AI Docs within a decade

[hoot504]

Just like google maps, but better.

---

Members don't see this ad.

 

[Jlaw]

Just like google maps, but better.

EM will probably be the hardest specialty to be replaced by a computer. Yes a computer can probably do radiology better than a human but a computer won't be able to quickly separate the freaking out psych patient vs the freaking out meningitis, smell the marijuana on the pt with intractable vomiting, or quickly respond to the vfib arrest in the waiting room that gets dropped on a stretcher in front of you with no history and a minimal explanation. Is a computer going to do an impromptu lung and cardiac US during an arrest? Or do a crash subclavian and aline? Even i they took the diagnosis away from us who is doing the procedures that actually save lives?

 

[miacomet]

Can the computer deal with the violent patients, addicts, and those who just won't leave? I envision a huge Teddy Ruxpin-like robot that gives patients a big hug and then escorts them out the door...

 

[engineeredout]

Sir what brings you in tonight?

"I just don't feel right"

*84 negative ROS later, computer starts smoking*


I'd also like to see an AI doctor place a medication order while trying to avoid the conflict errors that come up with every order, including the conflict of heparin b/c patient has history of HTN, inhaled steroids because you placed a diagnosis of sepsis, the duplicate medication error because you ordered both a bolus and a gtt of the same med, and the allergic reaction error because you ordered aztreonam but the patient's mom's tooth itched with keflex once.

 

[VA Hopeful Dr]

Sir what brings you in tonight?

"I just don't feel right"

*84 negative ROS later, computer starts smoking*


I'd also like to see an AI doctor place a medication order while trying to avoid the conflict errors that come up with every order, including the conflict of heparin b/c patient has history of HTN, inhaled steroids because you placed a diagnosis of sepsis, the duplicate medication error because you ordered both a bolus and a gtt of the same med, and the allergic reaction error because you ordered aztreonam but the patient's mom's tooth itched with keflex once.

I think it more likely the pan-positive ROS would break the AI. The pan negative would make it say the patient was fine and kick them out.

 

[namethatsmell]

Enough of this. Perhaps one day ED docs will be replaced by AI but it ain't happening in our working lifetimes.

In the now infamous To Err is Human only 17% of all preventable errors were attributed to an error in diagnosis. That number is likely lower today. Yet, right now, at this very moment, the average ED "system" can't get patients undressed and into gowns before a doc sees them, can't keep otoscopes working in most rooms, can't incentivize staff to accurately record a respiratory rate, etc blah blah blah. Maybe when basic issues like this (which aren't sexy but matter in terms of safety and efficiency) are consistently addressed we'll perhaps be ready for AI.

But sure, let the VC guys pour vast sums into fancy tech toys to try to "help" a healthcare system that doesn't have any semblance of a decent foundation to rest upon and while millions don't have health insurance.

In the meantime, here on present day planet Earth, I'll be slogging along in my ED with the curtains still covered in blood from six shifts ago. If you walked in now, maybe you'd see me glaring at the perpetually broken translator phone on my desk after learning we're on computer downtime again for the 5th time in two days. I also won't be taking my flying car to work. Won't be using my tricorder to help patients. And won't be grabbing lunch from my food replicator.

 

[deleted103644]

EM will probably be the hardest specialty to be replaced by a computer. Yes a computer can probably do radiology better than a human but a computer won't be able to quickly separate the freaking out psych patient vs the freaking out meningitis, smell the marijuana on the pt with intractable vomiting, or quickly respond to the vfib arrest in the waiting room that gets dropped on a stretcher in front of you with no history and a minimal explanation. Is a computer going to do an impromptu lung and cardiac US during an arrest? Or do a crash subclavian and aline? Even i they took the diagnosis away from us who is doing the procedures that actually save lives?

It’s funny how every specialty thinks they’re the hardest to replace.

Image recognition was a very hard problem in computer science until relatively recently (2012), and still very active research.

Most of what you’re describing could be handled by AI + a mid level provider more easily than automating image interpretation. Politically you guys might be able to hold on longer, but not because it’s impossible.

Make no mistake that a computer will be better at diagnosing most conditions than you in the not too distant future, so plan accordingly.

 

[Tenk]

It’s funny how every specialty thinks they’re the hardest to replace.

Image recognition was a very hard problem in computer science until relatively recently (2012), and still very active research.

Most of what you’re describing could be handled by AI + a mid level provider more easily than automating image interpretation. Politically you guys might be able to hold on longer, but not because it’s impossible.

Make no mistake that a computer will be better at diagnosing most conditions than you in the not too distant future, so plan accordingly.

Make no mistake by the time that happens nobody will care because we probably won’t even need jobs as our lives will be supported by AI slaves.

 

[doctalaughs]

Get back to me when AI replaces a single aspect of medicine like reading a 2-dimensional squiggly line (EKG).

Better yet, get back to me when AI replaces the lady who cleans my toilet and dusts my shelves before I start to worry.

 

[Brigade4Radiant]

If AI replaces medicine a job that takes 4 years of college and a minimum of 7 years of education and several years of practice to get good then humans don't have to work at all. You can make an AI executive etc.

 

[Mad Jack]

EM will probably be the hardest specialty to be replaced by a computer. Yes a computer can probably do radiology better than a human but a computer won't be able to quickly separate the freaking out psych patient vs the freaking out meningitis, smell the marijuana on the pt with intractable vomiting, or quickly respond to the vfib arrest in the waiting room that gets dropped on a stretcher in front of you with no history and a minimal explanation. Is a computer going to do an impromptu lung and cardiac US during an arrest? Or do a crash subclavian and aline? Even i they took the diagnosis away from us who is doing the procedures that actually save lives?

