Hi, everyone. My name is Ayala. I'm one of the fellows over at Maryland. It's my honor to introduce Dr. Levin. He is a practicing board-certified cardiac anesthesiologist. He is the associate chair of research and director of research informatics for the Department of Anesthesiology, Perioperative, and Pain Medicine, and the medical director of the clinical data science team over at Mount Sinai Health System. He has appointments in the Department of Genetics and Genomics and the Department of AI and Human Health. He graduated MIT and began his career working in the software industry in the dot-com boom. Throughout his career in medicine, he's continued to maintain an interest and expertise in technology.
board certification in clinical informatics and was served as a board member of the Society of Technology and Anesthesia. As the medical director of the clinical data science team, he's successfully overseen the implementation of real-time multimodal streaming data science pipeline that's used to deploy operational machine learning models at the point of care at the bedside. The clinical data science team recently won the prestigious Hearst Health Prize for its machine learning application called the NutriScan AI that facilitates faster identification and treatment of malnutrition in hospitalized patients. And I think a lot of us saw the editorial and trials called the Rescue Me trial, the results of which he's going to talk about today, demonstrating that machine learning alerts for clinical deterioration.
are associated with faster escalation of care and reduced mortality, which is really exciting. The accompanying editorial in critical care medicine emphasized that this is one of the few studies to truly examine what happens when prediction model is moved from bite to bedside. So without further ado, Dr. Matt Levin. Thank you very much for that very kind introduction. And I apologize that I'm not in person. It's a lot to go through, but I think you'll find it interesting. So the title of my presentation is taken unapologetically from the editorial because I thought it was so good, bite to bedside. Saving lives with Rescue Me. I do have one disclosure. I'm a technical advisor for a small company trying to build anesthesia databases.
It's still small, so I'm still talking. All right, so we're here today to talk about this article, which the title that we published under was Real-Time Machine Learning Alerts for the Escalation of Care. I had called it Rescue Me, which was a real-time streaming clinical use engine for medical escalation. So that was a acronym to try and make something that sounded good when I posted it. And this was a trial that we just got published in July, but really ran much earlier, as you'll see.
So before I go into the trial, I just want to briefly introduce or show a slide of the clinical data science team. So everything I'm presenting to you today, I could not have done without the support of this team. And it started out in about 2017, and it was just the four of us at the top, Robbie Freeman, who came from a nursing background myself as a physician, Dr. Kia, who was a physician who then became a data scientist, and Dr. Tim Sina, who was a computer scientist. So the four of us started this team in 2017, and it's now grown to be about 18 people over.
now seven years, and they are really the heart of the engine that enabled us to do what we did and do this trial. So this is a huge shout out to them with gratitude and, you know, the whole notion of having a team. Briefly, you know, CDS, which is clinical data science, this is our platform. So I don't want to get too technical just, but to show you that, you know, behind this study, there was a lot of fairly decent scale software engineering and data science just to bring all this disparate sources of data together, run predictions and push the predictions back into Epic or into dashboards or what we were doing for this project to a text messaging app. So it's not simple. And it took a lot of effort just to get it built before we could even move towards doing models. But now in 2024, we have 17 products at six hospitals, about 15 million predictions a year, and it spans a wide range of...
inpatient uses. So malnutrition was mentioned, falls, delirium, discharge planning, respiratory insights, which is more of a dashboard, some stuff with patient experience and surveys, stigmatizing language and notes, a pilot, a project that we're just about to launch a pilot of for whether you can predict someone's success with extubation in an ICU, avoiding admissions, pressure injuries, which is another great one to try and identify patients at risk for pressure injuries, and so on and so forth. So it's a big portfolio. that we've been very slowly growing, and it's very exciting work. So Rescue Me, let's get into this. So the study design, we'll go through the study design, inclusion criteria, statistical analysis, and then we'll talk about some of the results, limitations, and where we are now. So it was a single center study. It was what we call a pragmatic, non-randomized cluster clinical trial. And I found, you know, as the years go by, the journals want you to be more and more specific in your title of what exactly the study was.
