

Maybe in 120 years. Not in 20 years.

 

[tux4]

Machines make decisions based off of information. 90% of the information by 90% of the patients is incomplete, incorrect, or a lie.


Sent from my iPhone using SDN mobile app

 

[Zebra Hunter]

The day a computer can do my job is the day every single other job has been replaced by a computer. Is the computer going to talk to consultants regarding patients? Who is going to intubate, place a central line, etc on a crashing pt? Who is going to be gathering and inputting the information obtained from the patient or family member (a tech or nurse is not going to be able to do this)? Are patients going to be okay seeing a computer rather than a doctor? Who is going to determine disposition of the patient when most conditions do not have good risk stratification data regarding inpatient vs outpatient care? How is the computer going to be able to deal with more nuanced situations where things like social issues are brought into the picture? Is the computer going to be ordering d-dimers on every single chest pain or SOB pt that you can't PERC out, or are they somehow going to be able differentiate the population among these pts that requires testing? Is every questionable septic ESRD/CHF pt going to get their 30cc/kg bolus of fluid, or are there going to be nuances to this? Who is going to be at the bedside of the quadriplegic, demented nursing home pt talking with the family about how intubating their mother would only be prolonging her death?

Sure, their might be AI that comes along and assists us with our job, but most of medicine is still an art rather than a science, and even the science is very nuanced.

 

[Jlaw]

I'm 2 months from being a 2nd year ED resident and their are still patients who have normal vitals, a benign story and I say send home who my attending takes one look at and orders a full workup resulting in admission for something craz

 

[goodoldalky]

So they can't even make a decent computer system where you can write down all the information and look things up efficiently to act on it but suddenly they're going to replace doctors with their fancy algorithm learning?

Yeah okay.

gold

 

[n2b8me]

20 years ago in medical school an attending foretold us about how when we were in the wards there would be a geneticist and DNA specialist on all of our teams to round on patients. They would be giving recommendations about the person's dna profile and how it affects their disease process and recommending medications based on their genome. 20 years later that was obviously fantasy thinking, but DNA testing has revolutionized medicine in other ways. Cancer testing on a cellular level and for us in the ER PCR testing for all types of infectious diseases.

I remember a couple of months ago having dinner with friends on a Friday night in Mountain view and watching the autonomous google car drive through the streets. It looked like a granny was behind the wheel. Cars were swerving around it because it was taking so long to get down the block having to stop for double parked cars and pedestrians jay walking. I asked my friend if they always drive this way and she said one car was holding up traffic for blocks and the police officer pulled the car over to issue a ticket, but the person in the car said they weren't driving and so not responsible for the hold up... who is supposed to get the ticket then google?

As a practicing ER doc, I hope AI can help us avoid patients having catastrophic life events because of missed diagnoses. I have a feeling the computer output will look something more like popups for the practicing ER docs. Reminders, cautions, suggestions... etc. Someone always needs to be held responsible for their actions in the ER... I don't think the software company wants that responsibility - they will probably see their role as a consultant: "Recommend this that and the other, clinically correlate and reconsult prn if something changes. Thank you."

Last bit - us pit docs know how to tell if someone is sick or not sick often by just walking in the room. Its a gestalt that I think will be hard for software to emulate with sensors. We can immediately recognize and name huge spectrum of emotions- fear, pain, anxiety, drowsiness, elation, boredom, anger etc. It sounds simple and anyone is capable of this even a child... but for a machine it would be incredibly complex. Most often in the ER our actions are guided by this 'simple' pathway rather than a bunch of "data points" and extrapolation based on an algorithm.

And for the record, we already have AI in our ERs. Its called "Dr. Google"

 

[gman33]

Until lawyers can figure out how to sue a computer, my job will be safe.

 

[Arcan57]

Could you create a subset of ED patients that an AI could appropriately manage given a physical exam input from a midlevel? Yes. Could you create a subset of ED patients that a midlevel could manage without attending involvement at all? Yes, it's already been done. Could you create an ED where only those subset of patients present? That's the trick.

So much of what an experienced attending brings to the table is gestalt that is not easily coded using our current inputs. Looking at what you'd be deriving your algorithms from, the vast majority of decision rules are incredibly sensitive but non-specific. An AI based on those rules is going to be a high utilizer of testing.

 

[slatermd]

I love this question. It's medicine's version of the zombie apocalypse, and just a different iteration of "them darn midlevels are comin' to git are jobs!" I think instead of being afraid of robots, we should welcome anything that makes our jobs easier. Robots can't simulate a human impression of another human (is this patient sick or not? Is this patient full of **** or not?), but maybe Watson can communicate with Epic or Cerner better than I can, save me lots of time on chart review, and write notes better than your average scribe. If we can make that happen, we should all be happy. You'd think that would also satisfy the cost-saving illuminati played by hospital admins in these AI conspiracy theories.