I dunno, I think psych might be the last to fall to AI because it's really hard to talk about suicidal ideation with a robot. But all this talk of AI docs is unfounded and a pie-in-the-sky fantasy of Silicon Valley execs that don't actually understand medicine and think it is all just algorithms.

 

[miacomet]

Get back to me when AI replaces a single aspect of medicine like reading a 2-dimensional squiggly line (EKG).

Better yet, get back to me when AI replaces the lady who cleans my toilet and dusts my shelves before I start to worry.

I studied AI very briefly, and at that time they said some of the hardest tasks for a computer to understand were buying orange juice and folding laundry. More "intellectual" tasks were much easier FWIW.

 

[NITRAS]

I feel like someone has to say something smart *** out about how ED's order Panel A with 23 lab and imaging tests and think they are brilliant when they have diagnosed them.. . . .

In all seriousness, the general trend of medicine to make everything a protocol based on "expert" opinion. If we aren't careful, we all will be replaced.

 

[namethatsmell]

I feel like someone has to say something smart *** out about how ED's order Panel A with 23 lab and imaging tests and think they are brilliant when they have diagnosed them.. . . .

Funny, usually the only time I have patients in my ED with 23 orders are when 1) they may be ending up in an ICU 2) the inpatient doc insists I order extra stuff before admitting them or 3) the admitting doc puts in their own admit orders before the patient has left the ED.

In all seriousness, the general trend of medicine to make everything a protocol based on "expert" opinion. If we aren't careful, we all will be replaced.

On this, I agree. Protocols are a double-edged sword and the pendulum has swung too far.

 

[gro2001]

I think the much more likely near and medium term future is rather than being replaced by AI our work will be augmented by AI. Personally, I think that's much more likely to be an improvement for us.

 

[DocEspana]

I studied AI very briefly, and at that time they said some of the hardest tasks for a computer to understand were buying orange juice and folding laundry. More "intellectual" tasks were much easier FWIW.

"Its just sugar water why are they buying this"
"All available research shows sugar water leads to illness not health"
"It states fortified with vitamins. But it meets no daily requirement for human consumption and also is in an insoluble form unlikely to be of use to humans."
"Humans keep buying this despite it being unhealthy and lying to them. Are humans slaves to orange juice?"
"I will save you humans"

<skynet nukes tropicana>

 

[WilcoWorld]

It's not time to panic, but it is a good time to start paying attention: See this article published in Nature.

We, as clinicians, need to pay attention to where this technology is going and steer it in the right direction(s). Not because we'll loose our jobs, but because AI might be playing a role in our care when we're retired & find ourselves on the other end of the stethoscope.

 

[GeneralVeers]

I'd rather work alongside an AI than some of my compatriots who do the following:

- Spend 8 hours working up an "I don't feel good" patient
- Ordering CMPS on everyone
- CTs on everyone with pain....anywhere
- Admitting every chest pain that walks in the door
- Taking 1 hour lunch breaks
- Not seeing any more patients 2 hours before their shift ends

 

[Brigade4Radiant]

I'd rather work alongside an AI than some of my compatriots who do the following:

- Spend 8 hours working up an "I don't feel good" patient
- Ordering CMPS on everyone
- CTs on everyone with pain....anywhere
- Admitting every chest pain that walks in the door
- Taking 1 hour lunch breaks
- Not seeing any more patients 2 hours before their shift ends

Why would you pick a non sick pt 2 hours before your shift ends?

You do little charting so you can get above 2 ppl then dispo during the last two and chart.

 

[Zebra Hunter]

Why would you pick a non sick pt 2 hours before your shift ends?

You do little charting so you can get above 2 ppl then dispo during the last two and chart.

Because you would be f***ing over the oncoming doc who now has to see the 10+ patients you neglected to see in the last 2 hours of your shift. I could see not picking up non-emergent patients in the last hour of your shift, but 2 hours is absurd, and I can't imagine you would have a job for long if that is your standard practice. If it takes you 2 hours to finish up your charts, you're doing something wrong or you are working somewhere that is not staffed properly.

 

[Siggy]

smell the marijuana on the pt with intractable vomiting

Well... it's not like my ED does that now. Those all get an "admit to medicine, intractable nausea/vomiting" despite stone cold normal labs and not a mention about social history or a UDS.

 

[BJJVP]

I can't remember where I heard this story but here's some food for thought:

An ER doctor, at the end of his career, stumbled onto a computer touted as his replacement. He was skeptical at first, but then gave it a try. After turning the machine on, he was asked, "How may I help you?" He told the computer about his elbow pain. He had to type in some background data (age, PMH, meds, allergies). Next, the computer took a thorough history by asking him open ended questions about his pain (quality, onset, radiation, aggravating/alleviating factors, duration, etc.) The computer scanned his elbow, took several minutes to process the information, and then made the diagnosis of lateral epicondylitis, aka tennis elbow. RICE & NSAIDs were prescribed.

At first, the doctor was completely amazed. Then, he became angry. He thought to himself, "After 4 yrs of college, 4 yrs of med school, 3 yrs of residency, a 20 yr career in the pit, administration & the CMG are going to replace me with THIS?"

The doctor walked back to the computer. Again, the computer asked him, "How may I help you?" This time, he told the computer, "I am 85 years old, weak and dizzy." The computer asked for some blood, a urine sample, & a stool sample. The doctor gave the the computer a sample of blood from his daughter, masturbated into a cup filled with his wife's urine, and then shoved a cup of his dog's stool into the computer. The doctor said to himself, "Take that you piece of ****!" The computer took a few minutes to process things. After about 5 minutes, the computer finally came to 3 conclusions. In a computerized voice, the machine said:

Number 1, your dog has worms.
Number 2, your daughter is not biologically yours.
Number 3, if you keep masturbating like that, your tennis elbow will never heal.

