Latest / Code & Cure / #64 - Can A Timely Push Alert Prevent Hospital Deaths?

Transcript

#64 - Can A Timely Push Alert Prevent Hospital Deaths? · Code & Cure

• 0:00
  SPEAKER_01: What if the most important thing AI does in the hospital is simply know when to call for help?
• 0:15
  SPEAKER_00: Hello and welcome back to Coding Cure, where we discuss decoding health in the age of AI.
• 0:20
  SPEAKER_00: My name is Vasant Sarati.
• 0:21
  SPEAKER_00: I'm an AI researcher and cognitive scientist, and I'm here as always with Laura Hagopian.
• 0:26
  SPEAKER_00: I'm an emergency medicine physician and I work in digital health.
• 0:29
  SPEAKER_00: This is, I thought this paper was awesome because it was a simple solution and a very effective uh approach to something very important.
• 0:39
  SPEAKER_01: Absolutely.
• 0:39
  SPEAKER_01: I mean, basically, when they did this AI triggered rapid response, they saw a decrease in mortality.
• 0:47
  SPEAKER_01: Like fewer people died in the hospital.
• 0:48
  SPEAKER_00: Wait, wait, wait, wait, wait.
• 0:49
  SPEAKER_00: Let's let's start from the beginning.
• 0:50
  SPEAKER_00: Let's start from the very top.
• 0:52
  SPEAKER_00: We're talking about um rapid response and dying in the hospital.
• 0:57
  SPEAKER_00: And in my head, I go, uh, I think of the the movies or you know where they where they have like somebody in the hospital bed and the machine starts beeping all kinds of crazy and makes all these lights and has all this beeping, and people come running in and they help the person out um and save them, so to speak.
• 1:16
  SPEAKER_00: There's early warning systems, right?
• 1:17
  SPEAKER_00: These are all jargony words that that exist, but like I I want to situate it in my Hollywood mental model.
• 1:26
  unknown: Yeah.
• 1:26
  SPEAKER_01: Well, um, your Hollywood mental model is missing something very important where they come in, they shock the person, and then they're suddenly alive again.
• 1:35
  SPEAKER_00: Yeah, it's always a shock, right?
• 1:37
  SPEAKER_00: Always a shock.
• 1:38
  SPEAKER_01: Right.
• 1:38
  SPEAKER_01: Um, so what when we're talking about a rapid response team, that's pre that's pre-shock.
• 1:42
  SPEAKER_01: It's pre like cardiac arrest.
• 1:45
  SPEAKER_01: Um, a rapid response team is a team of people that come in when a patient isn't doing well to try to intervene.
• 1:51
  SPEAKER_01: Okay.
• 1:52
  SPEAKER_01: And so it's a team of people, it's often like an I an ICU doctor, a critical care nurse, um, someone who does respiratory therapy.
• 2:00
  SPEAKER_01: Someone who's at the hospital that that you know, don't you?
• 2:02
  SPEAKER_01: They're physically there in the hospital.
• 2:04
  SPEAKER_01: And it's not just one person, it's like a team of people.
• 2:06
  SPEAKER_01: And what what they do is they go to the bedside of a patient that's not doing well and they try to figure out hey, what's going on and how can we stop this person from getting worse?
• 2:18
  SPEAKER_00: Yeah.
• 2:18
  SPEAKER_01: What do we need to do for them?
• 2:19
  SPEAKER_01: And so the whole idea is like, hey, we want to get in there and stop them from dying, right?
• 2:24
  SPEAKER_01: Like we want to we want to stop them from maybe needing the ICU if possible.
• 2:30
  SPEAKER_01: So, so a lot of times these things are triggered by someone's vital signs or by someone like noticing that they're not doing well.
• 2:40
  SPEAKER_01: So maybe, maybe if someone's admitted to the hospital, they start having a bunch of seizures, you know, someone would press a button and call a rapid response.
• 2:48
  SPEAKER_00: Or so it's a person who presses the button to call the rapid.
