#63 - When AI Writes Back To Patient Messages

WEBVTT
00:00:00.239 --> 00:00:02.640
AI can draft a patient message.
00:00:02.879 --> 00:00:06.240
But what happens when the provider has to edit it?
00:00:14.960 --> 00:00:19.920
Hello and welcome back to Code and Cure, where we discuss decoding health in the age of AI.
00:00:20.160 --> 00:00:22.000
My name is Vasant Sarathi.
00:00:22.160 --> 00:00:26.640
I'm an AI researcher and cognitive scientist, and I'm here with Laura Hagopian.
00:00:26.800 --> 00:00:29.359
I'm an emergency medicine physician and I work in digital health.
00:00:29.679 --> 00:00:39.280
You know, I was thinking about um this article that we're going to talk about, and also thinking about how sometimes when you type Google like Gmail, I don't know if you use Gmail, but like, yeah, and I do know if you use Gmail.
00:00:39.520 --> 00:00:40.479
I do, I do use Gmail.
00:00:40.880 --> 00:00:41.200
Never mind.
00:00:41.600 --> 00:00:47.280
But but um, you know, if people use Gmail, Gmail offers suggestions for email responses.
00:00:47.439 --> 00:00:54.799
So like maybe you've gone back and forth with a friend about some like going out for a dinner or something, and at some point they say, All right, we got it all booked, all checked, whatever.
00:00:54.960 --> 00:01:00.159
And then Google just writes up the Gmail just writes up the message for you saying, Yep, sounds good, see you there, right?
00:01:00.240 --> 00:01:02.000
Or something like that automatically.
00:01:02.320 --> 00:01:08.000
And personally, I sometimes select that option, but I a lot of times I don't.
00:01:08.079 --> 00:01:08.879
I just write whatever.
00:01:09.200 --> 00:01:10.400
Sometimes it actually gets it wrong.
00:01:10.560 --> 00:01:13.040
Like it's like, oh, that's not actually what I wanted to say.
00:01:13.280 --> 00:01:14.799
Yeah, right, right, right.
00:01:14.959 --> 00:01:21.439
And it's good for like, you know, short, like two-word things like perfect, thanks, or uh sounds good.
00:01:21.599 --> 00:01:22.719
You know, like stuff like that.
00:01:22.879 --> 00:01:24.000
It's fine, I think.
00:01:24.159 --> 00:01:26.400
But even then, I I still don't just take the edits.
00:01:26.480 --> 00:01:27.920
I just write just write it up myself.
00:01:28.239 --> 00:01:31.040
I particularly have trouble when it's like trying to select a date for me.
00:01:31.120 --> 00:01:35.519
I guess it looks at my calendar and it's like, oh, these are the free times that I have.
00:01:35.680 --> 00:01:39.120
And like sometimes it's just not correct.
00:01:39.519 --> 00:01:40.879
Yeah, yeah, yeah, yeah.
00:01:41.040 --> 00:01:47.840
So I think we're not quite there in the world of edits, but um, but we might be pretty pretty pretty pretty good in general.
00:01:48.000 --> 00:01:49.359
I mean, this study talks about that.
00:01:49.519 --> 00:01:51.760
So that's I think what is very interesting about this paper.
00:01:52.000 --> 00:02:02.640
Well, I think the there's this whole idea like, hey, the patient portal is there, patients are using it, they're sending lots and lots and lots of messages, and the number of messages is increasing over time.
00:02:02.799 --> 00:02:06.079
And so it's a it's a big administrative burden for providers.
00:02:06.319 --> 00:02:10.800
It's it's great to have that sort of access to your provider.
00:02:10.960 --> 00:02:22.800
But then if they're overwhelmed and they're in the middle of like 27 very quick 15-minute office visits, and then they have all these portal messages that they may or may not actually get reimbursed for responding to.
00:02:23.039 --> 00:02:30.879
Now you have this like added workload that you're trying to figure out, hey, what do I do with this or how do I how do I handle this?
00:02:31.039 --> 00:02:38.000
Um, I think it can be great for things like, hey, following up on a lab result or needing a medication refill or whatever it is.
00:02:38.159 --> 00:02:42.800
Um, but it can take a while to respond to these, especially if you have a lot of them.
00:02:43.039 --> 00:02:53.599
And so I think there's been this idea that if an AI system can draft a message back to a patient, it might decrease the provider workload.