You know, to break it down, it was clinical. It was in the clinic. It was clustered because, as you'll see, we did it by patient location. So it wasn't randomized and it was pragmatic because it was just there was no active recruiting. It was just if you got admitted, you got you got assigned to a study arm. So our hypothesis was that real time alerts generated.
by some machine learning algorithm and sent directly to the frontline providers would reduce escalation of care. And again, as I started to say, so it was any adult patient admitted to one of four medical surgical units. So we did this because we had identified four units that were really kind of step-down units that were having issues and had already a good presence of our RRT team and intensivists. So it made sense as a place to try something as opposed to rolling it out to the whole hospital, which Mount Sinai is about a 900-bed hospital, a little bit more. We're trying to roll it out to...
um, really, uh, standard patient floors where not only the rate of events might be low, but, uh, the wrath of the response team doesn't often make appearances. Um, and so your assignment to your intervention or control arms was really solely determined by the initial unit admission. So if you roll into the hospital, um, or even as we'll talk about further later, if you got, um, surgery and you left the, uh, recovery room, or if you were in an ICU and got stepped down, um, if you landed in one of these four units, that was your assignment. And so there was two control units and, uh, two, um, uh, intervention units, and it was medical, surgical in each one. The primary outcome we looked at was the rate of escalation per 1,000 patient bed days. And so this is a fairly standard way to look at events per patient bed days. It's an aggregate type of measure. And it's a rate. It's not an incident. It's not really a percentage, although I'll present some percentages. And then a secondary outcome, we looked at sort of how often orders were placed, what kind of orders. And then the one that was the real reason that we do all this, which is that can you reduce in-hospital mortality and 30-day mortality?
So that was the study design. The sample size we had calculated, you know, we looked before we started at a historical cohort of patients going back five or 10 years, it looked like about 10% of patients in these units were escalated. And again, so there's this little mixing and matching of rate per patient days versus an overall prevalence or incidence. But we assumed a 20% reduction in escalations, so that 8%, and this gave us 6,426 subjects, or 3,200 each arm. So we'll come back to this number later.
The score, the driver of this was what we call the Muse++ score. So prior to even beginning the study, we spent about two years. And this was one of the first projects that the data science team did was to try and develop a better early warning score. And so we are hardly the first to do this. There are many, many, many early warning scores published, both sort of classical early warning scores, which are just purely.
you know, additive or rules-based, up to and including, you know, machine learning-based scores. So we were hardly the first machine learning score. But we did what everybody else does, everybody else does, which is we decided we had to make our own score. And this, you know, just briefly shows you that, you know, we sampled data in a 24-hour window prior to the prediction. And then we hypothesized this sort of golden hour of six hours that would be intervention. And then there was the time of escalation. So we built this model. We trained it. These were the features in the model. And it's everything you really expect. And I think the big difference between...
when we built this model in 2018 versus the models we're building now is that there would be a fifth column, which is the LLM, natural language processing, chat GPT, pulling in notes, free text data, which seems to, it does add something. It's not it's not game changing, but it does add something. But most of what you see here in the features are kind of what you expect. And as I said, I think lots of people have built models and they all sort of converge on the same set of features. And the demographics and vitals are pretty straightforward. The nursing assessments, you know, I'll just tell you. So we use Epic. And so the nurses have all these structured flow sheets where, you know, they say the patient is awake, sleepy, arousable.
or they say that their breathing is labored or comfortable. So we pulled all those in and then you have to normalize everything and turn it into categorical variables. And it all got thrown into the model. So the model performance is really good. And we chose our thresholds to get a balanced sensitivity and specificity and a good AUC. And on the right of my screen, you just see that sort of compared to the classical Muse, which is sort of these four measures that this Muse++ outperformed at all time points up to 24 hours before. So this was the basis. So now the question is, can we take this and can we deploy this into real life and actually use this to drive a workflow?
And again, this is one of the first projects we did. We got a little complicated, as you'll see, which turned out to make things complicated. But we thought it was good at the time. And so here you have what was the study protocol. And this is actually a simplified version of what was generated in our working groups. But to walk through it from left to right, patient gets admitted. There was sort of a two-hour, what we call refractory period, because we found early on in the pilot testing that if you send an alert to the RIT team, the minute that a patient gets stepped down from an ICU to...
A step-down unit, the RT team tends to get kind of annoyed because they just saw this patient. They just said that they were good to go to the floor. And that Muse++ score is based, it's kind of a lagging trend. So it's based on what was in the last 24 hours of Labs of Vital. So of course, when they first get admitted, they're going to have still a high and maybe downtrending score.