• 2:52
  SPEAKER_01: Yeah, that is interesting to me.
• 2:53
  SPEAKER_01: How do they get called?
• 2:54
  SPEAKER_01: Or it could be triggered by their vital signs.
• 2:57
  SPEAKER_01: So that's a common way people do it where it's like, hey, um, their heart rate went really low.
• 3:02
  SPEAKER_01: Got it.
• 3:03
  SPEAKER_01: Or their blood pressure is really low, or their oxygen level is really low, or they're suddenly like super confused.
• 3:09
  SPEAKER_01: We need help.
• 3:10
  SPEAKER_01: Something's going on with this person.
• 3:12
  SPEAKER_01: We need to figure out what to do with them.
• 3:14
  SPEAKER_01: And so if that happens, usually it's humans who will, you know, call, call a number or you know, page overhead and say, Hey, I need the rapid response team to this location, or they'll send it out on pagers.
• 3:30
  SPEAKER_01: There are still hospitals that have pager systems.
• 3:33
  SPEAKER_01: Um, and you'll get sort of beeped, or you'll get a push notification, or you'll get a little text message that says, Hey, like, come and see this patient.
• 3:40
  SPEAKER_01: Let's figure out what's going on with them.
• 3:43
  SPEAKER_01: But it's usually like a human that's noticed and does it.
• 3:47
  SPEAKER_00: Right, right, right, right.
• 3:48
  SPEAKER_00: Okay, so that's a rapid response team.
• 3:50
  SPEAKER_00: And in some sense, you've also talked about an early warning system, because that is what that is, right?
• 3:55
  SPEAKER_00: That that the the the vital signs going, you know, suggesting.
• 4:02
  SPEAKER_01: Something's not going well.
• 4:03
  SPEAKER_00: Right, right.
• 4:04
  SPEAKER_00: Something's off and being and discovering that um is the early warning system part.
• 4:08
  SPEAKER_00: And now we now there are systems out there that have that that do that do this, right?
• 4:15
  SPEAKER_01: Yeah, I mean, if you look at um if you look at electronic health records, a lot of them have tried to integrate and they even have proprietary systems that try to figure out like, hey, who who is gonna deteriorate without without necessarily having um vital signs that are falling or whatever.
• 4:36
  SPEAKER_00: Yeah.
• 4:36
  SPEAKER_01: Exactly.
• 4:37
  SPEAKER_01: So it's like maybe the vital signs, maybe the heart rate of you know, 38, that's too late.
• 4:43
  SPEAKER_01: Like, can we figure out earlier who's gonna deteriorate?
• 4:46
  SPEAKER_01: So Epic is like one of the most common medical record systems, and they've created this proprietary model where they like bucket people into like, you know, stop-like categories, green, yellow, red, in terms of how risky they are and how likely they think it is that the patient will deteriorate.
• 5:05
  SPEAKER_01: And so they take all sorts of inputs, not just they they do take things like your blood pressure and your pulse and your breathing rate and your oxygen level, but they also take into consideration a bunch of nursing assessments, uh, like how their cardiac rhythm is or their neurologic assessment.
• 5:21
  SPEAKER_01: And they take into account things like labs, like what's their white blood cell count, what's their sodium level, what's their pH, whatever.
• 5:28
  SPEAKER_01: And so they take all of that information together.
• 5:30
  SPEAKER_01: And what they've trained these models on is not just the the you know, those inputs, but also the the output, the outcome of like, hey, what happened to so now I have all this information.
• 5:41
  SPEAKER_01: What happened to this patient?
• 5:43
  SPEAKER_01: Were they transferred to the ICU?
• 5:45
  SPEAKER_01: Was a rapid response team called on them?
• 5:47
  SPEAKER_01: Was a code called on them?
• 5:48
  SPEAKER_01: That's how we started off, where you have someone shocking or doing CPR or whatever uh because their heart stopped, or did they die?