00:02:53.919 --> 00:02:54.159
Yeah.
00:02:54.240 --> 00:02:56.560
This is like the Gmail drafting your email.
00:02:56.800 --> 00:02:56.879
Right.
00:02:57.120 --> 00:02:57.759
It's the same thing.
00:02:58.080 --> 00:03:05.840
But like in practice, the studies have shown that there's like limited, very limited time savings.
00:03:06.080 --> 00:03:11.759
And then, you know, what this paper was really focused on is like, hey, what do what do providers actually edit?
00:03:11.919 --> 00:03:15.120
What do they change in the responses?
00:03:15.280 --> 00:03:17.280
How can we learn more about that?
00:03:17.599 --> 00:03:18.560
Yeah, exactly.
00:03:18.639 --> 00:03:29.919
And it was a large-scale study in a sense that they looked at a period of about one year where they looked at over 13,000, 14,000 um messages and from over a thousand physicians.
00:03:30.319 --> 00:03:32.240
From over a thousand physicians, right, right.
00:03:32.639 --> 00:03:43.840
And they, you know, generally these um emails kind of fell into one of four or five categories, maybe not one, but maybe multiples of four and five categories.
00:03:44.000 --> 00:03:53.599
They were either general emails or messages, or they were results-related messages, or they were medication-related messages, or they were just like paperwork requests.
00:03:53.759 --> 00:04:02.319
Like these are the types of messages that were uh being sent by the patients and had to be responded to, or you know, were the content of the response itself.
00:04:02.639 --> 00:04:02.800
Right.
00:04:02.879 --> 00:04:14.400
And so the AI model like categorized into these four areas, and I'm sure there was like a different background model like responding to each of these four categories.
00:04:15.520 --> 00:04:28.240
So the pipeline was that you would get these messages, the patient would send the messages in, the AI model would categorize it, and then for each of these four uh five categories, they had a different prompt, obviously, because that makes sense, right?
00:04:28.319 --> 00:04:34.560
If it's a paperwork request versus a medication question or whatever, you want the AI to respond differently.
00:04:34.800 --> 00:04:36.639
So they had a different prompt.
00:04:36.800 --> 00:04:48.240
And what that just means is they uh pre-pended or added on to the AI model just some text about what an ideal response should be for that category of question, right?
00:04:48.560 --> 00:04:57.759
And so they were able to generate these four or five different prompts and as a result of that uh run it through the system and have it produce um uh an email.
00:04:57.920 --> 00:05:13.439
Now, the system also had access to AHR and patient health information, and so was able uh the L the AI was able to access that information and was able to use that to draft a response that was more personalized, that was more accurate, more all of that, right?
00:05:13.600 --> 00:05:20.480
So it then produced the response um using that prompt and that medical information and the EHR information.
00:05:20.720 --> 00:05:33.759
And then the idea was that you would offer it up to the doctor as either I think the doctors either got a blank uh were allowed to either choose between a blank, you know, template versus a pre-drafted version.
00:05:34.079 --> 00:05:40.079
And once they had the pre-drafted version, they recorded how long these doctors took to make edits to them.
00:05:40.399 --> 00:05:42.000
And exactly what types of edits.
00:05:42.319 --> 00:05:42.800
Oh, yeah, yeah, yeah.
00:05:43.199 --> 00:05:47.839
So they classified, there were like 15 different kinds of classifications they made.
00:05:47.920 --> 00:05:54.639
Like, hey, was it that they wanted to um change something about scheduling or rescheduling an appointment?
00:05:54.879 --> 00:05:58.480
Was there information about referring to a specialist that got altered?
00:05:58.639 --> 00:06:05.439
Was there a different interpretation of lab results or a recommendation for testing, like an x-ray or lab imaging?
00:06:05.759 --> 00:06:10.240
Um, was there recommendation for an urgent care or ED visit?
00:06:10.480 --> 00:06:14.240
Was there um tapering a prescription medication or stopping it?
00:06:14.480 --> 00:06:22.240
There were 15 of these that they that they categorized as like ways in which the providers changed things.
00:06:22.480 --> 00:06:23.360
Yeah, exactly.
00:06:23.519 --> 00:06:29.519
And and just to pause for a second, we've talked about this in our uh podcast quite a bit about this notion of human in a loop.