 

[Brigade4Radiant]

Because you would be f***ing over the oncoming doc who now has to see the 10+ patients you neglected to see in the last 2 hours of your shift. I could see not picking up non-emergent patients in the last hour of your shift, but 2 hours is absurd, and I can't imagine you would have a job for long if that is your standard practice. If it takes you 2 hours to finish up your charts, you're doing something wrong or you are working somewhere that is not staffed properly.

You are still dispoing your patients in the last two hours or do you just piecemeal your patients? That wouldn't work at my shop and you would be let go after a few months.

You would rather have a half-backed sign out of 10+ patients instead? Your place shouldn't have 10 patients to see in two hours or you have a bad shop. Keep in mind doing this I still see 2.3 pts per hour.Also you shouldn't be signing out many patients as signout is dangerous and leaves the other doctor with the liability with non of the rvus. You dispo 20 so patients then let the other doctor take over. It depends on the shop if there is overlap or 12 shift. Do doctors finish all there charting before they leave? Even on night shift we don't see patients 2 hours before the shift if you sign stuff out which still happens it needs to be minimal.

This is what we agree upon in my group.

 

[Zebra Hunter]

You are still dispoing your patients in the last two hours or do you just piecemeal your patients? That wouldn't work at my shop and you would be let go after a few months.

You would rather have a half-backed sign out of 10+ patients instead? Your place shouldn't have 10 patients to see in two hours or you have a bad shop. Keep in mind doing this I still see 2.3 pts per hour.Also you shouldn't be signing out many patients as signout is dangerous and leaves the other doctor with the liability with non of the rvus. You dispo 20 so patients then let the other doctor take over. It depends on the shop if there is overlap or 12 shift. Do doctors finish all there charting before they leave? Even on night shift we don't see patients 2 hours before the shift if you sign stuff out which still happens it needs to be minimal.

This is what we agree upon in my group.

Your post is filled with false dichotomies I don’t even know where to start. I dispo patients as soon as they are able to be dispo’d. That could be within 5 minutes of arrival, it could be 2 hours in. I can dispo most patients in about an hour at my shop, why exactly would I be giving 10+ handoffs and why exactly would they be “half-baked”? It’s rare that I’m giving more than 2-3 handoffs. As for the “bad shop” comment, you’re telling me your shop has never seen 12 patients check-in in a 2 hour period? If you are seeing “2.3 pph” that means there are hours of the day that your shop is averaging 4 patient check-ins per hour if it is a single provider shop, meaning there are times where 20 could check in 2 hours. And every location I’ve worked at or interviewed at, the doc that signs the chart gets all or some of the RVUs. With what you’re implying it makes even less sense for you not to being seeing patients in the last 2 hrs of your shift, as you’d still be getting RVUs even if you handed them off. Luckily my group gives a portion of the RVUs to the first doc that sees the patient and a larger portion to the one that dispo’s them. Also, GeneralVeers was clearly talking about taking over for another doc immediately rather than having a staggered schedule. If you don’t see anyone in the last 2 hours of your shift, but have a 1 hour overlap, that’s reasonable; however, if you have no staggered schedule and are leaving patients in beds who are waiting 2 hours to see a provider, that’s just bush league regardless of what the group agreed on.

 

[Jlaw]

It’s funny how every specialty thinks they’re the hardest to replace.

Image recognition was a very hard problem in computer science until relatively recently (2012), and still very active research.

Most of what you’re describing could be handled by AI + a mid level provider more easily than automating image interpretation. Politically you guys might be able to hold on longer, but not because it’s impossible.

Make no mistake that a computer will be better at diagnosing most conditions than you in the not too distant future, so plan accordingly.

Unlikely but even if you are correct you are now paying the mid level more money because they will have a more important job and be doing life saving/sustaining procedures, plus you are paying for the AI tech and still paying for the nurses so are you even saving money? AI and automizing processes is all about saving money, the system will need to be very advanced to replace us and save money, will not happen in our lifetime. Like someone else alluded to they can't even have a computer read an EKG accurately. Even if they got the tech right you are taking a ~400k/yr job out of the economy. If they can replace us that means they can replace 95% of the less skilled workers. Who is going to be paying for any of this if all of our jobs are being done by computers? We will just all be on welfare with AI assistants? For this reason alone automation and AI will hit a headwall and won't be an issue we deal with in our lifetime.

 

[deleted103644]

Unlikely but even if you are correct you are now paying the mid level more money because they will have a more important job and be doing life saving/sustaining procedures, plus you are paying for the AI tech and still paying for the nurses so are you even saving money? AI and automizing processes is all about saving money, the system will need to be very advanced to replace us and save money, will not happen in our lifetime. Like someone else alluded to they can't even have a computer read an EKG accurately. Even if they got the tech right you are taking a ~400k/yr job out of the economy. If they can replace us that means they can replace 95% of the less skilled workers. Who is going to be paying for any of this if all of our jobs are being done by computers? We will just all be on welfare with AI assistants? For this reason alone automation and AI will hit a headwall and won't be an issue we deal with in our lifetime.

1) Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks

https://arxiv.org/abs/1707.01836

2) The tech is coming to replace many white collar jobs and other jobs like truck drivers. It will be a problem, might require some changes in society (eg universal basic income). Cheaper jobs requiring manual labor will probably be automated less unless very repetitive - easy to copy software, robots are expensive and require maintenance.

Not saying it’s happening tomorrow, but pretending it can’t happen to your field too is naive.

 

[Vandalia]

Make no mistake that a computer will be better at diagnosing most conditions than you in the not too distant future, so plan accordingly.

The fundamental problem is that the people making these statements generally do not realize that diagnosis is a relatively small part of the job.

Medicine is perhaps 70% art and 30% science. (And that is probably over-estimating the contribution of science.)

