Hello, and welcome to today's webcast, Safe De-escalation of Care, Automating Clinical Pathways. In this context, care is considered reducing time on mechanical ventilation and prompt early weaning of vasoactive medication. This webcast is sponsored by Etiometry, where we enable our hospital partners to discover the ICU's untapped QI potential by managing their care escalation and de-escalation with precision. Developed by clinicians for clinicians, Etiometry's clinical intelligence platform is tackling the toughest challenges in the ICU. The outcomes for more than 150 quality improvement and research initiatives demonstrate the impact of Etiometry utilization, including significant reduction in length of stay, bouncebacks, and more. See why some of the world's top medical centers rely on Etiometry. Visit etiometry.com slash get your view. My name is Teresa Borum. I am the LearnICU Program Development Manager at Society of Critical Care Medicine in Mount Prospect, Illinois. I will be moderating today's webcast. A recording of this webcast will be available within five to seven business days. Log on to mysccm.org and navigate to the My Learning tab to access the recording. A few housekeeping items before we get started. There will be a Q&A at the end of the presentation. To submit questions throughout the presentation, type into the question box located on your control panel. Please note the disclaimer stating that the content to follow is for educational purposes only. And now I'd like to introduce your speakers for today. Josh Thalben is an intensive care medicine specialist, pediatric cardiologist, Boston Children's Hospital in Boston, Massachusetts. Santiago Borossino is section head and medical director at Children's of Alabama and associate professor of pediatrics at University of Alabama in Birmingham, Alabama. And now I'll turn things over to our first presenter. Thanks, Teresa. I appreciate the introduction and thank you to the Society of Critical Care Medicine for inviting Santiago and I here today. I'm going to kick off and talk a bit about the use of our risk analytics algorithms to inform our weaning protocols in the ICU for patients having cardiac surgery. Again, my name is Josh Thalben. I work at Boston Children's Hospital in the cardiac intensive care unit. And we've had the privilege of working with etiometry for almost 15 years now in the development of some of the software and algorithms that I'm going to show here. I have just one disclosure. We were funded in conjunction with etiometry on SBIR phase 2B grant that gave us the funding to do some of the research that you're going to see here, but I do not have any additional financial disclosures. So, this is the familiar scene on the right side of the slide. This is our old ICU. We just moved in a new building, which I'll show you in a second, but you can see the traditional monitoring devices, the pumps along the side of the patient's bay, the oscillator, there's a renal dialysis machine here, and in the middle of this giant mess of cords and cables and numbers and data, there's a baby and the arrow points to it there. The baby almost gets lost in the mix here. We're challenged every single day by limitation of resources, very high acuity, low clinician to patient ratio. The multi-professional team experience is changing post-COVID at our center and across the United States with temporary workers filling in in places that we would not have expected to see them in prior eras. And all of this creates a challenge of information overload, potential to miss, subtle physiologic changes of the patient, and, you know, as you can see from this, lots and lots of distractions being created. And this, I love this slide because, I don't know where, I think I may have stolen this from Peter Lawson, who works with us on some of these projects. This is the inside of the space shuttle, and I love it because, first of all, the blue color is amazing. It's very soothing in contrast to what we saw there. And this is just as complex of an environment, but it seems to me that it's much more controlled and a much more peaceful period. And unlike the aerospace industry, we're really challenged by the fact that our most vulnerable periods are often the periods where the lowest staff is there, the patients are potentially the sickest, and the communication can be the biggest challenges. The biggest, one of the biggest things we find is that the cost of care delivery is often a factor where you wouldn't tell NASA, hey, you got to shut off that blue screen over there on the right, or we're going to take away your ability to communicate here. That's often the case of what we see in the healthcare field. And when you put all this together, what our really big challenges are evolving to be is that we can miss things in the ICU. There can be clinical deterioration of our patients that lead to adverse events without the proper early warning and recognition of those issues. So this is our new center compared to the old center, but the gist is essentially the same, and it actually brings up new challenges compared to that original slide. It's beautiful space, wide open. You can see that the monitor and the display systems are cleaner, they're nicer, they're easier to see from further away. Patients have more room, but it is an immense space. These are not taken with wide mode, like, you know, kids always say, oh, don't use wide mode when you take photos. This is our hallway. There's very little line of sight down both hallways at the expense of having lots of space where it really created a physical barrier to communication in the new ICU. And what we tried to do in this right-hand photo here is to, again, make the displays a bit bigger. You can see our electronic medical record here. You can see the in the elementary display with T3, which we'll go through in a second to try to counteract that. But this physical space, this giant distance between patients, care providers, and colleagues has made it very challenging, has created another set of challenges. This slide, the animation is not working, but it