And then what's left out of this is we ended up implementing all these heuristics where this score, as we rolled it out during this trial, was generated continuously. So every time you got a new lab, every time the nurse evaluated a patient, it generated a new score. So you can imagine this generated a lot of scores. So early on, we said, OK, if the score doesn't change more than 10%, let's suppress it. Then again, the nurses complained about too often alerting. So then we said, OK, well, if the score, we'll limit it to sort of like one alert per nursing shift. And then we even further limited it to only three alerts per patient stay in the unit. So all of these things.
as we'll talk to you later in the limitations. you know, probably affected the way the alert worked, but they were also practical in that our initial vision, which is, you know, just to walk through this, the rest of it, you know, the score was high. So the patient is really definitely sick. We're going to send an alert straight to the RT team and I'll show you. It went to an app called Curator, which was one of these HIPAA compliant instant messaging apps. I think that company now pivoted and doesn't even do this anymore. And it was supposed to send them running to the bedside. And that was supposed to be, you know, fairly specific so that there weren't a lot of false positives wasting the time of the team. But if the score was above a lower threshold,
he would just send an alert to the nurse by their vocera. So those are those little voice badges that the nurses wear. And so it would say, you know, use plus plus alert, you know, patient bed 12. And then the nurse would go and the nurse would take vital signs. And then the nurse would decide if they wanted to call the primary team or ROT. And so those were meant to have, you know, fewer false negatives because the cost of activating the nurse to go help take a look at the patient was relatively low.
And then, as you can see, then there was another refractory period, then you could send a second alert period, the same set of response, and there were a third alert. And then we ended up capping it after three alerts because most patients didn't stay in these respective units that long because they were step-down patients. And the ones who did stay a long time, sort of these, you know, regulars, there was alarm fatigue because you would just keep on getting alert on a patient that everybody knew was sick and maybe there wasn't anything to do about it. So I know I presented here a study protocol and I kind of jumped ahead a little bit to sort of some limitations, but I wanted to explain a little bit more in depth. you know, what's behind this low diagram and sort of the choices and compromises to make in doing an alert in this way, in a true real-time fashion. And so we'll get, we'll get back to that a little further. So then, as I said, they were real-time alerts. So they sent these nice little messages to the patient, the RT team's iPhone with the location and then this probability and then the top 10 factors. And you can see they were everything, you know, you might expect. False is 95. The patient is 95. You know, the INR is 1.5. The hemoglobin, they're breathing a little fast.
And the interesting thing about this trial, which I'll show you some word clouds later, but this was really an interactive system. So the RIT providers were always texting back to the computer, even though nobody was reading this. So it was like a weird, I mean, again, now this would be some sort of LLM or chat GPT and the chat GPT bot would probably respond to the physician. But here's a classic one. They had RVR, AFib, no need for upgrade. So that was the trial.
And then when we finish the trial, we have to analyze it. And this slide has a lot of text on it, and I apologize, and I'll try not to just read off of it. But, you know, very, very complicated because as I sort of started to describe in the study protocol, okay, you have many, many alerts. You've capped the alerts. The patients can be in a unit, leave a unit, come back to a unit. Patients can switch between units, although in the end, we didn't have much crossover. They can have multiple hospital missions. So what's sort of like a fair way to analyze this? It's not a randomized trial. So you're trying to eliminate bias.
You're trying to, and if you have multiple alerts and chains of alerts, you know, do you choose the first alert to analyze the last alert to analyze, you know, the first score on admission. So it was a very, very complicated analysis. I'll call out a few things, which is, you know, we use something called inverse probability of treatment weight, which is a form of propensity score matching because the units were not the same. And then even, and that's, even though we did that, then, you know, one of the reviewers came back and said, well, I still want you to include some sort of indicator variable for the unit that they were admitted to because as a fixed effect. because I think the unit is biasing your results. So there was a lot going on in the analysis. It took a really long time. The Poisson regression is what you do for rates. The log binomial is what you do for binary outcomes of mortality or even things like length of stay. And then we did look at the performance of the alert itself.