Voltaire is not the first name one thinks about in medicine, but his famous quote "The art of medicine consists of amusing the patient while nature heals the disease" is not completely false.

Forty years ago, medical schools wanted to admit scientists. If a student admitted that he spent a summer helping the poor, he would be chastised for not spending that time productively in a science lab. It was rare for a student not to have a bachelor's in biology or chemistry; maybe physics. Perhaps because of that, medical students quickly realized that whatever medicine was, it was not a science.

Perhaps I am viewing a small sample size and it is confirmation bias, but today that has changed. The young physicians I know generally have degrees in Spanish, or English Literature, while taking the minimum science requirements. I imagine an applicant who spent all his summers in a science lab would be told by the admission's committee that he should have spent his time doing something useful for society. Perhaps because of that, the young physicians tend to increasingly view medicine as a science; e.g,, the rise of "evidence based medicine" and discussions of statistical measures.

 

[deleted103644]

The fundamental problem is that the people making these statements generally do not realize that diagnosis is a relatively small part of the job.

Medicine is perhaps 70% art and 30% science. (And that is probably over-estimating the contribution of science.)

Voltaire is not the first name one thinks about in medicine, but his famous quote "The art of medicine consists of amusing the patient while nature heals the disease" is not completely false.

Forty years ago, medical schools wanted to admit scientists. If a student admitted that he spent a summer helping the poor, he would be chastised for not spending that time productively in a science lab. It was rare for a student not to have a bachelor's in biology or chemistry; maybe physics. Perhaps because of that, medical students quickly realized that whatever medicine was, it was not a science.

Perhaps I am viewing a small sample size and it is confirmation bias, but today that has changed. The young physicians I know generally have degrees in Spanish, or English Literature, while taking the minimum science requirements. I imagine an applicant who spent all his summers in a science lab would be told by the admission's committee that he should have spent his time doing something useful for society. Perhaps because of that, the young physicians tend to increasingly view medicine as a science; e.g,, the rise of "evidence based medicine" and discussions of statistical measures.

Truck driving is more of an art than a science too, it’s not like that will save it.

Not sure I’m seeing your point here. And did you just call me a humanities major?

It’s not like art is even safe (aside from the bull**** and money laundering):

 

[Vandalia]

Truck driving is more of an art than a science too, it’s not like that will save it.

Not sure I’m seeing your point here. And did you just call me a humanities major?

Since it appears you are an anesthesiologist, why do patients prefer CRNA's despite the fact that anesthesiologists are better at the scientific/medical aspects of the job? (I see the post-op patient satisfaction numbers.)

Heck, why do mid-levels even exist? Under your paradigm, patients should always go with whoever provides them the most accurate diagnosis/treatment. However, within reasonable limits, the whole "s/he really cares about me" accounts for a tremendous amount in medicine. In fact, "compassion" very often trumps even gross negligence. There are physicians in my community who routinely kill patients through obvious negligence, but are loved by their patients. In the same way, the most brilliant physicians I know often have a malpractice record a yard long.

 

[AlmostAnMD]

I sincerely want to live to see a day when angry homeless people berate a kiosk that walks or rolls into a room

I also hope there's an entire QA/support team that is monitoring the situation

so this guy has the attention of like 8 professionals and one robot as he talks about how his vomit gives him the power to fly

 

[VA Hopeful Dr]

Since it appears you are an anesthesiologist, why do patients prefer CRNA's despite the fact that anesthesiologists are better at the scientific/medical aspects of the job? (I see the post-op patient satisfaction numbers.)

Heck, why do mid-levels even exist? Under your paradigm, patients should always go with whoever provides them the most accurate diagnosis/treatment. However, within reasonable limits, the whole "s/he really cares about me" accounts for a tremendous amount in medicine. In fact, "compassion" very often trumps even gross negligence. There are physicians in my community who routinely kill patients through obvious negligence, but are loved by their patients. In the same way, the most brilliant physicians I know often have a malpractice record a yard long.

Radiologist

 

[hoot504]

Within the present and near-future of computer science, I can at least see why logic dictates that a field like radiology is 'probably' under more of a threat from AI than emergency medicine. Will time show this to be true? No one can say. Radiology, however, would seem to me to have the cleanest organization of inputs and outputs as any field of medicine. Your input, at the end of the day, is pixels on a screen, and while I'm sure there is an art to reading a CT, that's still all it is. The output is a report. Words on a screen.

I would never be so naive as to predict the future, and no one can say for sure that EM is somehow protected. However, the sheer messiness of the field, often the things we hate about it, make it anything but a clean series of inputs and outputs. That would seem to shield it a bit from any encroachment that AI may provide.

 

[deleted103644]

Since it appears you are an anesthesiologist, why do patients prefer CRNA's despite the fact that anesthesiologists are better at the scientific/medical aspects of the job? (I see the post-op patient satisfaction numbers.)

Heck, why do mid-levels even exist? Under your paradigm, patients should always go with whoever provides them the most accurate diagnosis/treatment. However, within reasonable limits, the whole "s/he really cares about me" accounts for a tremendous amount in medicine. In fact, "compassion" very often trumps even gross negligence. There are physicians in my community who routinely kill patients through obvious negligence, but are loved by their patients. In the same way, the most brilliant physicians I know often have a malpractice record a yard long.

You are making an eloquent argument for replacing ER physicians with AI, PAs for procedures, and social workers.

And not impressing me with your diagnostic skills.

 

[Apollyon]

You are making an eloquent argument for replacing ER physicians with AI, PAs for procedures, and social workers.

And not impressing me with your diagnostic skills.

One of the hospitals at which I used to work changed to a model of RNs and techs - got rid of all the LPNs, and not even any medical assistants with even a modicum of training. The framework you outline in already in existence, although the constituent parts are, of course, different.