gives you a sense of what traditional monitoring looks like, and I suspect it's similar across many institutions. We started out with a system that essentially shows in the back here, this Philips monitor has the waveforms, the traditional view of what you get from your patient. The nurse down here in the corner clicks on the EMR. Those values go into this really terrible Excel spreadsheet-type box, and then the fellows or the trainees in the morning will read off a list of data from the patients to try to draw a conclusion on rounds as to where the patients are going. And if you look here, these are either hourly entered by the nurse, or they enter them when there is an event or change or something going on. So a lot of selection bias in the data that goes in there, and the accuracy is pretty, it's not granular, but it's also a bit suspect. When you look at the graphs that we got from our traditional EMR, this is a blood pressure graph. I don't know about you guys, but I have a very hard time making any interpretation as to the trajectory of this patient based on these types of one hourly data, or maybe a couple of boxes here where the nurse decided something was relevant to click in that space. This is the platform that we developed in conjunction with the etiometry. This is, you can see, it's a very clean display where on the x-axis is the time, and on the y-axis are these different parameters, the IDO2, which I'll talk about in a moment, heart rate, oxygen saturation, blood pressures, filling pressures can all be seen along the display here. This little sliding bar, which again, I apologize, it's not showing up on the screen, but you can zoom in or zoom out to get as granular as five-second data. And then along the right-hand side, it gives you all the parameters that are hooked up to the patient. You see some of them here in a laboratory column, but you get all of the ventilator data, all of the hemodynamic data for the patient, and those can simply be dragged and dropped onto the graph, so you can look at those in real time. Since 2010, when we really started putting this platform into practice, we've collected a tremendous amount of data. This slide has actually been outdated by a year or so, but we've collected almost 150,000 patient data from our institution at this point. I think we're closer to 700,000 patient days at this point, and 15 million hours of data from our patients. These are sick, complex patients that have multiple sites that are being monitored, and it's all being collected in the etiometry system. So what do we do with all this? Aside from using it on day-to-day rounds and using it as a way to retrospectively look after an event, for example, we've developed these predictive analytics algorithms using the high-fidelity physiologic data that's coming from the patient. So we can now take the Q5 second physiologic data, combine it with the laboratory data, and put them into predictive analytics algorithms. The two that I'm going to speak about today are called IDO2 and IVCO2, going to focus mostly on IDO2. These are FDA-approved algorithms that are real-time models that show the probability of a mixed venous oxygen saturation being below 40 in the case of IDO2, or in the case of IVCO2, the PCO2 being above 50. These algorithms will be able to use, or they are able to use, up to 19 continuous data inputs, meaning the more numbers that you put into the model, the higher the predictive value of the model will be. And at this point, we're incorporating 19 possible values into these algorithms. And what we hope to show through our early research, and I'll go through some of that in a moment, is that we could potentially predict an evolving clinical deterioration. And then more recently, we focused on early de-escalation, which is what I'm going to focus on for the second part of my talk. So just very briefly, the predictive analytics models look something like this. This is an example of IDO2. It takes the 19 variables that are put into the physiologic model. We use a recursive Bayes theorem, we meaning the smart math people have come up with this machine learning algorithm, which then spits out an IDO2 calculation. And the IDO2 calculation, again, for orientation is the probability, in this case, of the mixed venous oxygen saturation at any given point being above, being below 40%. So this is how well the model works. If you use the full data set, meaning all 19 variables, and look at the receiver operating curve is the blue line here, you see a very high likelihood that your mixed venous oxygen saturation will be low. Not surprisingly, the AUC goes down as you use fewer and fewer variables, and the red dotted line is looking at just three continuous variables, heart rate, respiratory rate, and oxygen saturation. The model still performs quite well, but nowhere near as well as it does when you put in all of the inputs, all the potential inputs. So this is what it looks like in practice, and the IDO2 display is across the top part of the screen. I don't know if you remember, a couple slides back, it was entirely green. As the IDO2 rises, the display turns to red, and I have the line here where it's at 93%, which means there's a 93% chance that the mixed venous oxygen saturation is below 40. And as you can tell, the IDO2 was sort of smoldering here, got a bit better, and started to rise. And at the same time, you see some changes, but nothing dramatic in the heart rate, blood pressure measurements that we're seeing here. But then all of a sudden, you see, as the patient has an event, a narrowing of the pulse width and compression CPR started as the patient unfortunately suffered a cardiac arrest. The IDO2 very quickly goes back down into the green range because we went on mechanical circulatory support with ECMO, and you could tell that the oxygenation and the blood pressure ultimately went up, and the IDO2 reflects the improvement, so to speak, in the patient's status. So what does this do for us? We can use it as a bedside monitoring device, but how good is it at predicting or associating with an event in the ICU? So