All right. So now we have enrollment. So you'll see some little viruses floating around. And why? Because we had started this in July 2019. And obviously, as everyone probably lived through, in March 2020, COVID slammed into New York City and the East Coast. And so we just had to shut the study down because suddenly every unit in the hospital was in ICU. And we just had no time to run a trial. So what happened was, you remember I said that we thought we needed about 6,400 patients. So we ended up stopping at 2,700.
with only about 1,400 in the intervention arm. And I just also highlighted that not a lot of alerts were sent. And that's just because we throttled the alerts so much to try and avoid sort of alert fatigue. But as we'll see, even with this termination for COVID, we were able to find some results. And like I said, as you'll see, we felt pretty confident that the results would have stood had we completed. But we got shut down, like many others. And so here's table one and table two.
You're kind of classic, you know, this is basically just demographics. And I'll point out a few things. So, you know, column on the left intervention was you got the alert. The control group, no alert sent. So for every patient in the control group, we...
We were generating the Muse++ score. We were generating the alert. We just never sent it. So we knew what was going on in the background. And there's some things that stand out. And this is a function of the fact that different services admitted to different units. And you see that over here as well, where the liver service admitted more to the control unit. So you could argue that the control unit was more sick. The sort of bariatric service admitted more to the intervention unit. And so these are the differences that we tried to account for.
both with the propensity, the propensity weighting, and then using the ward or the unit as a fixed effect. But no doubt, these were different populations. And that's one of the limitations of doing these types of studies sort of in the hospital in the real world is you get what you get. Patients are not admitted to units at random. Units are designated for certain services. So that's something to account for.
The workflow was pretty good. So because we tracked all this sort of granular data about when people responded to alerts, because we had even with the vocera, when they clicked the vocera, it actually returned an acknowledgement that we stored in our database. So, you know, the nurses loved this. You can see 8.7 minutes to get to the bedside. So that's not bad considering a nurse, you know, likely is with another patient or at the nursing station. And even for the RRT providers, you know, because we, I think probably because we were sort of limiting to units that they kind of hang out on anyway, it was pretty quick. They got there pretty quickly.
Kind of a bigger spread than the nurses, but still there. And the positive predictive value of the RIT alert, meaning how often when the RIT physician showed up, did they actually think, yeah, this patient should be escalated? It was about one in six. So not amazing, to be honest with you, but not terrible. So what happened here with alerts and orders? So on the left side, you can see sort of the breakdown.
Most of the alerts happened right away when a patient got admitted to the units in both groups. And then it really dropped off. So that's, I don't know if you can see my pointer here, but that's sort of the table up here. So by the third day, you know, not very many people were getting alerts. And a lot of alerts were at the lower level, what we call the primary level. So these were mostly the alerts to the nurses. And about 50% of people got sort of both alerts. And then as mentioned, we capped the alerts and the majority of patients in the study actually didn't get alerts, which is, you know, potentially an issue of power. And most patients only got one primary alert.
On the right side, we sort of see orders placed. And we did find a difference in orders. And it was mostly driven by cardiovascular drugs. So whether that's a push of labetalol or some antiarrhythmic, everything else, you know, again, a lot of stuff in this table. But if you look at the standardized mean differences, most of them cross zero. So there wasn't too much difference. And this was just simply sort of a univariate analysis.
So how did the alerts do? Well, the alerts worked quite well. So again, how did we define alert performance? That was, well, when an alert went off, did the patient end up having an escalation? And so it seemed like they did. You can see that the primary alert's a little bit more sensitive than the RIT alerts. The RIT alerts are a little bit more specific than the primary alerts. So this was all kind of by design. The best was when you looked at patients who got both alerts. Then the performance got pretty good. So, you know, the positive predictive value for both alerts was...
a little bit more than a quarter. So that means one in four times. So you can ask yourself as an intensivist if one in four is enough to make you think that an alert is a value to you. I know for many of our other models, it's much better, which may be a function of our workflow, maybe a function of the model. But it is what it is. So this was the trial as was done, and this is what we saw. In terms of escalations, and this is just sort of the univariate raw data, the overall escalation rate didn't actually end up being very different.
But people did seem to escalate to the ICU, you know, two to three times more often within 12 or 24 hours in the intervention group. So that's promising. So that shows you that. you know, the alert was finding people who were sick, we think, and they were getting moved back to an intensive care location more often than the controls. But the time lag was kind of long. And this, but this is overall not including, you know, including the ICU escalations, but also including a non-ICU escalation. So when we did, so there was that. So then this, now we're going to pause and we're going to say, well, what did the clinicians think? Because I showed you that the clinicians could actually reply to the machine learning alert. And even though it was unmonitored, we logged all this. So then...