00:06:29.680 --> 00:06:35.360
In the past, we've talked about the idea of a human coming in and making sure that the AI systems are working correctly.
00:06:35.600 --> 00:06:39.920
And uh we've also talked about how maybe human in the loop is too broad a term.
00:06:40.000 --> 00:06:41.279
Maybe we need something more specific.
00:06:41.360 --> 00:06:42.319
What does it actually mean?
00:06:42.480 --> 00:06:46.639
How much of a burden it is for the human to go and evaluate the AI response?
00:06:46.879 --> 00:06:52.480
Is the trade-off of the human just like writing the email versus editing things worth it?
00:06:52.720 --> 00:06:56.639
Right, given the fact that they have to accuracy fact-check it and things like that.
00:06:56.800 --> 00:07:10.800
Um well, this is interesting because like there's some research that shows that when these portal messages, these drafts like come up, even if they're wrong, sometimes providers just send them through, right?
00:07:11.120 --> 00:07:16.560
It's like it's you feel like, hey, this is automatic, it's probably correct, like I don't have time for this.
00:07:16.879 --> 00:07:17.199
Go.
00:07:17.439 --> 00:07:18.079
Yeah, right?
00:07:18.319 --> 00:07:18.560
Right.
00:07:18.800 --> 00:07:24.959
I mean, you could imagine a situation like one of the categories they have is a medication related thing.
00:07:25.040 --> 00:07:30.160
So say someone, you know, types a message that says, Hey, I want my med refilled, right?
00:07:30.480 --> 00:07:30.720
Yeah.
00:07:30.959 --> 00:07:33.680
And so maybe your provider would say, Yeah, sure.
00:07:33.759 --> 00:07:38.879
Like, let me provide you with a 90-day supply, um, whatever, whatever.
00:07:39.199 --> 00:07:50.720
You could also imagine a scenario where your provider might say, Hey, I'd like to see you in the office to check on the status of this medication and see if it's continues to be a good fit for you.
00:07:50.959 --> 00:07:52.079
Let's make an appointment.
00:07:52.160 --> 00:07:54.560
I'm gonna give you a 30-day supply until then.
00:07:54.720 --> 00:07:54.879
Yeah.
00:07:55.040 --> 00:07:56.319
And so there's some nuance there.
00:07:56.399 --> 00:07:58.399
It's not like an automatic yes.
00:07:58.560 --> 00:08:02.319
There's some clinical decision making that may go into what the response would be.
00:08:02.480 --> 00:08:02.879
Yes.
00:08:03.120 --> 00:08:04.160
And so it's hard.
00:08:04.319 --> 00:08:09.279
Like it's it's hard for the AI to necessarily know that or do that.
00:08:09.680 --> 00:08:22.000
Well, there's two pieces to that, I to me, it seems like one is the AI might know that if that's standard practice, but if it's a physician-specific preference, then the AI might not know that, but in theory, could learn that over time.
00:08:22.240 --> 00:08:29.439
Uh, but yes, you make a very good point, which is there's a lot of variation, even in very simple things like refilling a prescription, right?
00:08:29.519 --> 00:08:32.080
What's seemingly simple might not be that simple.
00:08:32.399 --> 00:08:35.360
Um, and and that's that's that's interesting here as well.
00:08:35.519 --> 00:08:47.759
And so the the human in the loop it becomes a human on the loop when they stop like actually caring or not paying enough attention to the answer to the AI system and just kind of letting things go, right?
00:08:48.000 --> 00:08:50.240
And and so you know, we're gonna keep talking about this.
00:08:50.320 --> 00:08:59.360
I think this is a very interesting topic in general, but in this paper, uh, this is what they're operationalizing is trying to say, okay, the there's a human in the loop here.
00:08:59.600 --> 00:09:04.000
What amount of time are they taking to make the edits and what kinds of edits are they making?
00:09:04.080 --> 00:09:08.879
And is there a relationship between the kinds of edits and the time it takes?
00:09:09.120 --> 00:09:09.600
Yeah.
00:09:09.919 --> 00:09:17.440
And the number one um modification they made was around scheduling or rescheduling appointments.
00:09:18.159 --> 00:09:21.279
That's the number one that took the most amount of time.
00:09:21.679 --> 00:09:24.480
38 point, I'm sorry, the most amount of time overall.
00:09:24.639 --> 00:09:24.720
Yeah.