Craig Futterman and Avihu Kazit helped our group work on a study that looked back at patients who had a cardiac arrest compared to patients who didn't. We looked at the dose, the overall area under the curve of IDO2 for the 120 minutes preceding the event, so the event time you can see marked here. We had a censored interval because we decided that if you're going to have a cardiac arrest within 30 minutes, there may be other parameters that you see, and we're not really interested in knowing 30 minutes before. We like to know two and a half hours before so that we can potentially make an intervention, and we looked at how well IDO2 can show that this event's going to occur, and if you look at the receiver-operated curve for this, this is a random chance, coin flip, the cardiac patient's going to have an arrest or not. This is the IDO2 dose, that two-hour block of time, two and a half hour, I apologize, block of time preceding the event, how high the IDO2 is is associated with how likely they are to have an event. So this model, again, seemed to be quite predictive of a cardiac arrest event in our patient population. We're certainly not the only ones that are doing work like this, and I think it's important to note that there are other algorithms out there, other platforms out there. This is the group from Texas Children's who's developed a platform using their own real-time processing systems to do a very similar type of study that looked at prediction of severe deterioration in patients, and this was published several years before our study. So the next step that we looked at was to try to flip the script a little bit, sure, so we could tell potentially when patients are going to have an event using IDO2, but we now say, well, can you tell when you can take away care, de-escalate care, and it's a real frame shift for us. We use the funding from the grant that I mentioned earlier to put together a group of researchers from Children's Hospital of Philadelphia, from Boston Children's, Toronto Sick Kids, and Children's National Hospital, and we've looked retrospectively at first, and Michael Goldsmith led this effort, at patients who were able to successfully wean from vasoactive drugs. If you see a patient here, for example, patient A, they arrive in the CICU, they're weaned off inotropes, they don't go back on, that's a successful patient. Patient B weans off inotropes, they have a period, a honeymoon period, and then the dopamine in this case is restarted, that would be a failed wean from inotropes. We look back again and use the dose of IDO2, the amount of the red under the curve to make it easier to conceptualize, the prior to the wean in patients who were successful and patients who didn't, and in patients who were not successful, were not able to wean, they had a significantly higher IDO2, so the odds of needing vasoactive support reinitiated when your IDO2 was low was much higher than the odds of needing reinitiation when your IDO2 is high. And interestingly, these IDO2 doses, you can see the inflection point here, somewhere around 20 to 30% where the odds ratio starts to creep up for likelihood of needing to be put back on those drugs, is very low, so this dose is very low. So even a slight increase in your IDO2 may be a sign that it's not okay to wean, and the correlate of that is that a normal IDO2 may tell you that it's okay to go ahead and start weaning inotropes. So that was one aspect that we looked at. Danny Hames is a colleague of mine here at Boston Children's, did a great study that was published in the last 12 months that looked at associations between extubation failure and using these risk analytics algorithms as well. Danny's case used both IDO2, and I mentioned earlier in the talk, IVCO2, which is the likelihood of the CO2 being above 50, and when you combine both the IDO2 and the IVCO2 into our existing algorithms for extubation readiness, the predictive value of those readiness tests goes up, meaning that patients are less likely to need to be reintubated, and patients are less likely to need non-invasive, unexpected non-invasive mechanical ventilation. So we've been able to show, at least retrospectively, that the use of these two types of variables in an existing model may aid in our ability to tell which patients are ready for extubation and who is not. And Danny has gotten funding through several sources, including the Brett Boyer Foundation, to fund a study, a prospective study, to look at these weaning algorithms incorporating IDO2 and IVCO2, and those studies are now underway and hope to report some data for you soon. And then finally, we did, I mentioned earlier, we did a prospective trial at three sites at St. Louis and Children's National and Boston Children's Hospital, where we set up these temporary displays at the bedside on an iPad, these are touchscreen interfaces, and the provider was prompted when the IDO2 was low, so we took all patients coming back from the operating room after cardiac surgery that were under one year of age, we put these devices at the bedside, and we asked the provider simply to look at the IDO2 at the bedside and make decisions about potential weaning of vasoactive drugs. This is analogous to what Danny's going to be doing with the ventilator, but this was the vasoactive study. And that's all we asked them to do, look at the IDO2 in the context of the patient and consider weaning the vasoactive drugs two times per day, and we recorded the results from that. No recommendations were made based on this, but just prompted a discussion at the bedside. And what we found, and these are data that are in submission right now for publication, but we did present these data at the 2023 World Congress in Washington, D.C. a few months back. And our hypothesis, again, was that after initiating this algorithm, the math, we call it, that the infusion times, we hypothesized, would be shorter when you had simply added in the discussions of incorporating IDO2 in the decision-making. And post that intervention, we did show a reduction in the length, the duration of infusion times in our patients. And