• 5:55
  SPEAKER_01: And so you have all of these inputs that you can get from the electronic health record, and then the outputs that you're looking for is like, did the patient do really poorly?
• 6:04
  SPEAKER_00: Right, right.
• 6:04
  SPEAKER_00: And then how many hours since the, you know, certain, you know, uh levels were dropped or something, right?
• 6:11
  SPEAKER_00: Yeah, exactly.
• 6:12
  SPEAKER_00: Yeah, yeah, yeah.
• 6:13
  SPEAKER_01: And so the idea here is like we we know more information about a patient.
• 6:20
  SPEAKER_01: Can we create this predictive model of who's gonna do poorly with the concept that we can intervene on them?
• 6:28
  SPEAKER_01: You can imagine if someone has like a really, really high score that maybe you can't intervene, they're just gonna do poorly no matter what.
• 6:36
  SPEAKER_01: But is there a segment of patients that, hey, we could intervene and actually change the course on by getting an earlier warning sign than a heart rate of 38 or an oxygen level of 82%, you know?
• 6:48
  SPEAKER_00: Right, right, right.
• 6:49
  SPEAKER_00: And that's what this paper's all about.
• 6:51
  SPEAKER_00: I to me at least.
• 6:52
  SPEAKER_00: It seems like it's all about the intervention because these models exist.
• 6:55
  SPEAKER_00: The models do flag the red, green, and yellow uh levels.
• 6:59
  SPEAKER_00: Um, but how they go about then alerting the human or the humans, um, the human teams, is how it's different.
• 7:07
  SPEAKER_00: Like in the past, it it was it was not done with calling out on a rapid response team.
• 7:11
  SPEAKER_01: It was done with maybe a a flag in the chart or something like that, where it's less sort of in your face.
• 7:17
  SPEAKER_01: But what they did in this paper was they said, Hey, anyone who's in the red, anyone who has a high risk of deterioration, um within like 20 hours or 20, 20, 30 hours or something.
• 7:28
  SPEAKER_01: Yeah, exactly.
• 7:30
  SPEAKER_01: We're going to send a push notification to this rapid response team so that they show up at the bedside and figure out what to do.
• 7:39
  SPEAKER_01: Right.
• 7:40
  SPEAKER_00: Right.
• 7:41
  SPEAKER_00: And you know, this seems so obvious that this is what should be done, but it hasn't been done.
• 7:46
  SPEAKER_00: And that's why I love this paper because in hindsight it's obvious, but it wasn't, right?
• 7:50
  SPEAKER_00: And it's such an important thing to do is yes, you have a system that seems to predict correctly.
• 7:55
  SPEAKER_00: Why not then call the people who yes, yeah.
• 7:58
  SPEAKER_01: And so you need the clinicians there too, right?
• 8:00
  SPEAKER_01: By calling a rapid response, like what they did find was that when they did this, it led to um a lower mortality, right?
• 8:10
  SPEAKER_01: So the the risk-adjusted odds of in-hospital mortality were lower in this group, like um 0.82.
• 8:20
  SPEAKER_01: So odds ratio less than one.
• 8:22
  SPEAKER_01: And that um the in-hospital mortality went down by 4.5 percentage points.
• 8:28
  SPEAKER_01: So the whole idea here was like, hey, let's do this rapid response thing, let's send a push notification, and then let's see what happens, and people were less likely to die.
• 8:39
  SPEAKER_01: That's awesome.
• 8:40
  SPEAKER_01: I mean, that's like a great outcome.
• 8:42
  SPEAKER_01: But I I want people to understand that like it's not the, it's not just the alert.
• 8:47
  SPEAKER_01: The alert is important, like, let's go check on this patient.
• 8:50
  SPEAKER_01: But then you have all these clinicians convening on the patient, figuring out, hey, what's going on?
• 8:56
  SPEAKER_01: What can we do to prevent this trajectory?
• 8:59
  SPEAKER_01: And uh, and how can we treat this patient better, right?