00:09:24.960 --> 00:09:39.759
When you add it up overall, because even though it may be quick to like change change that copy in the in the message, 38.5% of the encounters had a change around scheduling or rescheduling an appointment.
00:09:39.919 --> 00:09:46.879
I'm not particularly surprised by this because I have seen some of these AI drafted notes, and most of them say, like, I recommend you come into the office.
00:09:47.039 --> 00:09:48.480
Like I see that all the time.
00:09:48.639 --> 00:09:51.679
It's like, well, you can't actually have everyone come in for an office visit.
00:09:51.759 --> 00:09:52.960
That's not actually practical.
00:09:53.200 --> 00:09:53.440
Right.
00:09:53.600 --> 00:09:56.960
And so maybe that has to do with prompting strategy as well.
00:09:57.200 --> 00:10:00.320
But overarchingly, that was like the number one.
00:10:00.559 --> 00:10:01.840
Um, and then the number two.
00:10:02.240 --> 00:10:02.960
But that was bulk, though.
00:10:03.039 --> 00:10:03.519
That was overbulk.
00:10:03.759 --> 00:10:04.000
Overarching.
00:10:04.240 --> 00:10:08.080
Let's just take a moment to talk about per email what took the patients more time.
00:10:08.240 --> 00:10:11.360
And that was things like reviewing radiology charts, right?
00:10:11.519 --> 00:10:12.879
That was like very technical.
00:10:12.960 --> 00:10:23.519
It sounded to me like it was very technically specific and potentially something that the doctor had to read very carefully to make sure it was accurate to make and then made the edits accordingly.
00:10:23.600 --> 00:10:29.039
So they took more time to work on those things, but those types of messages were much less frequent.
00:10:29.279 --> 00:10:37.039
So overall, it didn't, it didn't like it didn't cause the same amount of workload, time load as scheduling and other stuff.
00:10:37.360 --> 00:10:40.720
Right, because the scheduling was like almost 40% of the messages needed that edited.
00:10:40.879 --> 00:10:41.039
Right.
00:10:41.200 --> 00:10:52.480
Whereas the things that, I mean, and this makes sense, the things that are sort of requiring clinical judgment are the things that the providers are gonna take the most time to change and adjust and edit.
00:10:52.639 --> 00:11:00.639
Things like you said, interpreting a radiology result, uh, clarifying or ruling out a diagnosis, interpreting a lab result.
00:11:00.799 --> 00:11:10.960
It's like the stuff that gets complex that the AI maybe, I don't know, it's not able to give the right information to the patient.
00:11:11.120 --> 00:11:14.799
And so these are the areas where they required the most editing.
00:11:15.039 --> 00:11:21.679
And so it's like per in the messages that had those things, they required a lot more editing time.
00:11:21.919 --> 00:11:31.440
But because those were less frequent, if you add up over time, it's the administrative stuff that took the longest because it was in the most percentage of these messages.
00:11:31.840 --> 00:11:32.320
Yeah.
00:11:32.720 --> 00:11:35.279
So, you know, I think that that was an interesting result.
00:11:35.360 --> 00:11:45.519
And, you know, the the point they were making was that this is a nice way to measure the performance of these systems, which in theory could be improved by better prompting.
00:11:45.679 --> 00:11:57.840
Maybe that would result in fewer edits needed because the messages were probably the uh the assumption there, of course, is that the messages were, you know, could be improved accuracy or could improve, like be more personalized to the physician or whatever.
00:11:58.000 --> 00:12:01.279
Yeah, or maybe they just like shouldn't recommend a doctor's visit every single time.
00:12:01.440 --> 00:12:01.919
You know what I mean?
00:12:02.000 --> 00:12:07.440
It's like little things that if you see it over and over and over again, maybe it's like an adjustment that could be made.
00:12:07.679 --> 00:12:13.600
It should remember, yeah, any time any kind of stuff, any kind of um preferences that the doctor might have as well.
00:12:13.840 --> 00:12:18.480
Um, I I you know, I think this I have I have a few questions that I feel like haven't been covered in the study.
00:12:18.559 --> 00:12:30.000
I mean, I don't know if you have more to say about the study particularly, but there were questions that I had as I was reading this that were um had to do again with the with sort of this human AI interaction piece.