interestingly, I found this quite interesting, that the sicker the patients were based on their stat category, the more prominent the effect was. You can see the IRR for different stat categories here. And the biggest impact was in the sickest patients. So we showed overall 18% reduction in duration of infusion times once we used IDO2 as part of our weaning algorithm. So this is all great and reassuring, but can it improve outcomes? So what we didn't show in that last study is it didn't necessarily reduce the length of stay or the number of infections or some of the other secondary variables. But of course, many things go into those outcome variables. So how do we tell whether it's going to help improve the outcomes for our patients? And I think one of the things that we really need to focus on and we're working on in studies now is to understand why decisions are made. Putting a clinical decision support system in place is great, but the implementation and understanding the human factors is really key, and understanding the workflow and the environment is essential for this. And Sarah Thiel – I apologize that the citation got left off of here, but Sarah Thiel recently published a manuscript last month, actually, that looked at the – what we say we do and conjunction with these types of technologies when we put them in place. Sarah interviewed the faculty prior to the initiation of our study about how they think they make decisions, and then observed all of the interactions on rounds between the patients, the providers, and the technology, and tried to draw some conclusions about what she saw. And what she found was that we say we make decisions that are based on patient characteristics, and this chart, I know it's a little bit challenging to read, but it does make a ton of sense. The factors that go into the decision-making are bigger or smaller based on the size of the triangles here, and the associations, how near or far away they are, are based on the lines that you see on the chart. The predominant features of this that I'd like to highlight are that we say we make things based on the patient's physiology, and then incorporating technology and thinking about the laboratory values. But when the ethnographic observations were done, you could see that the technology pieces, the TES technology, is way farther away from the patient decision-making than we say. And there are many other factors, who the bedside nurse is, how much experience the team has that day, how many patients the respiratory therapist is covering. All of those factors really came into how we're making decisions in the ICU. So I think what this highlights to me is that we need to have a better understanding of how these technologies are interfacing with the care teams so that we can maximize the benefits that we're seeing in the retrospective and even the prospective studies. Modification of these human factors, again, is really critical here and needs to be the next phase of how we think about this. And once we understand the human factors that go into the decision-making for all of these patients, the implementation of these decision support tools is more likely to have a positive effect. So, in conclusion, I'll wrap up. The predictive analytics algorithms, such as IDO2 and IDCO2, may help us with early warning indicators, looking at adverse events, but I think the exciting potential component of this is that it could help us de-escalate care. And using a better framework for incorporating the human factors and the methodology of how we care for patients is really essential as we implement these clinical tools, the analytics tools. I just want to take a moment. I think there's a few key people that I'd like to thank. Craig and Avihu were instrumental in getting some of the early studies done. Michael was the lead on the study looking at the cardiac arrest data. And Danny Hames, of course, as I mentioned, is doing the excavation studies here. Peter was one of the early developers that worked with the audiometry on creating the platform and has been instrumental in helping us do some of the research as we move forward. And Sarah Thiel, who's the medical director of our unit, has looked very closely at the human factors and continues to do so to maximize the benefit of using these algorithms. And that is all I have. I'll turn it over to Santiago. Josh, you make it hard to follow you. I want to thank the Society of Critical Care Medicine for inviting me to participate. I want to up front say thank you to the team that helped me here. Ashley Mullinger, our lead nurse practitioner. Hayden Sankani and Ahmed Asfari. Those are my colleagues. They helped me out with this project in conjunction with the people of audiometry who helped us introduce the application to our platform. I'm going to be talking about a similar idea of what they've done in Boston Children's, but we decided to tackle on sort of making a more automated extubation readiness tool that will help us with the process of both deciding when to do extubation, yes, no, but also sort of reminding us that the patient was ready for an extubation spontaneous breathing trial kind of test. I'm going to go through this fast because Josh has done a great job at it, but basically a picture does not tell the whole story. You would think this is a very romantic moment between two people, but it was not. One is providing mouth-to-mouth resuscitation to the other one who was electrocuted a second ago. So you can get the wrong idea. to the other one who was electrocuted a second ago. So you can get the wrong picture. And the same thing about patients. You walk around and you look at the monitor and you might see a perfect heart rate for age, a perfect respiratory rate for age. But if you knew that the heart rate was like 50 or 60 points higher before or lower and the respiratory rate was a lot higher, it might give you a completely different picture. And that's what a platform like etiometry where you see almost second-to-second vital signs instead of every hour vital signs like in the flow sheets might help you. So when we