• 9:02
  SPEAKER_01: Like what how can we help them improve?
• 9:05
  SPEAKER_01: Yeah.
• 9:05
  SPEAKER_01: And so there's a huge clinical aspect that goes along with it.
• 9:08
  SPEAKER_01: So there's the alert part, and then there's the the clinicians coming in and deciding what to do with the patient.
• 9:14
  SPEAKER_00: Yeah, yeah, that's right.
• 9:15
  SPEAKER_00: And and I think this is again going back to our terminology prior, or the notion of humans in the loop.
• 9:20
  SPEAKER_00: This is not an example of humans in the loop.
• 9:22
  SPEAKER_00: This is humans, you know, doing the thing, right?
• 9:25
  SPEAKER_00: But just being alerted to doing it when they needed to do it.
• 9:28
  SPEAKER_01: And it's probably earlier, right?
• 9:30
  SPEAKER_01: And one of the things that I first thought about when I read this article was like, oh gosh, you're gonna have all of these rapid response team activations.
• 9:38
  SPEAKER_01: Like, do we even have the capacity for that?
• 9:40
  SPEAKER_01: Do you have the people to do that?
• 9:41
  SPEAKER_00: Do you have false positives and you have to deal with all of that?
• 9:44
  SPEAKER_01: Yeah, exactly.
• 9:44
  SPEAKER_01: And they did find that um the activations went up a little bit, that the rapid response team was called a little bit more by 12.2% percentage points more.
• 9:55
  SPEAKER_01: Um, however, that's not like a huge amount.
• 9:58
  SPEAKER_01: Right.
• 9:58
  SPEAKER_01: It's not like they were called like double or triple what they had been called before.
• 10:04
  SPEAKER_01: And I think they put some um, they put some parameters in place to help prevent that because alert fatigue is one of the things that you just start ignoring stuff because you're overwhelmed, right?
• 10:15
  SPEAKER_01: If you get too many alerts, then none of them mean anything anymore.
• 10:19
  SPEAKER_01: And so what they did was they said, okay, we're gonna send the push notification if if this deterioration index score is in the red, but we're not going to send it if they're already in the ICU, right?
• 10:34
  SPEAKER_01: Because they're already getting that level of care.
• 10:36
  SPEAKER_01: We're not gonna send it if they're on comfort care, meaning, you know, this person is probably going to die and we're just making them comfortable.
• 10:46
  SPEAKER_01: And um, we're not gonna send it if they had uh a rapid response team event very recently, um and and an alert was already triggered, or and we're not gonna send it if they're in the middle of a procedure, because sometimes that can throw things off too.
• 11:00
  SPEAKER_01: Yeah, yeah.
• 11:01
  SPEAKER_01: And so uh, you know, when they did the the trial arm of this study, it's not like it it triggered on everyone, it triggered on about half half the patients because they had said, okay, we're not gonna exclude exactly.
• 11:16
  SPEAKER_01: And so, and they did this across different types of hospitals, academic, community, etc.
• 11:21
  SPEAKER_01: And they were able to say, like, here's how often we think this is going to trigger.
• 11:25
  SPEAKER_01: And it wasn't a crazy amount.
• 11:27
  SPEAKER_00: Yeah, yeah.
• 11:28
  SPEAKER_00: Which again, you know, really pushes for a tool like this to work effectively in these settings.
• 11:33
  SPEAKER_00: Uh I know that they had um other uh um uh other types of uh early warning system uh uh I think they had reviewed, they had surveyed some other ones too.
• 11:45
  SPEAKER_00: And I I just found it intriguing that a lot of the other uh approaches were not calling the rapid response team, but they were doing things like putting a red flag or calling or calling a central hub that then had to call somebody else, right?
• 12:01
  SPEAKER_00: So then you had you you placed all these like steps in between and slowed things down.
• 12:06
  SPEAKER_00: I mean, those were not anywhere close to as effective as this.