00:12:30.159 --> 00:12:57.279
Now I was just imagining these doctors using this, and this was only a one-year study, and there was a thousand physicians, and I forget how how I don't know that I saw any results about um sort of longitudinally, like one taking one physician and seeing what they did over the course of multiple um responses, like did their edit times go down over time for a certain category or collectively?
00:12:57.679 --> 00:13:18.159
Because to me, I wonder sometimes with these studies is you get some result like this, and then you wonder if in fact, over time, do these physicians start to start to kind of uh assume that the AI is doing the right thing, or maybe it does it a right thing a couple of times, and then you kind of step back a little bit because you're like, okay, I think it seems to know what it's doing, I'm just gonna let it go, right?
00:13:18.399 --> 00:13:28.000
Yeah, and they don't they don't I they didn't do that in this study, but it on the flip side, like so you could say maybe the physicians will say, Oh, it's automated, it's doing fine, I'm gonna send it through.
00:13:28.159 --> 00:13:32.399
Or maybe they just get faster at editing because they know the areas where it errors, right?
00:13:32.480 --> 00:13:40.720
Like it's like, oh, it always messes up this appointment thing, or like anytime there's an x-ray result include included, I'm gonna have to spend time at editing it.
00:13:40.879 --> 00:13:45.600
They get to know like the nuances of where it does a good job versus where it doesn't.
00:13:45.679 --> 00:13:48.559
Like maybe it's pretty good at at infusing empathy, yeah.
00:13:48.720 --> 00:13:49.679
Um, etc.
00:13:50.080 --> 00:13:54.080
So I think there are components of it that yeah, the provider could get to know.
00:13:54.240 --> 00:14:01.600
But I do think it makes sense that people could kind of develop this automation bias even more over time.
00:14:01.679 --> 00:14:01.919
Yeah.
00:14:02.080 --> 00:14:04.879
Because it's like, hey, that took the administrative burden off me.
00:14:04.960 --> 00:14:06.559
I'm gonna just like move it along.
00:14:06.799 --> 00:14:07.039
Yeah.
00:14:07.200 --> 00:14:11.919
And I think the one thing that counters that a little bit is the fact that new models come out.
00:14:12.159 --> 00:14:31.600
And new models, you know, you could easily imagine your hospital or whatever changing from Claude Anthropic to OpenAI, or open AI releases a brand new model, or something else happens, and as a result, the types of emails, the types of messages that are created are totally different.
00:14:31.840 --> 00:14:32.799
So then what do you do?
00:14:32.960 --> 00:14:35.360
It's almost like you're working with a different person, right?
00:14:35.600 --> 00:14:36.000
Totally.
00:14:36.399 --> 00:14:43.679
So you've built over time as a physician, you've built an understanding of the strengths and limitations of the AI system you're working with.
00:14:43.840 --> 00:14:46.559
But if that just changes all of a sudden, now you have to start again.
00:14:46.720 --> 00:14:52.799
So that now, you know, now you have to figure out okay, this new AI system which reads differently.
00:14:52.960 --> 00:14:56.960
I have to now read this thing again very carefully to understand what the strengths and weaknesses are.
00:14:57.120 --> 00:15:02.080
Um, that could be an increased burden as you know, as we get new and new models, and new models come out every year.
00:15:02.159 --> 00:15:04.960
So it's like it's going to be interesting when that happens.
00:15:05.039 --> 00:15:06.000
And they're all different.
00:15:06.080 --> 00:15:09.919
The interaction modalities are different, their prompt expectations are different, all of that is different.
00:15:10.159 --> 00:15:13.600
Yeah, I think it's interesting because like newer doesn't always mean better, right?
00:15:14.080 --> 00:15:16.559
It could be worse for this use case.
00:15:16.799 --> 00:15:23.840
And also there has to be some component of like education of these providers to say, like, hey, we've changed this.
00:15:23.919 --> 00:15:25.919
Here are the changes you can anticipate.
00:15:26.080 --> 00:15:32.799
Um, and here's where, you know, maybe that means the underlying prompts need to get changed too when the model gets updated.
00:15:32.879 --> 00:15:35.360
These things need to get tested, these things need to get governed.
00:15:35.440 --> 00:15:47.200
You can't just snap your fingers and say, Oh, there's it's not like when your uh, you know, computer updates the operating system and you kind of expect it to just like keep working as is, maybe run a little faster.