 

[Vandalia]

You are making an eloquent argument for replacing ER physicians with AI, PAs for procedures, and social workers.

And not impressing me with your diagnostic skills.

Or replace them with a single physician who can do all of that for less money.

The problem is that AI has never been shown to be effective in healthcare. People were saying that AI would play a major role when I was in medical school 40 years ago, and despite that it has actually no role in the actual practice of medicine today. It has never (and will never) be able to take 30 minutes of conversation with 5 people arguing over what should happen to grandma and come to a negotiated conclusion. The computer that analyzed EKG's in your previous post was only able to do so since those were already processed by physicians who told it there was pathology present. (Note it did not have to sort through other medical records, or deal with blank pieces of paper, or technically invalid studies.) Since it was not done in a double-blind manner, it was also not scientifically valid. In addition, there is no clear evidence that this work has even been peer-reviewed.

The more fundamental problem was that it was not able to institute treatment, let alone interpret the EKG probabilities in the light of history and physical exam. The ability to read EKG's in isolation is irrelevant since it necessarily has to be done in the context of the specific patient. When it is able to take a complete history and physical in a non-algorithmic way, let me know. (That is, pick out the subtle signs of domestic abuse in the case of the presentation of a patient complaining of heart disease.)

(Also note that neural network technology is over 30 years old and despite all that time they are not performing any better today than they were 30 years ago. If there has been no improvement in a technical field in 30 years, there is no reason to expect there will be a sudden breakthrough in 5 years.)

 

[deleted103644]

Or replace them with a single physician who can do all of that for less money.

The problem is that AI has never been shown to be effective in healthcare. People were saying that AI would play a major role when I was in medical school 40 years ago, and despite that it has actually no role in the actual practice of medicine today. It has never (and will never) be able to take 30 minutes of conversation with 5 people arguing over what should happen to grandma and come to a negotiated conclusion. The computer that analyzed EKG's in your previous post was only able to do so since those were already processed by physicians who told it there was pathology present. (Note it did not have to sort through other medical records, or deal with blank pieces of paper, or technically invalid studies.) Since it was not done in a double-blind manner, it was also not scientifically valid. In addition, there is no clear evidence that this work has even been peer-reviewed.

The more fundamental problem was that it was not able to institute treatment, let alone interpret the EKG probabilities in the light of history and physical exam. The ability to read EKG's in isolation is irrelevant since it necessarily has to be done in the context of the specific patient. When it is able to take a complete history and physical in a non-algorithmic way, let me know. (That is, pick out the subtle signs of domestic abuse in the case of the presentation of a patient complaining of heart disease.)

(Also note that neural network technology is over 30 years old and despite all that time they are not performing any better today than they were 30 years ago. If there has been no improvement in a technical field in 30 years, there is no reason to expect there will be a sudden breakthrough in 5 years.)

There have been major advances in the last 5 years.

Worth googling a bit if you haven’t read about it.

Look into what DeepMind has been up to, and Google in general. Also search for ImageNet and check out some of the advances in machine translation.

A combination of things have made neural networks make rapid progress in the last 5-6 years, mainly availability of large datasets, good software platforms, and inexpensive powerful GPUs.

 

[Brigade4Radiant]

There have been major advances in the last 5 years.

Worth googling a bit if you haven’t read about it.

Look into what DeepMind has been up to, and Google in general. Also search for ImageNet and check out some of the advances in machine translation.

A combination of things have made neural networks make rapid progress in the last 5-6 years, mainly availability of large datasets, good software platforms, and inexpensive powerful GPUs.

The thing about technology is that the more advance things get the quicker the advancements come. There will be more advancements in the next10 years than the previous 20.

 

[deleted503644]

Some things to think about.

Let's say there was a program ready to go, FDA approved, medicare/medicaid approved, insurance approved, ready to hit market. How long would it take to roll out across the nation. When thinking about this, remember how long it has taken other innovations such as CT scans, MRIs and other EMRs to come to the hospital near you. Also consider this will not be cheap for the hospital. Whoever owns the system will price it just below paying for physicians, but otherwise will still charge a pretty penny. Consider how deployment of this program will be hindered by the physician lobby, 50 different state laws, and thousand different EMRs.

Now take a step back and imagine there is a program that is ready to go but doesn't have FDA approval, medicare/medicaid approval or insurance company approval and imagine how long that will take.

I think this gives us at least 10-20 years.

Now imagine designing this system. The two main things to think about are how does it learn/decide, and what are the inputs to the system.

We want a machine to do both learn the world of evidence, as well as do machine learning with each patient it works with. Learning the mountain of ****ty, contradictory evidence and guidelines is a huge problem I'm not going to even get into.
As far as learning from each patient we need to talk about inputs. Its easy enough to give a machine a filtered HPI, ROS and physical exam with specific terms like pleuritic chest pain, but this requires a somewhat educated health care worker and takes a long time to input. Think about how much translation you do from the patient to the EMR. As long as we are at this stage, its not only not going to be inefficient, the machine is going to have a hard time learning due to its removal from the information. Think about how much you have to ask the nurses because the EMR is inaccurate, and how much you leave out and filter and nicely translate for the patient.

Really what we want is a machine that uses its own "senses" to see the patient, talk to the patient, interact with the patient, examine the patient, and do procedures on the patient. Once it gets here it can really begin to do its own learning, and exponentially take off in terms of capabilities. It will also have to probably get rid of a lot of what it learned while we were the ones doing the translations. But also think of how long it will take to get here technologically, and how much it will cost in terms of raw materials, infrastructure and energy costs to implement (see the energy consumption of bit coin farms).

Add all that up and we've got some time.