In writing the paper, I made a word cloud, and you can see that the biggest, most common phrase was clinically stable. And I include this because I thought it was very interesting that, you know, regardless of whether patients needed upgrade or not, most of the time, people thought they were stable at the time being seen. But around the edge, you know, and a lot of stable, stable intervention just downgraded, remain on the floor, stable interventions will remain, remain, you know, just downgraded from SICU. So this was, I thought, this is a very qualitative analysis, but this kind of tells you something about the patients that were being seen or the impressions that people had. And again, the impression is a little bit different than the reality of what the study showed. Because when we did the full multivariate analysis with the inverse probability of treatment weights and the ward as an indicator variable and all the standard stuff you do in a multivariate analysis, it did look like patients who got the alert were more likely to be escalated. This is the top. We're more likely to have orders. And then...
just shy of one, seemed to have a reduction in hospital or three-day mortality of close to 25%. So this, I think, is what got the attention of the editors. And this is what the point of the paper was. And this is actually fairly in line with other recently published studies of... machine learning type alerts or command center type things, getting clinicians, and I include both mid-level providers and physicians, getting clinicians in front of patients is a good thing. It helps patients. And then you look at the ad hoc outcomes, which is sort of the things that we analyzed that weren't formally specified in our protocol. And again, you see that you're more likely to go back or go to an ICU within 12 or 24 hours. And this faster hospital discharge was marginal. So that was the outcomes. And here I just sort of highlighted it.
And we were very happy with that. And this was despite the fact that the whole project was terminated early because of COVID. And in fact, analysis of it was delayed for two or three years because of COVID. And again, during the review process, the editors asked us, well, your study was terminated early, how can we trust the results? So it actually, and we can see only 44% of the patients projected were enrolled. But actually, during the height of COVID, there was an article published in JAMA about what do you do if your study has been terminated due to COVID? And so they had some suggestions, one of which was to do this post hoc analysis of power, which we did. And you can calculate something called the futility index, which is also used in clinical, in randomized controlled trials with...
intermediate or interim analysis to decide if it makes sense to continue. And what's called the futility index was high. And a high futility index means that it's likely that even if we continued to our full enrollment numbers, we would not have seen a clinically significant difference in the escalation rate. And again, that's just the raw escalation rate. And so it's likely that despite the early termination, the results we found would have held. So that was one limitation. The other, which again is the nature of the study, is it wasn't randomized at a patient level. And this was a pragmatic decision because
It was fairly easy to randomize at a unit level. We identified the units we wanted. We engaged with the nursing leadership on those units. We had... information flyers created in four languages that were available on the unit that were placed in the patient's chart. And we convinced our IRB, and I didn't really make a slide on this, and this gets to the point that we'll come up a little bit, but we convinced our IRB that since this was an advisory alert only and didn't recommend a specific course of action, that the harm to the patient was minimal. And so therefore, we didn't have to explicitly consent every individual patient because we weren't actually doing anything to the patient. The only thing we were doing was having a provider show up at the bedside.
And there was nothing to prevent the provider from showing up at the bedside in the control units or even in the intervention units on patients who didn't get the alert. Another limitation was really that, you know, we didn't specify escalation prevention intervention. So we didn't say give 250 cc's of fluid to the patient who looks, you know, who meets these criteria. You know, give a bolus of labetalol to this patient. Give a, you know, glucose to the patient with a low sugar.
So it was really up to the provider, up to the RIT team as to what they did. So this is the downside of a pragmatic trial. You don't control everything that happens. But this is what happens in reality in the hospitals. You don't control everything that happens as much as you try to protocolize the care. The workflow that we ended up coming up with was super complicated. And I showed you that slide. And it differed a lot from what we did in the training, which was simply six hours before. You look back and you decide.