00:15:47.360 --> 00:15:49.120
It's like it could change everything.
00:15:49.360 --> 00:15:49.600
Yeah.
00:15:49.840 --> 00:15:50.159
Yeah.
00:15:50.399 --> 00:15:55.039
And and honestly, to me, it seems like much more than response times.
00:15:55.120 --> 00:16:04.799
I mean, is response time even the right metric for measuring, in fact, if um, you know, there's an appropriate level of load, cognitive load on the on the doctor.
00:16:05.039 --> 00:16:10.799
Because ultimately what you're saying is, hey, I, doctor, don't want to respond to all these emails.
00:16:11.039 --> 00:16:20.879
Um, I want to rely on AI system to help my burden so that I can I can I can be uh responsive to my patients and have get them the answers they need, but not like overburden myself.
00:16:20.960 --> 00:16:22.480
Ultimately, that is the goal, right?
00:16:22.720 --> 00:16:26.559
So then that means at some level, the question is one of trust.
00:16:26.960 --> 00:16:32.879
Do you, as a doctor, trust the system to write the right emails or write the right responses?
00:16:33.120 --> 00:16:34.879
And I how do you measure that?
00:16:35.039 --> 00:16:43.759
Like, how do you um ensure that the pay physician is not overtrusting the AI system, um, or undertrusting for that matter?
00:16:43.840 --> 00:16:56.960
So, like, if they you know, I don't know if response time captures that accurately because you could imagine a situation where a doctor who's so busy or so tired uh responds quickly or has a low response time, but that's got nothing to do with trust.
00:16:57.120 --> 00:17:02.480
On the other hand, you could say the doctor completely trusts the system and it does great all the time, which is why the response time is low.
00:17:02.720 --> 00:17:09.039
So it feels like we're not quite capturing the core aspect of this, which is the trust aspect, which to me I think is very important.
00:17:09.440 --> 00:17:22.319
Well, it's interesting that you bring it up because I think that brings us back to this idea that, or what they found in the study, which is that the editing burden for anything requiring clinical judgment was longer.
00:17:22.720 --> 00:17:30.240
So that's the area where you're like, hey, people don't people aren't trusting it because it's not necessarily giving the correct result.
00:17:30.480 --> 00:17:30.559
Yeah.
00:17:30.720 --> 00:17:38.640
It's like when it came to interpreting a lab or a diagnosis or interpreting a radiology study, saying, here's what this means for you.
00:17:38.880 --> 00:17:41.200
That's the stuff that required clinical judgment.
00:17:41.359 --> 00:17:45.680
And that's where it didn't do a great job because you see that people are editing there.
00:17:45.839 --> 00:17:46.079
Yeah.
00:17:46.240 --> 00:17:57.519
So it's like, well, maybe some more information needs to be built into the system to make that better, perhaps with prompting or macros or I don't even know what.
00:17:57.920 --> 00:18:05.279
But I I think this study is showing that the trust isn't in those places where it's most clinical.
00:18:05.440 --> 00:18:05.839
Yeah.
00:18:06.160 --> 00:18:09.519
And and so how like how do you move forward from that?
00:18:09.839 --> 00:18:10.160
Yeah.
00:18:10.319 --> 00:18:22.720
I I also there's another piece of it which I think is more systemic, which is aren't at some level, if you have AI responding to your patient's messages, are you then losing the connection with that patient?
00:18:22.960 --> 00:18:37.519
Because that patient, I I understand that there's a burden of of just volume, but outside of but just thinking for per patient perspective, the doctor is developing this relationship with the patient to understand what their needs are, what their medical situation is, it's unique for everybody.
00:18:37.680 --> 00:18:45.119
Um, and by having the AI respond, they're they're not gonna see the signals that maybe they need to see as the human doctor.
00:18:45.279 --> 00:18:48.319
And I worry sometimes that maybe that's skipping a step.
00:18:48.480 --> 00:18:56.559
That is, they're not building that that understanding of the patient, and they're gonna over time going to get more and more separated from what the patient actually needs.
00:18:56.960 --> 00:19:10.079
Yeah, I mean, this this brings me back to some research I've read like ages ago about how when you take notes, like in class, in a in a college class, if you were to take notes, yes, you like remember the information better than if you were to type it.