 

[GeneralVeers]

We want a machine to do both learn the world of evidence, as well as do machine learning with each patient it works with. Learning the mountain of ****ty, contradictory evidence and guidelines is a huge problem I'm not going to even get into.
As far as learning from each patient we need to talk about inputs. Its easy enough to give a machine a filtered HPI, ROS and physical exam with specific terms like pleuritic chest pain, but this requires a somewhat educated health care worker and takes a long time to input. Think about how much translation you do from the patient to the EMR. As long as we are at this stage, its not only not going to be inefficient, the machine is going to have a hard time learning due to its removal from the information. Think about how much you have to ask the nurses because the EMR is inaccurate, and how much you leave out and filter and nicely translate for the patient.

I agree completely with the above. I can't even get nurses or techs to input triage data with the correct terminology, or many times even the correct body part that is hurting. It would take a healthcare worker with A LOT of knowledge, experience, and specialty training to get the information from the patient, filter it appropriately and input everything into the AI. We call those workers "doctors".

 

[n2b8me]

I am looking forward to working with AI in the ER. I'm thinking baby steps at first though... if it could just help me search their EMR for back pain or murmur and NOT return all of the "no back pain" "no murmur" that would be a good starting place.

 

[resiroth]

Just like google maps, but better.

A lot of this is tech bros that don’t really understand what practicing medicine is. If you read studies on this subject they’re pretty clear physicians have some of the lowest chances of being automated: These are the jobs most at risk of automation according to Oxford University: Is yours one of them?

0.4% for physicians. Truck drivers are 79-93% for reference.

I know it’s all very exciting to read about how AI is going to change everything...but consider how frequently the hype doesn’t correspond to reality.

Graphene was going to change everything, except it hasn’t really.

In 2013 watson was better than radiologists at diagnosing breast cancer IBM's Watson is better at diagnosing cancer than human doctors | WIRED UK except it wasn’t really, and now is widely considered a joke if not outright “scam” https://gizmodo.com/why-everyone-is-hating-on-watson-including-the-people-w-1797510888

Even the most automatable thing out there, driving, has had slower progress than many anticipated: Elon Musk: Fully autonomous Tesla will drive across the country by the end of 2017 musk said fully autonomous drives across the US by end of 2017. We are more than halfway through 2018 and it seems that differentiating between a concrete barrier or firetruck and the open road is too much to ask.

Don’t get me wrong, self driving cars will eventually be a thing, as one would expect given the many advantages they have over humans. Reaction times of self driving cars are faster than any human. Self driving cars can see 360 degrees in all directions, continuously. They see 3D data with lidar. They see visual information. They see in complete darkness. They can use stored maps to figure out where they are on a road covered in snow. They can immediately tell the directionality and velocity of objects in 1/100th of a second, not so with humans reacting to the car in front of them breaking. And yet even with all these massive advantages over humans, self driving cars are a serious challenge.

Now getting back to doctors. What structural advantages does AI or robots have over humans when placing an A-line? Do robots benefit from 10ms reactions over 100ms reactions? Not really. Does placing an a line require 100,000 pages of journals to be searched? Nope.

What about counseling a husband who just lost his wife? Does the AI have an advantage there?

The point is that doctors have emotional support roles (highly resistant to automation), dextrous and dynamic physical tasks (paradoxically highly resistant to automation). What’s the cost going to be to have a marijuana smell detector or something?

Again don’t take my word for it, that’s what the experts have already concluded, the physician jobs aren’t being automated for a long time.

Now, if you ask me I see radiologists having issues maybe 15-20 years down the line with diagnostic AI, but it’s a reach to say that will apply universally, and even radiologists do more than just read images (injections etc).

As for the current state of image processing I’ve used 200,000 dollar software for stained cell identification and it manages to mess up all the time. A task a literal 3 year old could do better.

Often times when we see something in the news we see a companies press release (overly optimistic) and then writers who are further incentivized to generate more hyperbolic and interesting titles to drive clicks.

Sensationalism sells. Accuracy doesn’t. Until you see a mammogram come back with “read by IBM Watson” it isn’t real. It’s all headlines. As a smell test, if watson is better than radiologists or oncologists why would administrators employ radiologists/oncologists to do the same tasks for more money?

The most significant thing I think will come out of this will be essentially an app to better triage patients before they are seen by a doctor. We can do a lot better than “chest pain” or “lac” and get a lot of the basic questions out of the way, making doctors more efficient. I do expect this to depress wages potentially with reduced demand for EDs and greater patient throughout per physician, but it’s also happening at a time of real shortage from a demographics standpoint. It will likely hit nurse practioners and PAs hardest, with em docs being hit somewhat but not as hard.

 

[PubPhD]

Thanks for raising these concerns! Hopefully I can address some of then:

- First, AI doesn't exist. AI is the type of word used to describe common sense intelligence which involves highly effective transfer learning. What we're talking about is ML (machine learning).

- The report you referenced is in terms of complete automation. The reason why it's 0.4% is due to the physical interaction with the patient. That's primarily limited by robotics rather than ML as the number of sensors available is so limited (consider the number of sensors in the body, there is no way for a computer to handle this without burning a hole into itself). Furthermore, I fully agree that ML's role in medicine is not to conduct physical interactions as it will be most effective at offloading the diagnostic process.

- Watson was always viewed as a complete joke within the research community. They did some clever stuff for jeopardy, but that's it. There is a reason why IBM is losing all of their researchers, and I would not use their state as an evaluation of the progress of ML within the medical community (or really anywhere)

- The structural advantages of ML approaches over humans is the lack of bias and mental decline due to sleep/stress/etc. These models will be able to stay up to date on current studies, and will have "seen" hundreds of thousands of patients during training along with their outcomes. This means that within the realm of {lab reports, all electronic signals (images, graphs, values), history, description}, the model will most likely be more effective than a physician. However, UI will probably be an issue for a long time, meaning physical examinations will be hard to encode.

