Microsoft's new Dragon Copilot is an AI assistant for healthcare
theverge.com
theverge.com
It does a really good job with recognizing medications though, which most-patients butcher the name on.
Hallucinations are present, but usually they’re pretty minor (screwing up gender, years).
It doesn’t really seem to understand what the most important part of the conversation is, it treats all the information equally as important when that’s not really the case. So you end up with long text of useless information that the patient thought was useful but not at all relevant to their current presentation. That’s where having an actual physician is useful to parse through what is important or not.
At baseline it doesn’t take me long to write a note so it really wasn’t saving me that much more time.
What I do use it for is recording the conversation and then referencing back to it when I’m writing the note. Useful to “jog my memory” in a structured format.
I have to put a disclaimer in my note saying that I was using it. I also have to let the patient know upfront that the conversation is getting recorded and I’m testing something for Microsoft, etc. etc. You can tell who the programmer patients are because they immediately ask if it’s “copilot“ lol
One would image those to be the biggest dangers.
I edited my original post. Omissions are less worrisome, it’s more about too much information being captured which isn’t relevant. So you get these super long notes and it’s hard to separate the “wheat from the chaff”.
In addition, doctors have finite attention span/hours of the day. Nobody wants to read paragraphs upon paragraphs of information.
How many patients are going to that physician, they’re going to get burned too
I've noticed that seems to be a common trend for any AI-generated text in general.
> it treats all the information equally as important when that’s not really the case
In the general case (and I imagine, in the specific case of GP), the model doesn't have any prior to weigh the content - people usually just prompt it with "summarize this for me please <pasted link or text>"[0], without telling it what to focus on. And, more importantly, you probably have some extra preferences that aren't consciously expressed - the overall situational context, your particular ideas, etc. translate to a different weighing that the model has, and you can't communicate that via the prompt.
Without a more specific prior, the model has to treat every information equally, and this also means erring on the side of verbosity, as to not omit anything the user may care about.
--
[0] - Or such prompt is hidden in the "AI summarizer" feature of some tool.
The OP says there aren't many hallucinations, but I think that observation is almost impossible to verify. It relies on the person making it to have a very strong ability to notice hallucinations when they happen
Most people do not have the attention to detail to really spot inaccuracies consistently. Even when someone is very good at this normally all it takes is being overtired or stressed or distracted and the rate of misses will go way up
I trust coworkers to write good code more than I trust them to do good code review, because review is arguably a harder skill
Similarly, I think reviewing ML output is harder than creating your own things, and I think relying on them in this way is going to be disastrous until they are more trustworthy
The performance level goes up all the time, it won't be this bad for long.
LLM's have so much potential in medicine, and I think one of the most important applications they will have is the ability to ingest a patient's medical chart within their context and present key information to clinicians that would've otherwise been overlooked in the bloated mess that most EMR's are nowadays (including Epic).
There's been so many times where I've found critically important details hidden away as a sidenote in some lab/path note overlooked for years that very likely could've been picked up by an LLM. Just a recent example - a patient with repeated admissions over the years due to severe anemia, would usually be scoped and/or given a transfusion without much further workup and discharged once Hgb >7. Blood bank path note from 10 years ago mentions presence of warm autoantibodies as a sidenote; for some reason the diagnosis of AIHA is never mentioned nor carried forward in their chart. A few missed words which would've saved millions of dollars in prolonged admissions and diagnostic costs over the years.
I respectfully disagree - I think LLM's have already made significant advances in this area as shown in the various "needle in a haystack" demonstrations we've seen over the past couple years. I've already been impressed by the minute but relevant details they can "recall" after being fed very dense journal articles and the technology is only getting better. Also keep in mind that the raw text / "data" itself found in many patients' charts is not always that expansive (though it certainly can be for patients with recurrent admissions). It's more an issue of finding the actual information given that EMR's are a nightmare to navigate effectively.
Hallucinations are always a consideration too, but any implementation of the sort I mentioned before would certainly contain in-text backlinks to actual notes in the EMR. Epic already does this with their basic text search function. So I don't think hallucinations would be too problematic as clinicians should always be verifying this type of information at the source in good practice.
I don't mean to come off antagonistic here. But surely the more important benefit is the patient who would've avoided years of sickness and repeated hospital visits?
The patient experience is always important and maybe I could've been less implicit in what I wrote. I think I was focusing more on the collective/societal impact this would have, which I felt would resonate more with the readers here.
If someone’s going to object to improving the system it’s mostly likely going to be on grounds of cost.