make decisions in medicine, again, Josh spoke about this very well, you have the standing monitor and you're just standing in front of that room like the picture that Josh had, where it's like there's so much coming at you and your brain has to try to process it, all of it. And yes, with experience, we get better at it, but is there a better way to do things? In the old days, we had the flow sheets on paper and those were great because they put everything in one place. Different devices, different screens, everything in one place. But nowadays that we have the electronic medical record, the computer screen is just not big enough for everything we have on a patient. And there are new monitors that didn't used to exist when I started doing critical care back in the 2000s. So with that, etiometry has provided us with a really good platform to display all of them, not only the vital signs from the monitor, but also the ventilator, the vital signs from the monitor and other monitors like NIRS monitors into one place where you can see them and not only see them at one point in time, but see them over time. So you can use trends instead of just moments in time. And that's good. On top of that, Josh spoke about the data analytics. They have four different indexes that they've been developed and calculated and FDA approved that help us predict very specific things like low oxygenation index for delivery, cardiac output, high CO2 ventilation index, high lactate, lactate index and acidemia index. Now, but we wanted to move this a little bit further. It was sort of how can this helps us make decisions or improve our ability to make decisions by helping us introduce all the other data that was available. So the common question that we always ask ourselves is when to extubate. Extubation is one of those key moments in the care of patient where you start getting better and failed extubation is associated with bad outcomes of all types. So improving our ability to extubate successfully and improving our ability to make decisions about extubation earlier, both will be great for our patients. Extubation in general, there's very strong recommendations about how to liberate the patients from the ventilator. And the PALISI network has a document recommending protocolized screening, the standard ERT bundles, spontaneous breathing trial to be part of the ERT, to have a standardized spontaneous breathing trial, and have a spontaneous breathing trial of a certain duration. We thought that we could use the platform from etiometry to automate a lot of these things and make them happen faster or, you know, dare say, maybe even better. So, when we decided to do this, to create the algorithm to calculate the possibility of extubation failure or extubation success, there was a, the team in our center wanted to make it automated. We did not want it to have to depend on any human being introducing any new variable or having to calculate anything because we thought that that will just hinder the process. And those are the kind of times when you have a human, one of us in the team, just forget about doing it and all of a sudden you don't have it. So, on purpose, we did not include a weight at the time etiometry was not able to do weight. So, we wanted to just basically do vital signs. The other thing that we wanted to move away from was every patient is a little bit different, they're special in their own way. So, we wanted to compare the patient to itself when they were on a higher ventilator setting and then compare what they did during the spontaneous breathing trial. So, this on purpose is the patient comparing, compared to themselves with more support to see what the changes were, changes were on key vital signs, changing from one to the other, to use that information to help us decide, is the patient ready to extubate? These are the target rules, variables that we use. Some of those variables had different cutoff points and all of this was basically based on our experience. There's no magic to these numbers. We thought that this could be good numbers to try and that's what we did to create the tool. So, then we wanted to, again, introduce eligibility criteria so that the platform will remind us that somebody was ready. Our rules are ready to do spontaneous breathing trials, we just introduced them into the software to alert us that a patient was ready. And then we, luckily for us, we already had a spontaneous breathing trial that was very, very much standard, especially the one on the left with pressure support. We don't do the volume support too often, but the one with pressure support is basically what we do on every patient. The one with pressure support is basically what we do on every patient. So, it was very easy for us to, A, know when a patient was ready and then it was very easy for us to identify, even retrospectively or prospectively, patients that were getting a spontaneous breathing trial and then the decision to extubate. We do a spontaneous breathing trial on a majority of our patients, not all of our patients, and you'll see when I show you the data of how frequent that was. So, how do we calculate our probability, our compliance? Well, we established a baseline that was about 30 minutes. We calculated the baseline for the different variables that we included, and then we looked at every minute of the whole length of the spontaneous breathing trial. If they met, if they stay within the, if they were compliant with the rules or not, and if one variable was not compliant one minute, that was a minute that was not compliant, and then we calculated a percent of time that the patient, a percent of compliance with the test. So, if a patient remained within every single variable, didn't violate any of the rules above or below our variable's thresholds, the patient had 100% compliance. If the patient, the moment we went on spontaneous breathing trial, violated all the rules immediately, all 30 or 40 minutes of spontaneous breathing trial, the probability would be zero. Those are extremes. And here in this slide, you can see the little Ts of compliance, a