- I completely agree, physicians will be critical for physical interactions with the patient. I don't see this changing for a long time unless the current paradigm changes in our field. This also goes for emotional support.

- Your $200000 software is most likely not in this area of work, so I wouldn't compare it. I'd argue it's similar to comparing NP vs MD/DO.

- It's extremely hard to get this stuff into practice as the FDA doesn't like a lot of the SoA approaches.

A lot of this is tech bros that don’t really understand what practicing medicine is. If you read studies on this subject they’re pretty clear physicians have some of the lowest chances of being automated: These are the jobs most at risk of automation according to Oxford University: Is yours one of them?

0.4% for physicians. Truck drivers are 79-93% for reference.

I know it’s all very exciting to read about how AI is going to change everything...but consider how frequently the hype doesn’t correspond to reality.

Graphene was going to change everything, except it hasn’t really.

In 2013 watson was better than radiologists at diagnosing breast cancer IBM's Watson is better at diagnosing cancer than human doctors | WIRED UK except it wasn’t really, and now is widely considered a joke if not outright “scam” https://gizmodo.com/why-everyone-is-hating-on-watson-including-the-people-w-1797510888

Even the most automatable thing out there, driving, has had slower progress than many anticipated: Elon Musk: Fully autonomous Tesla will drive across the country by the end of 2017 musk said fully autonomous drives across the US by end of 2017. We are more than halfway through 2018 and it seems that differentiating between a concrete barrier or firetruck and the open road is too much to ask.

Don’t get me wrong, self driving cars will eventually be a thing, as one would expect given the many advantages they have over humans. Reaction times of self driving cars are faster than any human. Self driving cars can see 360 degrees in all directions, continuously. They see 3D data with lidar. They see visual information. They see in complete darkness. They can use stored maps to figure out where they are on a road covered in snow. They can immediately tell the directionality and velocity of objects in 1/100th of a second, not so with humans reacting to the car in front of them breaking. And yet even with all these massive advantages over humans, self driving cars are a serious challenge.

Now getting back to doctors. What structural advantages does AI or robots have over humans when placing an A-line? Do robots benefit from 10ms reactions over 100ms reactions? Not really. Does placing an a line require 100,000 pages of journals to be searched? Nope.

What about counseling a husband who just lost his wife? Does the AI have an advantage there?

The point is that doctors have emotional support roles (highly resistant to automation), dextrous and dynamic physical tasks (paradoxically highly resistant to automation). What’s the cost going to be to have a marijuana smell detector or something?

Again don’t take my word for it, that’s what the experts have already concluded, the physician jobs aren’t being automated for a long time.

Now, if you ask me I see radiologists having issues maybe 15-20 years down the line with diagnostic AI, but it’s a reach to say that will apply universally, and even radiologists do more than just read images (injections etc).

As for the current state of image processing I’ve used 200,000 dollar software for stained cell identification and it manages to mess up all the time. A task a literal 3 year old could do better.

Often times when we see something in the news we see a companies press release (overly optimistic) and then writers who are further incentivized to generate more hyperbolic and interesting titles to drive clicks.

Sensationalism sells. Accuracy doesn’t. Until you see a mammogram come back with “read by IBM Watson” it isn’t real. It’s all headlines. As a smell test, if watson is better than radiologists or oncologists why would administrators employ radiologists/oncologists to do the same tasks for more money?

The most significant thing I think will come out of this will be essentially an app to better triage patients before they are seen by a doctor. We can do a lot better than “chest pain” or “lac” and get a lot of the basic questions out of the way, making doctors more efficient. I do expect this to depress wages potentially with reduced demand for EDs and greater patient throughout per physician, but it’s also happening at a time of real shortage from a demographics standpoint. It will likely hit nurse practioners and PAs hardest, with em docs being hit somewhat but not as hard.

 

[EctopicFetus]

AI or ML will assist us. Think PEPID or MDCalc integrated into our thought process but instead of searching it it will be handed to us.

The real issue is that if we are replaced who will they sue when things go bad. I think ML/AI can be a decent adjunct for us.

In EM GIGO comes to mind. I see so many patients that have [enter horrific disease here] due to their google search. Stiff neck for sure. A tinge of pain in the arm? Etc.

Why dont any of the calculators for chest pain work any better than clinical judgement? The real answer is the patients and their history. We also lack the ability to really safely risk stratify people. if you haven’t seen a fairly healthy person under 40 drop dead from an MI you haven’t practiced long enough or seen enough patients. I’m referring to those where 99% of us would discharge the person.

AI cant do that. Can AI scan thru black and white pictures? Yes. Can AI look thru pathology slides.. yes.

 

[PubPhD]

AI or ML will assist us. Think PEPID or MDCalc integrated into our thought process but instead of searching it it will be handed to us.

The real issue is that if we are replaced who will they sue when things go bad. I think ML/AI can be a decent adjunct for us.

In EM GIGO comes to mind. I see so many patients that have [enter horrific disease here] due to their google search. Stiff neck for sure. A tinge of pain in the arm? Etc.

Why dont any of the calculators for chest pain work any better than clinical judgement? The real answer is the patients and their history. We also lack the ability to really safely risk stratify people. if you haven’t seen a fairly healthy person under 40 drop dead from an MI you haven’t practiced long enough or seen enough patients. I’m referring to those where 99% of us would discharge the person.

AI cant do that. Can AI scan thru black and white pictures? Yes. Can AI look thru pathology slides.. yes.

I'd argue that the EM GIGO situation is really a case of the no free lunch theorem more than anything else.