00:19:10.319 --> 00:19:21.759
Here, I mean, presumably I'm extrapolating, but like if you write the message, you probably remember it better and build your relationship more than if you let AI write it and you just press send, right?
00:19:22.160 --> 00:19:40.000
It's like when you spend the cognitive load thinking about your patient and what X, Y, or Z means for them as a human being in a personalized fashion, then you are working on building that relationship and building that trust between the two of you.
00:19:40.319 --> 00:19:40.559
Yeah.
00:19:40.720 --> 00:19:48.799
And at some level, what I'm saying, so I guess that goes back to the categories that they had, which was all about the scheduling versus like the medical heavy pieces.
00:19:48.960 --> 00:19:53.039
You know, one one would one might argue that scheduling is a task that that's okay.
00:19:53.119 --> 00:19:54.160
You can have it automated.
00:19:54.319 --> 00:20:07.440
But on the other hand, a short conversation with a patient and you realize that the reason they want to reschedule is because something that was medically related is now relevant to you, could actually surface when you have that conversation, when you have that direct email.
00:20:07.599 --> 00:20:21.359
So it's uh, I mean, even that, even a task that is seemingly not like medically relevant, like rescheduling an appointment, could have information in it that might be relevant to the doctor's understanding of the patient.
00:20:21.519 --> 00:20:33.200
Um, so I I don't know what the right answer is because I do appreciate the burden and the and the you know the right answer is we need more primary care physicians, and the right answer is we need We need to give them more than 15 minutes to do an office with it.
00:20:33.839 --> 00:20:37.279
And we need to reimburse them for responding to portal messages.
00:20:37.519 --> 00:20:41.839
Like there are so many systematic issues that this is trying to get at.
00:20:42.240 --> 00:20:54.480
And I think the truth of the matter is that AI is can be used as a as a tool, but it doesn't fix the systemic underlying problems that are so difficult right now.
00:20:54.720 --> 00:20:56.400
People don't have access to care.
00:20:56.559 --> 00:20:57.920
There are not enough providers.
00:20:58.079 --> 00:21:02.960
The providers don't have enough time with patients, they don't get reimbursed for certain things.
00:21:03.119 --> 00:21:06.960
Like a lot of practices actually don't bill for these portal messages.
00:21:07.200 --> 00:21:11.599
So there are all these systemic problems that you're like really not fixing with AI.
00:21:11.839 --> 00:21:12.240
Exactly.
00:21:12.400 --> 00:21:13.920
And I think that's the point, right?
00:21:14.000 --> 00:21:15.039
I think that's exactly right.
00:21:15.119 --> 00:21:23.920
And I think the takeaway then is it might seem like AI is helping with handling workload, but the workload is a systemic problem in the first place.
00:21:24.079 --> 00:21:27.680
And it's not a solution to just say, hey, I'll fix it with technology.
00:21:27.839 --> 00:21:31.920
It's a solution to you may have to go back to the root cause and fix the root cause first, right?
00:21:32.160 --> 00:21:32.400
Right.
00:21:32.559 --> 00:21:38.000
And I think the research here is like, in terms of is this actually helping, right?
00:21:38.160 --> 00:21:41.119
There's like limited to no time savings in these studies.
00:21:41.359 --> 00:21:45.440
And then there's erroneous drafts that are just being passed through.
00:21:45.599 --> 00:21:49.279
And then there's a lot of edits that need to be made on the clinical side of things.
00:21:49.440 --> 00:21:53.519
So I think like the the concept of this is an interesting one.
00:21:53.759 --> 00:21:56.079
Some providers may feel like this is helpful.
00:21:56.240 --> 00:22:03.599
Some providers may not use this at all, but it it doesn't help with the systemic underlying issues that still do need to be addressed.
00:22:03.839 --> 00:22:11.200
Yeah, and I think we always need to keep track of that with any kind of new technology that's being integrated into our socio-cultural fabric, right?
00:22:11.279 --> 00:22:17.920
It's in it's a piece of technology that comes into our lives, and as a result, going to be in factored into the whole system.
00:22:18.079 --> 00:22:22.000
And I think we need to think more about that as a society.
00:22:22.400 --> 00:22:23.039
Absolutely.
00:22:23.200 --> 00:22:24.400
All right, thank you for joining us.
00:22:24.480 --> 00:22:26.000
We'll see you next time on Code and Cure.
00:22:26.400 --> 00:22:27.200
Bye for now.