little F of failures, and you can see the different times. So, each minute, the computer, the algorithm will calculate is the patient passing or failing, passing or failing on each one of the variables. So, initially, we needed to see if the algorithm had any hope to work, and choose which variables were going to be used for and what cutoff points. So, we looked at the data we had in etiometry, and we found that we had 572 excavations, 470 since etiometry was installed, 459 that had data on etiology. Then we found that 347 actually had a spontaneous breathing trial prior to the excavation. So, we were able to analyze that data. And then we had enough data for 339 of those excavations. We had 17 failures and 322 successes. And the failure was defined as 48-hour reintubation. And with those tools, we then tried out all of our rules. Initially, we had four rules that they were limited to then four rules for spontaneous breathing trial. And then we started looking at the different thresholds of each of these four rules. And we ended up with a four-rule pulse ox, mean arterial pressure, minute ventilation, upper, and respiratory rate as the four rules that we introduced in our model. And we had AUC of 0.72. With a compliance up to 25% compliance, predictive excavation failure with odds ratio of 5.75. And then it went up, went down to 3.63 and then to 2.09 and then to 1, which was the reference 75 to 100% compliance was the reference value. So, you can see that there's a very nice dose effect. The less compliant, the more likely of excavation failure. And with that, we had a tool that we wanted to sort of introduce to our attendings, to our respiratory therapists, to our nurses. And we were beta testing. And so, we just put it out there and see what usage we had. And we wanted to see, you know, and then sort of like a QI project, PDS cycle it and improve it over time. After about a year of having it out there, we decided to look back and see how we perform from the standpoint of total ventilation time. And length of stay. And again, here we just introduced the tool. We didn't make any big push or mandatory interventions related to the tool. And I cannot tell you how often it was used versus not used, unfortunately, because we were sort of like a preliminary beta testing period. And this ERT pathway tool was embedded inside of the etiometry. So, it was everything automated. We did not need to introduce anything. We did not need to turn anything on and off for every patient, for every extubation trial. Everything was automatic. So, we deployed it. Pre-ERT was December 1st, 2021. And post was up to January 1st, 2022. We decided that it will be very, very important to do some controlling of the data. So, we used STAT category, age, chromosomal anomalies, or syndromes, and those were things that we sort of extrapolated from PC4 studies looking into extubation failure. And those also were data that we had through our internal collection of data that will be very easy to obtain. We wanted to make sure that we kept it simple. And again, the goals were to see if our total ventilation time or hospital length of stay were affected. And then, in my experience, neonates behave completely different than the rest of the population in cardiac critical care. So, we set up that we were going to look at a sub-analysis of neonates only. These are patient demographics, the pre and the post. And there's definitely a difference in the number of patients we have pre versus post. And that's related to time, but there were no major differences in the composition of those groups. So, once we analyzed the data, we had 500 versus 193 patients. And what we found was that the post-deployment period had a reduction of 22.5% in total ventilation time and a reduction in 19% of hospital length of stay. And when we look at the neonates, we saw a significant reduction, even stronger signal, but wider conference interval of reduction in total ventilation time, but no association with prolonged hospital stay. This is the table with the adjusted analysis of the whole population. You can see the odds ratios are below one for total ventilation time and hospital length of stay, a decrease of 22% and 19% for all ages. And, of course, there's other variables that are well known to be associated with both length of stay and total ventilation time. And then we did an adjusted analysis for the neonates only, and here you can see the strong signal, but with wide conference intervals for total ventilation time, but not for hospital length of stay for the neonatal group. So, we concluded that introducing this ERT pathway application to our radiometry platform was associated with decreasing total ventilation time and hospital length of stay in our center. But we take this with a really grain of salt because, unfortunately, I don't have data for you to tell you that it was used at 90-something percent of the time, that it was used by everybody. It was turned on and off, and that's, unfortunately, all I can tell you. To me, it's very exciting because a lot of times the thing that delays patient improvement or weaning of the patient is not the patient itself. It's sometimes the human factor of being distracted or not being reminded of having multiple patients at the same time or just thinking out of fear that the patient is not ready. And a lot of times the patients are ready faster than we think. I think this is very similar to what Joshua showed in his inotropic talk where the patients might be ready faster and you can actually wean them faster than you would think without a tool that tells you or to give you confidence that the patient is ready for it and that reminds you, hey, the patient is ready, go do it. Thank you for the opportunity. Thank you, doctor, so much. We do have, it looks like, a couple of questions. First question here says, thank you for this presentation. What were the absolute reductions and comparative in TBT and length of stay? The absolute reductions, I don't have, I will have to get out of here. They definitely, they were not significant without the correction for the different