Historically, medicine is incredibly bad at utilizing technology. How many decades did we spend using amazing word processing software, email, and instant messaging in our daily lives, but refusing to let go of paper charts? Where I work now, the most sophisticated automation is a herd of wheeled carts that move dirty laundry around the bowels of the hospital. I can tell you, those guys are a long way from being able to deliver a baby, cric a patient, or do a double lumen intubation through massive hemoptysis. The rest of the hospital may think ED docs are developmentally delayed Neanderthals that slipped through the cracks in medical school, but if you consider the more sophisticated things we're trained to do, there's no combination of robot + midlevel that could replace us. When you start talking about the low-value, monotonous tasks in our day-to-day, many of which could be done by a person or machine without 7+ years of medical education, robot +/- midlevel could feasibly take that off our hands, and we should be eager to let them.

For the time being, we haven't figured out how to replace scribes, bed managers, case managers, unit clerks, housekeeping, transport, RT, pharmacy...you get the idea. We haven't fully replaced any human position in the hospital with robots, and it doesn't feel like we're anywhere close to it. The technology we use to monitor patient vitals is plenty sophisticated enough to record and store information, but when I go into the EMR, I don't see vitals for a patient unless a nurse or tech takes the numbers from one robot and puts them into the other. If we can't automate that step, I can't fathom how long it would take to replace everything an ED doc does in a given shift.

 

[Fox800]

Machines make decisions based off of information. 90% of the information by 90% of the patients is incomplete, incorrect, or a lie.


Sent from my iPhone using SDN mobile app

Damn, I was thinking this was gonna be a GeneralVeers post.

 

[witzelsucht]

my EMR claims that every patient on a low dose SSRI can't have any medicines because of the risk of long qt. normal saline and lexapro? torsades. motrin and paxil? VT arrest. tylenol and zofran? may as well just mainline a bolus of K

 

[gro2001]

If the current system worked in a way that computer scientists think it works, then yes, we would be not too far away from replacing (or at least significantly changing) at least some aspects of MD work with AI. However, we still live in a world where one hospital's EMR can't talk to another hospital's EMR. AI needs data. As long as you sometimes need to use a freaking FAX machine to collect the data, it's just not going to work.

I don't think we should be afraid of the change though. The future is going to look different, but I don't think it includes crowds of starving ER docs.

 

[Mad Jack]

If you've been keeping your ear to the ground you may have noticed that we're possibly crossing a "threshold" at which weak AI starts being functionally useful. There has been a tremendous spike in both public interest, research papers, and real world achievements (image recognition, Go, Poker etc) in machine learning. Ironically, despite the assurances of people in the medical field that doctors will be among the last occupations to be automated away, the current push into machine learning is primarily directed into two fields:

1) Autonomous vehicles
2) Healthcare

That's right, Silicon Valley has identified the two lowest hanging fruits for disruption by AI and they seem to think they are truck drivers and doctors! It is easy to imagine how AI might do away with a significant chunk of radiologists, for example, by increasing throughput so much that only a small fraction of the current radiology workforce is needed to handle the same workload.

How about EM? Based on the nature of the work you guys do, do you consider a typical ED shift to be vulnerable to...er..augmentation by AI?

How much of an EM doc's workload consists of being handed information gathered by say a midlevel, and then generating an output based on that information such as labs to order, treatments to give, etc? And how much of the information you guys use to generate your "output" cannot be gathered by lower trained, lower paid providers?

I think you know where I'm going with this. If a large part of what an EM doc does is data analysis and decision making based on information gathered by third parties, then a sufficiently advanced machine learning algorithm could theoretically do that part of the job. Instead of gathering data and presenting it to a EM doc, an ED provider could feed the data into an algorithm and the algorithm could generate a list of further actions to be taken based on that data.

So I wrote all that just to set the stage. Obviously, a large part of EM consists of making decisions and physically intervening in a time critical manner that is not conducive to the model I described above. If you're running a code or performing a critical intubation then obviously that is no time for a midlevel-AI-midlevel loop.

The question I have is, based on the nature of your average shift, do you think a "competent" AI in combination with a midlevel to provide the "human" capabilities could be a potential game-changer in the field?

(In case anyone is interested here is a longish recent article on the topic:
A.I. Versus M.D.)

Yeah, not gonna happen in our lifetimes.

The trouble with all of these predictions is they don't understand that algorithms are only as good as the inputs. A machine can't do a thorough physical exam, so a clinician is needed. And that clinician needs to know what to look for and what questions to ask. So basically, you need a doctor for quality inputs to work, regardless of how good the AI is, and at that point, doctors are merely being assisted by machines, not replaced by them.