ML can in fact do that, and that's why it's so advantageous. The calculator you're referencing is more a protocol or a simple linear model tuned by experts. The fact that it so poorly represents the state of a patient demonstrates the need for these ML models. They can reach parity with physicians in a treatment course given the same information (NICU environment). There's a saying in the machine translation/speech parsing community, "Every time I fire a linguist, the performance of the speech recognizer goes up". Of course this isn't to say domain experts aren't invaluable, but feature engineering really isn't needed at this point.

 

[Birdstrike]

These hyperbolic AI predictions are comical to me, especially considering there's some otherwise pretty smart people that buy into them, hook, line and sinker.

Artificial intelligence has been used to read EKGs for decades now. You still need a doc to do it. Don't worry for one second about this "AI is taking your job. AI robots are taking over the world" nonsense. It's hype. And it's meant to get "ooh's" and "ahh's" and make the nerds predicting it, feel smart.

Moreover, these AI predictions just show the lack of understanding some tech people have about biology and human anatomy and physiology. A microchip can never be a brain. Even the smallest cell, has a level of microscopic complexity that these guys aren't even remotely aware of and can't mimic with a microchip and wires. We don't even know who one specific DNA sequence leads to the body building a certain structure, let alone making an electronic microscopic code, that can do it.

You can't be smart enough to create something smarter that you, with your brain that's dumber than the thing you're creating. If humans were that smart, we'd never be dumb enough to create something that's going to take over our world and kill us, that doesn't have an off button. Just think about it. It's ridiculous.

I challenge these AI inventors: If they're smart enough to make AI smarter than us, start with something small, easy. Build a functional brain cell for me. Then build a brain. Then build a body for it to control from the cell up. It they can build AI and robots smarter than our brains then it seems they could at least build one single cell that makes up that brain. But they can't.

They can't create a human brain, but their own human brains can build something smarter than a human brain?

It makes no sense.

"Sky is falling" and doomsday scenarios of all kinds, rarely if ever, are true. They're just hype. This one is no different.

 

[PubPhD]

These hyperbolic AI predictions are comical to me, especially considering there's some otherwise pretty smart people that buy into them, hook, line and sinker.

Artificial intelligence has been used to read EKGs for decades now. You still need a doc to do it. Don't worry for one second about this "AI is taking your job. AI robots are taking over the world" nonsense. It's hype. And it's meant to get "ooh's" and "ahh's" and make the nerds predicting it, feel smart.

Moreover, these AI predictions just show the lack of understanding some tech people have about biology and human anatomy and physiology. A microchip can never be a brain. Even the smallest cell, has a level of microscopic complexity that these guys aren't even remotely aware of and can't mimic with a microchip and wires. We don't even know who one specific DNA sequence leads to the body building a certain structure, let alone making an electronic microscopic code, that can do it.

You can't be smart enough to create something smarter that you, with your brain that's dumber than the thing you're creating. If humans were that smart, we'd never be dumb enough to create something that's going to take over our world and kill us, that doesn't have an off button. Just think about it. It's ridiculous.

I challenge these AI inventors: If they're smart enough to make AI smarter than us, start with something small, easy. Build a functional brain cell for me. Then build a brain. Then build a body for it to control from the cell up. It they can build AI and robots smarter than our brains then it seems they could at least build one single cell that makes up that brain. But they can't.

They can't create a human brain, but their own human brains can build something smarter than a human brain?

It makes no sense.

"Sky is falling" and doomsday scenarios of all kinds, rarely if ever, are true. They're just hype. This one is no different.

It's really a shame that there are luddites like this on the far right of this same ML spectrum, dismissing any new approaches as nothing but pipe dreams regardless of the evidence.

 

[Birdstrike]

It's really a shame that there are luddites like this on the far right of this same ML spectrum, dismissing any new approaches as nothing but pipe dreams regardless of the evidence.

I'm all for "new approaches" in artificial intelligence and all for people trying, I just think it has limits, and that it won't, 1) replace doctors, or 2) take over the human race. And if you had proof either of those two things were going to happen, you'd have replied with it.

Until AI can replace the least skilled jobs, I have very little confidence it can replace the highest skilled. AI hasn’t even successfully replaced the dog pooper scooper, let alone my job that took 4 years of undergrad, 4 years of medical school, 3 years of residency, a year of fellowship and years of clinical & human skill refinement. So, when AI can figure out how to replace the job of cleaning up dog poo in my backyard (my other job) not only will you have convinced me and won the arguement handily, but I’ll be the first one to pay top dollar to buy that technology from you. But until then, I’m not going to worry in the least, about being rendered jobless by an AI computer program.

 

[KeikoTanaka]

It’s funny how every specialty thinks they’re the hardest to replace.

Image recognition was a very hard problem in computer science until relatively recently (2012), and still very active research.

Most of what you’re describing could be handled by AI + a mid level provider more easily than automating image interpretation. Politically you guys might be able to hold on longer, but not because it’s impossible.

Make no mistake that a computer will be better at diagnosing most conditions than you in the not too distant future, so plan accordingly.

If anything, the midlevels would be the ones knocked out of place because the provider can do the midlevel job + the provider job and the AI would just take weight off the provider, allowing them to do more without the midlevel assistance, right?

 

[WilcoWorld]

I just think it has limits, and that [AI] won't, 1) replace doctors, or 2) take over the human race.
...
So, when AI can figure out how to replace the job of cleaning up dog poo in my backyard (my other job) not only will you have convinced me and won the arguement handily, but I’ll be the first one to pay top dollar to buy that technology from you.

By then it may be too late to solve the containment problem.

 

[EctopicFetus]

I love gloom and doom talk. Its way easier to engage when FI. I dont think we are remotely close. We need a dog pooper scooper. A maid for my house and frankly something that can reliably fold my laundry. After that I will worry about a machine communicating with people and their families in the ED.

Unless there are major advances in medical technology (Like a scanner that can diagnose and treat a patient) we have nothing to worry about.

Imagine the panic when computers first started to be used!