variables, but they were there about, I'm trying to remember, maybe like three or four or five days for length of hospital stay and maybe five, six hours for ventilation. But I don't have in front of me. Thank you. Next question for Dr. Salvin. These results are focused on pediatrics. Can this approach with the IDO2 be applied to adults as well? Yeah, it's a good question. I think our data, our preliminary data is focused on our pediatric patients mostly because of the homogeneity and the types of operations that we're performing and the age of the range of the patients that we saw. I don't see any real barrier at all to applying this to adult patients. We have patients in our ICU who are adult congenital patients who are operated on our institution. We do use IDO2 in those patients, we just haven't included them in the studies. One of the potential limitations in adult patients, which is not specific to the algorithm, is that the monitoring that we use for the adult patients tends to be a bit less robust. So, where we may have 19 or so inputs into the IDO2 model for a neonate, a pediatric patient who is under a month of age, in an adult patient, we'll often limit that to four or five potential variables just from a practical standpoint of what type of monitoring they're using, how long they're expected to stay intubated, what their vasoactive requirements are, not in all cases, but in many adult patients. So, that's why we chose sort of the sicker cohort of the younger patients that have more monitoring. But from a conceptual standpoint, no, I don't see any limitation to using these algorithms as part of decision support systems for adult patients as well as pediatric patients. Great. How about the actual implementation? Looks like I see there are lots of monitors informing what you're seeing in the platform. What was the effort to even get it up and running? Santiago, do you want to go first? I guess the effort for us was doing all the meetings to come up with an algorithm that will work and all the analytical aspect of coming up with an algorithm. But once we came up with an algorithm and we deployed it, it was up there for it to be used and it's very easy to use. I personally use it very frequently for my excavations. It not only gives me the percent, but also it allows me to look at the whole period of the spontaneous breathing trial, and it highlights the times when the patient was out of range for those rules, which are useful. I think the rules were very intuitive and everything is automated. So we don't lose a single spontaneous breathing trial that doesn't activate this because it's all done by the computer. We don't have to do any clicking or reminding. It's always there. So from that standpoint, it's very easy. Would those numbers or those rules apply for other hospitals? Not necessarily, because there's no standardization across hospitals of what a spontaneous breathing trial should be. But I think that the process was relatively easy to do to create our own rules and to apply them. In response to... Oops, sorry. Go ahead, Dr. Salvin. Oh, no, sorry. I was just going to add to what Santiago said that I think this is a very good question, and it kind of comes back to the human factors component of this. I think once people see the value in utilizing the system and the algorithms, the weaning protocols, the adoption rate goes up exponentially. So we started out, as I showed you, with the little iPad system, because we didn't have the devices at the bedside mounted on the walls. And that brought attention to the utility of the platform. The providers would see the iPad on the rolling carts at the bedside, and they would say, okay, well, what is this? And then you interact with it, and then you start to use it, and then you see a difference in outcome. And people really started to adopt. The adoption rate went up quite a bit. So that when we moved into our new building last June, same as Santiago, there's countless meetings and funding and all of that stuff that goes along with it. But what we convinced the hospital to do is put in these large touchscreen displays at every bed space that has the T3 platform with the algorithms up there. And you walk by the bed space. It's on a swinging arm. You could pull them out. You can sort of face them outside the room. So you walk by the bed space and see them from the door. You can interface with them on rounds. So I think the adoption and understanding how people are utilizing them is key, and then getting the administration from the standpoint of the cost of installing those, which is relatively inexpensive. But the commitment to that was pretty easy once the adoption rate went up. They're at every bed space. Now there's a permanent, what we call persistent display at every bed space in our new tower and the new ICU all across all 48 beds. Thank you. This question looks to be in follow-up to applying this approach to adults. What variables in adults would you use? It's a good question. All of the variables are the same. So the model doesn't distinguish between adults and pediatric patients. But just back to the slide, if you recall, with the area under the curve, the power of the algorithm goes up. The statistical power of the algorithm goes up as you put more variables in, which is not surprising. So the neonates tend to have the normal blood pressure, invasive blood pressure, oxygen saturation, heart rate that every patient, almost every patient in the ICU will have those. We'll tend to have often two intercardiac lines, one measuring right-sided and one measuring left-sided pressures in every single patient that's in the ICU. There's laboratory data that's drawn every couple of hours that gets incorporated into the algorithm itself. And there's other data like temperature, NIRS monitoring that goes into almost every single neonate patient, small patient, for example, that may just not be followed in adult patients. So again, the statistical power of the model in adults would be the same if we had all those monitoring devices. But because the adult patients tend to be less