 

[Fox800]

I, for one, welcome our new DilaudidBot and TurkeySammichatron overlords. Press Ganey scores of 6 out of 5 are not unreasonable.

 

[slatermd]

However, we still live in a world where one hospital's EMR can't talk to another hospital's EMR. AI needs data. As long as you sometimes need to use a freaking FAX machine to collect the data, it's just not going to work.

Interestingly, you mention the exact problem AI is starting to solve for the ED. I agree with you completely that machines will only replace portions of our work, not the entirety, and we should welcome them to do so. None of us are likely to still be working by the time Skynet becomes self-aware.

A machine can't do a thorough physical exam, so a clinician is needed. And that clinician needs to know what to look for and what questions to ask.

Let me start by saying I agree with you, so I don't accidentally come across as if I'm arguing. Just to enhance the discussion (because I really do love these zombie apocalypse conversations), I think the question the OP was raising is will it ever be feasible that the clinician the robot needs could be a midlevel. By the end of my second year of med school I was pretty damn good at the physical exam...I just needed some time to understand the sensitivity/specificity of the maneuvers I was doing, as well as some clinical experience to really see positive tests (instead of trying to elicit rebound tenderness on a healthy classmate). If physical exam were the only thing the robot needed a human for, it might be feasible to replace the MD or DO with a less expensive clinician. But as many are saying, there are a lot of things doctors do or decide based on years of experience, and those things can't be explicitly taught or programmed.

@Arcan57 brings up another interesting idea, that if we could create an ED/urgent care setting (and a prehospital routing situation) where all midlevel-appropriate patients presented only to the robot/midlevel clinic instead of coming to the hospital, you could have some level of "ED" where doctors have been "replaced."

We could feasibly create that type of ED/urgent care (Walgreens is probably already working on it), but it will never catch all of the appropriate patients. For one thing, there's a fair volume of patients in the grey area of acuity, where you don't know they're appropriate for a lower/outpatient level of care until you work them up. Secondly, as long as people can walk into any ED they choose and seek care, I imagine there will always be people who opt out of the robot clinic. Unless...

I, for one, welcome our new DilaudidBot and TurkeySammichatron overlords. Press Ganey scores of 6 out of 5 are not unreasonable.

What if, for greater efficiency and throughput, the turkey sando had the dilaulau IN IT? Sounds like a good QI project.

 

[DrWhozits]

Just for sake of argument, mid-level + machine seems like it could feasibly work (given huge advances in AI); however, in this world, I would imagine EM physicians would still be preferred for the robot assistant over a NP/PA.

In my opinion, every "oh no, midlevels are gonna take our jobs!" conversation falls apart once salaries equalize. If mid level salaries went up and physician salaries went down until they were equal, a physician would be chosen over a mid level, every single time. So even in a physician dystopian future, we would just end up making less money, never unemployed.

Edit: never unemployed until Skynet becomes self-aware and replaces everyone.

 

[TrumpetDoc]

I can see a funny YouTube spoof right now with the dilaudid bot seeing a patient and refusing to give narcotics.
"I'm sorry, you do not get narcotics as you do not have a fracture or acute surgical emergency evident." Next, patient violently breaks the multi million dollar robot.
All hospitals press Gainey scores Dive and and the program is scrapped.

 

[dadaddadaBATMAN]

If you've been keeping your ear to the ground you may have noticed that we're possibly crossing a "threshold" at which weak AI starts being functionally useful. There has been a tremendous spike in both public interest, research papers, and real world achievements (image recognition, Go, Poker etc) in machine learning. Ironically, despite the assurances of people in the medical field that doctors will be among the last occupations to be automated away, the current push into machine learning is primarily directed into two fields:

1) Autonomous vehicles
2) Healthcare

That's right, Silicon Valley has identified the two lowest hanging fruits for disruption by AI and they seem to think they are truck drivers and doctors! It is easy to imagine how AI might do away with a significant chunk of radiologists, for example, by increasing throughput so much that only a small fraction of the current radiology workforce is needed to handle the same workload.

How about EM? Based on the nature of the work you guys do, do you consider a typical ED shift to be vulnerable to...er..augmentation by AI?

How much of an EM doc's workload consists of being handed information gathered by say a midlevel, and then generating an output based on that information such as labs to order, treatments to give, etc? And how much of the information you guys use to generate your "output" cannot be gathered by lower trained, lower paid providers?

I think you know where I'm going with this. If a large part of what an EM doc does is data analysis and decision making based on information gathered by third parties, then a sufficiently advanced machine learning algorithm could theoretically do that part of the job. Instead of gathering data and presenting it to a EM doc, an ED provider could feed the data into an algorithm and the algorithm could generate a list of further actions to be taken based on that data.