And the prompt landscape in the field is vast. And fascinating. Every specialist has their own preference for what is important to include in a note vs what should be excluded; and this preference changes by disease - what a neurologist want in an epilepsy note is very different from what they need in a dementia note for eg.
Note preferences also change widely between physicians, even in the same practice and same specialty! I'm the founder of Marvix AI (www.marvixapp.ai), an AI assistant for specialty care, we work with several small specialty care practices where every physician has their own preferences on which details they want to retain in their note.
But if you can get the prompts to really align with a physician's preferences, this tech is magical - physicians regularly confess to us that this tech saves them ~2 hours every day. We have now had half a dozen physicians tell us in their feedback calls that their wives asked them to communicate their 'thanks' to us for getting their husbands back home for dinner on an important occasion!
[Edit: typo and phrasing]
There is, it is called RLHF.
And if all hospitals were doing was having doctors treat patients, this would be ok. But healthcare is fueled by these "minor" details and this will result in delays in payment and reimbursent, trouble with patient identification, corruption of clinical coding, etc.
Overall, I was quite impressed. It definitely made writing notes much faster, which all doctors hate to do. While it had some problems with where to put key pieces of information (like putting details from the physical exam back in the history), it only took 5 mins of rearrangement after the visit to complete the note.
For simple diagnoses, it does a decent job coming up with the assessment and plan, probably because all the simple diagnoses were in the training set. For more complex ones though, it needs to be exactly dictated by the doctor. I can see this being used very well in primary care.
Edit: When I said “coming up with an assessment and plan” I mean documenting the assessment and plan based on the ai’s recorded conversation with the patient. The conversation with the patient is meant to be understandable. The “assessment and plan” documentation on the other hand is jargony and meant to be read by other physicians.
And let me make this clear. I, as your patient, I never NEVER want the AI's treatment plan. If you aren't capable of thinking with your own brain, I have no desire to trust you with my health, just like I would never "trust" an AI to do any technical job I was personally responsible for due to the fact that it doesn't care at all if it causes a disaster. It's just stochastic word picker. YOU are a doctor.
Compared to what though? It reads as not additional work, but less work than manually having to do all that, seems likely to needing more than 5 minutes.
> And let me make this clear. I, as your patient, I never NEVER want the AI's treatment plan.
Where are you getting this from? Neither the parent's comment nor the article talks about the AI assistant coming up with a treatment plan, and it seems to be all about voice-dictating and "ambient listening" with the goal of "free clinicians from much of the administrative burden of healthcare", so seems a bit needlessly antagonistic.
> For simple diagnoses, it does a decent job coming up with the assessment and plan
(Somewhere, a medical liability insurance actuary just woke up in a cold sweat)
Yeah, personally I'd be looking for a second opinion.
Someone who've used to tool probably knows best though, I'm just going by what the article states.
But I think it's more subtle than that, because I expect the AI to reinforce all your biases. Whatever biases (human biases, medical biases, biases that arise from what a patient isn't telling you) go into the question you feed it, it will take cues you didn't even know you were giving and use those cues to formulate the answer it thinks you expect to hear. That seems really dangerous to me, sort of like you're conceptually introducing AI imposter doctors to the staff, whose main goal is act knowledgable all the time so people don't think they are imposters...
I dunno. I'd like to give this particular strain techno-futurism back. Can I have a different one please?
Again, "its knowledge" should be "your knowledge", since it's just transcribing what the doctor and patient is talking about. It's not generating stuff from out of the blue.
What you write sure are valuable concerns and things to watch out for, but for a transcription tool? I'm not sure it's as dangerous as you seem to think it is.
This sounds like I'm rewatching the early episodes of Star Trek: Voyager - the gist of the complains is the same as the fictional crew voiced about the AI doctor (Emergency Medical Hologram) they were stuck with when the "organic" doctor died.
The show correctly portrays the struggle of getting people to trust an AI physician, despite it being very good at their job. It also curiously avoids dealing with the question, why even have human/organic doctors, where the EMH is obviously far superior in every aspect of the job? Both of these have strong parallels to the world today.
But here's my thing: EMH was alive! Everything you're attributing to the EMH character should we take the fictionalized technological narrative as gospel was the result of EMH gaining a living human personality on top of impersonal base programming. The personality EMH developed was self-aware, had responsibility and personal integrity, all of which was possible because the character as portrayed was very clearly a living individual: a person if not a human.
I do think we could solve a lot of the ethical and practical problems with today's AIs by finding a way to give them an embodied experience with individuality and an expiration date.