But you can't do that. And that's the difference really between a model that you build in a lab versus a model that you deploy. And I can tell you that the majority of the work, although the data science team spends a lot of time building models, the much, much harder part, the majority of the effort intellectually, I think at this point, goes into identifying a good workflow that will enable you, enable the clinicians to actually use the prediction in some functional way. So that's tricky. And there was a lot of learning from this project around the fact that, you know, the complexity of what we tried to do with real-time alerts.
real-time feedback was hard. Now, this was an interesting point that was raised by one of the reviewers that does activating RRT expose other patients on the study unit to higher risk? And at first I didn't understand this, but then I read this article that was cited. And the theory is that, you know, if you're spending all your time looking at a patient that you got alerted on, you might be ignoring some other patient down the hall who is slowly or even rapidly deteriorating, and therefore that patient may suffer harm. We disagreed with that. We didn't see that when we looked at patients on the units who didn't get alerts as a separate cohort versus those who did. We didn't see any particularly higher rate of escalation or death or mortality. But this has been raised, and I think this was a JAMA publication as well. So it got some visibility, but we didn't feel that that was that.
And then finally, you know, we built a model and like every other large health system, you know, we validate the model at our health system. Now, we happen to have a big health system. We're in a city that has a very diverse patient population. And so we think that the model was pretty well trained, pretty well validated and pretty well calibrated. But everyone's going to say, well, you know, you didn't validate it at 100 other sites across the country on 100 other different patient populations. So how do we know that this would actually scale or apply to some other system? And I think the reply to that is I think there's a convergence of all these models.
whether it's a random virus, whether it's going to be a chat GPT LLM model, they all kind of converge on the same set of clinical factors. And I think they're all going to basically have more or less the same predictive power. So they probably more or less could all be, as long as they've been, I think an important point is validated and calibrated across a sufficiently large and diverse patient population, they're probably generalizable. So that was the study. Yeah, I was just wondering, were there a group of patients who the model identified as high risk? The humans said they're not high risk and then they had a bad outcome, like mortality, intubation or something. I don't know if that was tracked. You know, I'm sure there certainly were. In what we ended up publishing, we didn't focus on that group. But that's the group of sort of false negatives where the model missed them and a clinician caught them.
But generally speaking, in the control group, which is all the places that the model didn't alert, we didn't see much difference. But it's a good question. Great talk so far. I just had a couple of questions. I think one, what was the, I guess, provider perception of all this? So actually here at Maryland, we actually had this rescue score too. We have to call it rescue score or patient decomposition risk assessment where it was an epic flag. And we had our own algorithm for trying to do these bedside huddles. And at that time, I remember, I think there was a lot of like provider fatigue. And I think you explained really well how to just narrow some of that stuff down.
But I know more often than not, it was a lot of complaining and whining from the providers. So we found the same thing. And over and over again, we found this, which is that nurses, PAs, case managers, social workers, they love this stuff. They love it. They think it's great. They will use this all day long. They're so happy. The nurses, when this was rolled out, they'd be like, this is great. I got an alert on Mrs. Jones. The physicians hate it. They hate it. They feel like it's a waste of time. And again, I put this slide back up because this is what everybody said. It didn't matter. And this was to y'all's question too. That patient's fine. Oh, their hemoglobin just dropped three points and they're fine.
I mean, it was really that. And it was it was herding cats. And everyone thinks that they're the best doctor and better than the machine. And so it's hard. And I'll jump forward to, you know, you know, rescue me 2024. Where are we now? And you'll see what happened. But yeah, we've had that. You know, we have another model that predicts 48 hour discharge. And again, the case managers love this. If this thing is down for even 30 minutes, they file a ticket. They're like, the model's down.
In two years, we haven't been able to get a hospitalist to look at this thing once. And so if the hospitalist is not going to look at it, there's only so much that the case manager can do. So it's a very interesting situation that we find ourselves in that we're trying to solve. But I'll take that to jump ahead to your point, which is that we did all this. Oh, my God. It was such a work. And so now where are we? Well, what happened was we finished this. We got through COVID. We wrote the paper up. And then hospital leadership said, well, we want to roll this out now to everybody. And how do we do it? And so then we decided.
You know, first of all, we got rid of the health system, decided to, you know, no longer use this third party text messaging app that we had integrated with because it was expensive and then it disappeared. And they want you all to use Epic Haiku and that built-in Epic chat functionality, which had some real limitations in terms of, you know, when we were using sort of our own text app, we could create our own custom groups like WhatsApp. Like you could create groups, but like in Epic, like in Haiku or Rover or whatever, you couldn't create groups that spanned specialties or care teams. It's like, it was quite difficult. So like we basically threw out the entire workflow and ended up with this, which is exactly what you said.