So I wrote all that just to set the stage. Obviously, a large part of EM consists of making decisions and physically intervening in a time critical manner that is not conducive to the model I described above. If you're running a code or performing a critical intubation then obviously that is no time for a midlevel-AI-midlevel loop.

The question I have is, based on the nature of your average shift, do you think a "competent" AI in combination with a midlevel to provide the "human" capabilities could be a potential game-changer in the field?

(In case anyone is interested here is a longish recent article on the topic:
A.I. Versus M.D.)

I am not particularly worried about what a group of people in silicon valley think will happen. It is difficult to defend your profession against a mythical and God-like AI which does not exist. Any argument is met with "well no one ever thought AI would beat us at chess" or "nobody predicted the internet" or "yeah well AI will be able to do THAT in 10 years." The claims are baseless or irrelevant, which makes the defense pointless and frustrating.

I would be concerned if I saw clinical trials specifically showing equivalence between physicians and AI within the ED. Even with EKGs, which are literally a graph on 2 axes, current AI is unable to beat an experienced human. It has difficulty recognizing atypical presentations and filtering out crappy data.

No individual task that a physician does is that difficult, but this is frequently only apparent in retrospect. It isn't hard understand an attending went into a room and clarified that the "chest twinge" was crushing substernal chest pain accompanied by diaphoresis and shortness of breath. It is difficult to figure out how to be that attending who knows the right questions to draw that specific history out of a patient, and to do so consistently. It is also difficult to decide what workup to pursue when a patient reports a large variety of positive complaints.

These are the kinds of tasks that limit automation. There are many others. They require detailed language processing, experience and high emotional intelligence.

Added barriers include high risk procedures (intubation, lines, etc) and communication skills (with patients, consultants, staff) that are present in the ED. Many of these tasks are time sensitive and required numerous times per shift, especially in high volume high acuity EDs. They are each tailored to the individual situation.

Finally, I would argue that subspecialty clinics, which deal with an extremely selected/filtered population and a far narrower differential, are signficantly easier to automate. The highly specialized physician or nurse has an extremely narrow scope of practice. The tasks are usually less time sensitive and far less intimidating as a problem than the completely undifferentiated patient in a high acuity setting.

 

[emergentmd]

Dr Emergent comes in, takes a history and it goes something like this with a Chief complaint of Sore throat.

After my typical 1 min history and physical, they are out the door with a script.

Dr AI sees the same 45 y old pt and it will go like this.

DR AI - what brings you in
Pt - sore throat
Dr AI - do you have any other complaints
Pt - I have Chest pain, shortness of breath, abdominal pain, vomiting, body aches, headache, and neck pain. I obviously ignored all of this as the pt is smiling while posting the ER visit on facebook.
Dr AI after a 30 min hx/exam dealing with all of the other complaints orders Cardiac enzymes, CXR, EKG, Pan Labs, LP, CT head, CT chest, CT abd.
After 3 hr of evaluation, pt is recommended to be admitted for Chest pain rule out.

In the 8 hr shift, Dr AI has seen a whopping 8 pts with full workup on everyone and have admitted 7 pts for Chest pain Rule outs, Abd Obs.

EM doc need not be worried. The lowest hanging fruits to be replaced with be pathologists, Radiologists, and even anesthesiologists.

 

[NinerNiner999]

The clinical side of our jobs is the smallest portion of actual "work" that we need to complete our day. I definitely see a role for AI in medication reconciliation, patient billing, insurance collection, transfer paperwork and verification, patient bedding, nursing home referral/rehab placement and other non-clinical tasks that are fumbled daily by inept underpaid minions who hate their jobs but make ours incredibly difficult.

 

[Interpolfanclub]

I wish a robot could've seen most of the patients on the shift I just worked.

 

[Arcan57]

I'd like to see our AI once the programmers realize they have to deal with cross-cutting core measures that are traditionally outside of EM. Somebody gets an NIH grant for preventative medicine in the ER and all the sudden the robot is having a scripted conversation about weight loss with a pre-contemplative overater with a BMI of 60 or a chain smoker with a COPD exacerbation.

"WTF did you just say to me, computer?"

 

[gro2001]

I think the first version of the AI ER doc is a widget that will observe your patient interaction and style of writing notes until it is able to document on your behalf. And I can't wait for that day.

 

[deleted77919]

…

 

[BoardingDoc]

Guys, none of us needs to worry about this until 2371.

We need to start as game a la the Super Troopers meow game where you have to see at least 10 patients and introduce yourself with: "please state the nature of the medical emergency"