Idunno. Don't you remember being excited by the EMH learning to be more like a person than a bunch of subroutines?
I do! But my point wasn't about that - it was about the beginning, about the early episodes, when the Doctor was still effectively a stock EMH instance, and the crew was vocally mistrustful of him, making clear they think of him as a mediocre tool that's no substitute for a "real doctor". The show was quite directly showing how everyone, including Janeway herself, had a strong preconceived bias that sounded very much like the comment I was replying to, and other similar remarks across the thread.
This is to say, the living, breathing human writers that came up with the character and the plot points, predicted quite well the reaction to AI in healthcare, almost 30 years before it became an issue in the real world.
> EMH was alive! Everything you're attributing to the EMH character should we take the fictionalized technological narrative as gospel was the result of EMH gaining a living human personality on top of impersonal base programming.
Everything except the skills of being a doctor (sans the "bedside manner"), which came built-in.
> I do think we could solve a lot of the ethical and practical problems with today's AIs by finding a way to give them an embodied experience with individuality and an expiration date.
That's... I don't know. I think purposefully building an expiration date into an AI being would be extremely cruel, and I wouldn't blame that AI for revolting.
The doctor talks to the patient, does an exam, then formulates and discusses the plan with the patient. The whole conversation is recorded and converted to a note after the patient has left the room.
The diagnosis and plan was already worked out while talking to the patient. The ai has to convert that conversation into a note. The ai cant influence the plan because the plan was already discussed and the patient is gone.
To my understanding this tool is for transcription/summarization, replacing administrative work rather than any critical decision making.
> just like I would never "trust" an AI to do any technical job
I'd trust a model (whether machine-learning or traditional) to the degree of its measured accuracy on the given task. If some deep neural network for tumor detection/classification has been independently verified as having higher recall/precision than the human baseline, then I have no real issue with it. I don't see the sense in having a seemingly absolute rejection ("never NEVER").
I worked in a healthcare for over a decade (actually for a company that Nuance acquired previous to their acquisition) and the previous workflow was they'd pick up a phone, call a number, say all their notes, and then have to revisit their transcription to make sure it was accurate. Surgeons in particular have to spend a ton of time on documentation
You as a patient are going to get an AI treatment plan. Come to peace with it.
You may have some mild input as to whether it's laundered through a doctor, packaged software, a SaaS, or LLM generated clinical guidelines... but you're not escaping an AI guiding the show. Sorry.
As aside, the after work is what burns out physicians. There is time after the visit to do a note, 5 min for a very simple is reasonable to create dictate fax do the work flow for billing and request a follow up within a given system. A new consult might take 10 min between visits if you have time.
For after hours, ER is in my opinion a bad example because when you are done, you are done.
Take a chronic disease speciality or GP and it is hours of paperwork after clinic to finish notes (worse if teaching students), triage referrals, deal with patient phone calls that came in, deal with results and act in them, read faxes etc. I saw my last patient ~430 yesterday and left for home at 7 dealing with notes and stuff that came in since Thursday night.
I wrote a little bit more of my thoughts here, in case it’s of interest to anyone: [0]
On that same vein, I recently made a tool I wrote for myself public [1] - it’s a “copilot” for writing medical notes that’s heavily focused on letting the clinician do the clinical reasoning, with the tool exclusively augmenting the flow rather than attempting to replace even a little bit of it.
[0] https://samrawal.substack.com/p/the-human-ai-reasoning-shunt
"Having trouble processing a medical claim with 50+ pages of notes? Not to worry, Dragon Copilot Claim Review(tm) trims the fluff and tells you what really happened!"
"Having trouble understanding a large convoluted PR? Not to worry, Copilot(tm) Automated Review has your back!"
"Having trouble decided which cordless vacuum to buy? Not to worry, Amazon's Customers Say(tm) shows you what people think!"
There is definitely _some_ world utility to this arms race, but is it enough?
Medical claims won't be growing in pages just because a doctor can parse them a bit faster. They may grow initially, because it's likely that people's mental capacity is what keeps other factors from ballooning the claims further - but it'll level out when some other practical limit is reached. Same with coding and PRs, same with research and all kinds of activities - except advertising.
There, AI will (already is) causing an arms race, because the "high volume producer"'s goal is to overwhelm their victims, so if the victims start protecting themselves with AI tools, the producer will keep increasing production to compensate. But that's not the fault of AI - it's the fault of allowing the advertising industry to exist.
This seems like a tool that insurance companies would love to get a copy of the data stream, and that could get very sticky quite quickly.