• 12:09
  SPEAKER_01: Right.
• 12:09
  SPEAKER_01: I think this this is like, hey, we found something that works, let's operationalize it in a way where we can intervene, because this is the situation where you wanted to want to intervene earlier before the person deteriorates, right?
• 12:23
  SPEAKER_01: And I think that was key is they were like, hey, in this situation only, we're gonna send it as a push notification.
• 12:29
  SPEAKER_01: So if the if the you know, going back to the traffic light model, if it was green or yellow, there wasn't going to be a push notification.
• 12:35
  SPEAKER_01: There might be a flag in the chart, but if it's red, it gets the flag and it gets the push notification so that people show up at the bedside and figure out what to do for the patient.
• 12:46
  SPEAKER_00: Yeah.
• 12:46
  SPEAKER_00: And and also just thinking about from the human perspective, this is also where this helps with the cognitive burden, right?
• 12:53
  SPEAKER_00: Because you don't have to worry about getting to these things um or thinking about these things too much.
• 12:59
  SPEAKER_00: Like you just it tells you when, you know, um, when you need to show up in these cases.
• 13:04
  SPEAKER_00: And it's it's not something that you are going to be burdened with additionally.
• 13:09
  SPEAKER_00: I mean, yes, you have probably more 12.5% more visits, but for the most part, that's all real, right?
• 13:15
  SPEAKER_00: To some degree.
• 13:16
  SPEAKER_01: Yeah, and the median number of daily alerts per hospital was between one and ten.
• 13:21
  SPEAKER_01: It's not like it was like, oh, there are 50 rapid responses all of a sudden.
• 13:24
  SPEAKER_01: Yeah, sure, there were 12% more, but it wasn't like a ridiculous number overarchingly.
• 13:31
  SPEAKER_01: Um yeah, yes, there were more, but maybe that was maybe that's a good thing.
• 13:36
  SPEAKER_01: Is that is that necessarily a bad thing?
• 13:38
  SPEAKER_01: It's like, well, we maybe helped more people too.
• 13:41
  SPEAKER_01: And that's when you go to the mortality and you say, hey, what happened?
• 13:45
  SPEAKER_01: And basically they found that the in-hospital deaths before this push notification intervention were at the uh in this population of 23%.
• 13:56
  SPEAKER_01: And then post-intervention, 18.63%.
• 13:59
  SPEAKER_01: So negative 4.5%.
• 14:01
  SPEAKER_01: They were less likely to die in hospital with this triggering system in place.
• 14:06
  SPEAKER_00: Yeah, yeah, yeah, yeah.
• 14:07
  SPEAKER_00: That should make people feel pretty good about uh this is a great example of AI use that actually works because it's so limited, right?
• 14:14
  SPEAKER_00: It's not doing everything, it's doing one thing really well.
• 14:17
  SPEAKER_00: Um, and so that's always always going to be a good um good result.
• 14:21
  SPEAKER_01: But it's not just the algorithm, right?
• 14:23
  SPEAKER_01: You can put the algorithm anywhere you want.
• 14:25
  SPEAKER_01: It's like the fact that it's linked to the push notification, which is then linked to humans showing up at the bedside to provide their judgment and their action before the patient physiologically deteriorates.
• 14:38
  SPEAKER_00: Yeah, yeah, yeah.
• 14:40
  SPEAKER_00: I can almost imagine sort of follow-on studies in which it's not just alerting the human response team to come in, but maybe even preparing ideas for what they can do.
• 14:50
  SPEAKER_00: Because honestly, the AI system probably has access, because it has access to everything, the health records, the patient, and everything else, might even know what's going on with the patient more than just alerting the the them to come and show up, which might be all they need, but you know, when they show up, they're gonna still need to know who this patient is, right?
• 15:07
  SPEAKER_00: So, what actually happens when the response team shows up?
• 15:10
  SPEAKER_00: Like they're gonna they're it's gonna be to me at least, it seems like they're gonna be out of context.