And this is the workflow of most of our other models, which is that it pushes the score into Epic. The score gets color coded based on the thresholds that we tell them. You can hover over the score and it tells you what it is. And then each day there's proactive rounding. And initially when we launched this, which was maybe last year, fall of last year, maybe, or like late summer of last year, 2023. Initially we had, there was a, we created a new team that was not ROT. We called it SST. So step down support team. And their job was just to do proactive rounding and look at the top 10 or 15 patients.
And then we created this feedback loop where there's a little flow sheet form in Epic where they do this. And this was based on work group feedback and things that we realized from the initial study where we weren't quite perfectly able to close the loop if we wanted to do model training to say how to make the model better. And so, again, this seems like it was working. And then what happened was, like many other institutions, we have a staff shortage. And everyone in critical care was being asked to do more and more and whatever. So about maybe six months ago.
the head of critical care said, okay, as of now, the physicians are not going to round on these patients. It's going to be an NP-driven service only. And so now it's been that. And actually, we haven't done any follow-up analysis recently. So we went from this super complicated workflow where it sends a real-time alert. It waits. If the prediction is less than 10%, it doesn't send it. If it only sends three alerts a day, but it was still sending three alerts a day. So now it's like you essentially generate one prediction a day. The NP is round in the morning on the 10 or 15 sickest patients.
And that's that. So it's in some ways quite a fall back to earth. But it's the reality of the way most of the workflow is, at least in our hospital, and just how epicentric the workflow is and how hard it is to do non-epicentric workflows. And again, we still we added these questions about palliative care, level of care, because, you know, you find that even on the step down, you get the patients who have been on a step down unit for 30, 40, 50 days. They're just their families don't want to let go. They can't find placement. They just they're these, you know, liver patients or.
whatever leukemia patients, and they just don't move. And so they would always be the ones at the top of the list. So we implemented some filters that once an NP or someone clicked on this, it would take them off the list, basically. And so that's the workflow we have now. And I think it still works. Again, we haven't analyzed it in a while because even if you take the futility analysis, we still have 2,400 patients that we looked at. So you have to let these things run for a while before you have enough events to be able to determine whether something is actually moving the needle.
you know, this brings to the last point, which is the great editorial. And I really liked this editorial by Dr. Weiss and McGarry. I mean, we chatted a few times when he was writing it, you know, about going from in silico to in clinical. So, you know, there's just so many people publishing models, right? But it's deploying the models that is so hard and where things fall down. And so there was just a few quotes that I thought were great. You know, the first one was that, you know, there's just so much challenge around deploying this. Nobody knows the answer really. There's no real consensus. The second one is fascinating because
So the FDA supposedly classified all this software as a device, right? But they actually, in the review, they didn't have any examples of deployed models that they used or considered or any studies of equity. And so it's like sort of a black hole, and I'm hardly an FDA expert or regulatory expert. But if you go and look at the FDA site and try and comb through this, it's kind of a total mystery. Like there are companies and health systems that are submitting their models and getting FDA approval or FDA clearance.
And then there are tons of places that are just rolling out models. And, you know, Epic itself, right? Epic itself is building lots of AI models and just pushing them out into their products. And as far as I know, there's no FDA oversight of that as well. So I don't even know what this middle paragraph means in reality for using this stuff. And then the last point is that, you know, so you don't really know what happens. In this case, he was talking about our model, which is clinical deterioration. But for any of the models and, you know, for some of our other models, you know, malnutrition, which is the one that won this awesome, this great prize that we were so happy and honored to win.
We then went and did a calibration analysis after we had run it for a couple years to see, is it biased against African-Americans versus non-African-Americans or age or ethnicity? And kind of the only way to tell this is to run it in the real world and see what evolves. All of the models that we build so far, they don't prescribe any specific action, which is why in our mind, to me, they don't really feel totally like devices. But, you know, who's to say, you know, the pulse ox is a device. It doesn't provide any action, prescribe any action. So I thought these were really, really interesting points. And I just love the bite to bedside. Like that was just a brilliant catchphrase. So I still look for the title. So it's been a very interesting, it's been a very interesting several years doing this. So that's it. So I'm happy to, you know, take further questions, have some discussion, whatever works for everybody.