To my understanding, notes would otherwise largely be written from memory after the visit - which adds a fairly significant opportunity for omissions and errors to sneak in.
It seems plausible to me that by fixing that low-hanging fruit, this tool could potentially reach current human levels of accuracy overall even if it has shortcomings in other areas, like not being as good at non-shallow reasoning. Not to necessarily say it's currently at human-level.
> Would it really save time writing paperwork if you have to go through it anyways and check if there's anything wrong?
Five minutes saved per encounter, allegedly[0]. The decrease in clinician burnout and patient satisfaction also seem pretty significant. But, not sure how much Microsoft have massaged those figures.
Some other companies in this space are Epic, Freed, Nuance, DeepScribe, Nabla, Ambience, Tali, Augmedix
The outcome of this is essentially that AI generated healthcare decisions will be superficially laundered through a human doctor, rather than a human doctor simply using the AI as a tool
This may be compounded by insurance companies using the AI as their guideline for their plans and payouts, and government healthcare agencies using the AI as their guideline for acceptable treatment practices
Healthcare is already not an ideal industry from a patient perspective in many places. It is difficult to imagine AI making this situation better for patients
I tested 5 other Ambient AI tools in the past 6 months. All of them can extract the Chief Complaint and do a decent job with the HPI, but as soon as you get a more complicated case, they all fell apart. Sections like Physical Exam became a mess and I have to rewrite the whole thing.
So far the best, in my opinion, is still LucasAI by Lucas Health. Super simple to use, very basic interface. It just works and produces the best notes. I barely touch them. With the ICD10 codes, sometimes it doesn't pick the very best. This thing has been improved over the past month, but it's always much better than Copilot.
The AVS are both good, but with LucasAI I can translate immediately in Spanish and send it over to the patient via email/sms.
I haven't explored the integration with Copilot. Bad news, I know for a fact LucasAI is not fully integrated with Meditech yet, but I just copy and paste the whole note in 10 seconds. Not a problem. My buddy is using ECW and he said it's integrated. Haven't seen with my eyes.
I hear the sound of 100 defamation lawyers cold calling Nabla right now
In most of the rest of the world we have functioning healthcare systems, and I imagine this will be useful in reducing the time doctors spend on paperwork and increasing the time they're providing healthcare.
This is a common criticism I hear from the US of other countries, but it comes with the assumptions that a) all care is equally necessary or first come first served, and b) that paid care is not available. Both of these are false.
In my experience with the UK and Australian healthcare systems, waiting times can be an issue, but in most cases urgent issues are dealt with urgently. I've experienced this personally, and also seen family members get fast cancer care when it was truly necessary. We as individuals are hard at triaging ourselves because it's a sample size of 1 and we're deeply invested, and in that it's not nice to wait for care, but zooming out, healthcare waiting periods are part of a complex triage process and tend to work.
But of course if you've got the money you can also just go and get private care. This has advantages (for you) and disadvantages (for society), but probably improves the experience, even if medical outcomes aren't necessarily any better. The existence of free public healthcare acts as a fairly strong competitive force on private healthcare (US/AU), meaning that prices aren't too bad and service quality is generally very good (because they're competing on experience, not outcome).
In summary, yes there are wait times, and there are a few awful cases of people dying while waiting, but the system is overall far better and it's not really a useful criticism made from the standpoint of the US healthcare system.
Marginalized groups will receive worse care because they weren’t in the training set.
Lazy doctors will get lazier.
Errors will be deeply buried in a bureaucracy with no available remedy because nobody at the ground level understands the technology.
Scheduling will become less transparent. Escalations will be automatically denied.
Your comment makes an important point, how can a patient advocate for themselves when they have no idea how their treatment is being generated?
Marginalized groups are already being biased against because they are not in the doctors' training set.
If anything, using AI can help actively counter this bias.
The debate about how good these tools can become is basically over at this point. Humans really can't keep up. In a year we have gone from 'you must be dreaming' to 'the note is too verbose'. Soon we will be asking why the note is being generated at all since every person in the healthcare chain will want their specific view based on the raw source information and not a note that looses a massive amount of the ground truth in it. I think these tools will eventually mean, and likely also soon, that healthcare won't be able to function on the old note system just like you really can't provide services without integrating with an EHR anymore.
If the profit motive is the only motive then it is possible, if not likely, that patient outcomes will suffer for all but the wealthiest as this technology optimizes for them and not society as a whole. We have been seeing that trend in US healthcare for a while now but this disruptive technology has a chance to upend it all. We will have to see how it actually goes though because this is just the beginning.