• 15:15
  SPEAKER_01: Well, they may not know the patient, so having like a a nice summary in the electronic health record would be useful.
• 15:22
  SPEAKER_01: Trends that have been um you know picked out, maybe um, you know, uh an explainable reason that a rapid response was triggered.
• 15:33
  SPEAKER_01: Every patient's different, right?
• 15:35
  SPEAKER_01: If you have a patient with pneumonia that's deteriorating because their oxygen level is low versus a patient with dizziness whose now heart rate is 34, those are gonna be different interventions that you do for them.
• 15:47
  SPEAKER_01: Maybe the patient with a heart rate of 34 is in a third-degree AV block and they need to go to the A ICU and get a temporary pacemaker put in.
• 15:55
  SPEAKER_01: That is very different than, you know, putting oxygen on someone with pneumonia and starting uh, you know, some breathing treatments.
• 16:04
  SPEAKER_01: So this is where, sure, there could be suggestions, there could be good summaries, because you're right, a lot of times the rapid response team is coming in without context.
• 16:14
  SPEAKER_01: They haven't treated this patient, they don't know this patient.
• 16:16
  SPEAKER_01: They're coming in to try to figure out, hey, what's going on and what can we do about it?
• 16:20
  SPEAKER_01: And it's and it may be, it may be different based on, you know, the reason they were called and why that patient is in the hospital.
• 16:28
  SPEAKER_00: Yeah, and that's why I think it would be also interesting to note like of how many of all the people, all the rapid response teams that were called and showed up, what percent of them could have saved the patient had they had information faster, right?
• 16:40
  SPEAKER_00: Or had they been up to speed, like how much time was wasted in in getting up to speed, um, and how much like thought was made, you know, there's obviously pros and cons to this.
• 16:50
  SPEAKER_00: The AI recommendations could be wrong, and you know, some of these could be wrong, and human presence there is going to be able to tell a lot more about what's happening with the patient than any AI system potentially.
• 17:00
  SPEAKER_01: But the humans may need to intervene in a way that AI can't.
• 17:03
  SPEAKER_01: Now you have you have someone in respiratory distress, maybe they need to have a breathing tube put in, right?
• 17:07
  SPEAKER_01: The AI is not doing that.
• 17:09
  SPEAKER_00: No, no, and that might be the the first problem to solve before any other problem, because that's the most important one.
• 17:15
  SPEAKER_01: Exactly.
• 17:15
  SPEAKER_01: Right.
• 17:16
  SPEAKER_01: Exactly.
• 17:16
  SPEAKER_00: Right.
• 17:17
  SPEAKER_01: Um I I do think this is a cool study though, because it's like, hey, there's a tool.
• 17:22
  SPEAKER_01: We're gonna add one thing that's seemingly small, which is just like a push notification when there's a trick when the trigger is in place.
• 17:31
  SPEAKER_01: And by doing that, by creating that very timely trigger, they are able to allow for a more timely clinical intervention that reduced the chance of death.
• 17:43
  SPEAKER_00: Yeah.
• 17:43
  SPEAKER_01: That's like huge.
• 17:45
  SPEAKER_01: Wow, that's huge.
• 17:46
  SPEAKER_01: That's a really great use of AI.
• 17:49
  SPEAKER_01: Yeah.
• 17:49
  SPEAKER_01: It's a really great use of operationalizing it in a way that humans can intervene.
• 17:56
  SPEAKER_01: It's a great example of how um AI and humans can work together where clinical judgment and action really is necessary to create this reduction in mortality.
• 18:09
  SPEAKER_00: Yeah.
• 18:09
  SPEAKER_01: Yeah.
• 18:10
  SPEAKER_00: I agree.
• 18:11
  SPEAKER_01: Awesome.
• 18:12
  SPEAKER_01: Well, thank you for joining us.
• 18:13
  SPEAKER_01: We'll see you next time on Code and Cure.
• 18:15
  SPEAKER_01: Bye for now.