unstable, acutely unstable in the early post-operative period, we tend not to use the intercardiac lines, for example, or a continuous temperature probe or a NIRS monitor on those patients in the immediate post-operative period. So it's not so much that the model's different, it's just how many inputs are going in, and that affects the sensitivity. Thank you. Next question, would it be possible, or is it already implemented, are these protocols already implemented through the EMR? So, sorry, Santiago, you have... No, I will let you answer. The EMR talks to, sorry, the laboratory data that goes into the EMR is also sent to the platform. So that's the way it's sort of connected. I think that there's ongoing attempts to get more of the data from the EMR that can help us sort of establish that a patient is sick into the platform. But it's more like the data from the EMR comes into, sorry, the data that goes to the EMR will also go to the to the geometry. I think that's the way it works and not that the data that a geometry has comes from the EMR. I think that there's still ongoing work to get more and more data that could help improve the algorithms out of the EMR into the platform. So more is coming. But the platform can be turned on through the EMR, at least in our center with the EMR we have, when you are on your own computer, which is different than the ability to look at the platform while you're standing on the bedside with the monitors that Josh described. I don't know if that's what they meant with the question. Josh? Yeah, similar here. The EMR, if you go into the patient's chart in the EMR, there's a tab for the T3 platform that you could pull up. So there's the screen I showed you before, that Excel spreadsheet type screen, which I find fairly not helpful. And then there's a tab for the geometry platform, which then you can populate right in that same space and get the more granular data that the T3 is pulling in directly from the monitoring devices. So there is integration between the EMR and the platform, in addition to the persistence displays that I mentioned. Thank you. The next question here says, if you're being more efficient, do you see your vasoactive weaning approach with etiometry having an impact in staff numbers or how your unit is staffed? Yeah, that's a great question. I think the staff numbering or staffing in the units right now is a real critical point that we need to focus on as healthcare providers. We've seen, again, as I alluded to, a real change in the potential working pool of nurses, physicians, respiratory therapists, pharmacists since COVID. And it's something that we really like to focus on. I don't know that something like a weaning protocol necessarily that we've seen an impact at this point, although we haven't really studied it. What I will say is that with the different team dynamics, either less experienced providers in some cases, or more likely just providers that haven't worked together as much, these teams are somewhat temporary. And what I've found is that the IDO2 and IVCO2, and just having the T3 platform up, has allowed us to get a 30,000-foot view of the units. You can go to a screen that shows all 48 patients, shows their most recent parameters, IDO2, for example, on the right side of the screen. And we talk about it, about nursing assignments, for example, or which providers are going to care for each patient. So it's not necessarily that we've reduced the ratios. We haven't. We haven't gotten to that point. But it allows us to target levels of experience, potentially doubled up assignments for a nurse, for example, based on the real sense of how sick the patient is using the algorithms that we've spoken about today. It'd be great to think that one day we could back off on the number of providers. Well, I don't think we're quite there yet. I want to comment on the automated part of things, like having a platform remind you to do tasks that are important for the patient. In this case, both a ventilator weaning of our extubation decision and weaning of the inotropes. When you're young, and we all were young physicians, nurses, or respiratory therapists, you're sort of in survival mode. You're trying to make sure that the patient stays OK until the end of your shift. You're nervous. You don't have enough of bandwidth in your brain to remember to move the patient along because you're so afraid that something bad is going to happen. A lot of your fear and attention is on that aspect of care instead of weaning. I think that having a system that tells you, hey, your patient is OK. It reminds you, hey, it's time to wean, will be very useful for any of us when we're new to a task. Any task. I think that automating the reminder alone, it's a plus in this whole endeavor of trying to liberate the patients from inotropes and ventilators and getting them out of the ICU. Staying in the ICU for too long is not a good thing if you don't need it. I agree. The last thing I'll say about this, just to build on Santiago's point, which is really important, is that these tasks are new to many of our providers and orienting them to the importance of that or automating some of that. It takes away that sort of mental load when there's so many other things going on. One of the things that I found is really important to thinking about staffing is we have newer staff, younger staff instead of people like me in Santiago. These guys coming out of school that are maybe less experienced in the clinical world, what they do very, very well is interface with the technology. This stuff is seamless to them. They love interfacing with a platform like this, getting these automated reminders. It's what they're used to in every other aspect of the technological parts of their lives. The adoption is much simpler, paradoxically much simpler adoption and utilization in our providers who are younger and potentially more experienced. Have there been any attempts to use this platform to help guide weaning from MCS? No. When a patient's on ECMO, the algorithms essentially are tricky to use because the physiologic parameters are somewhat, they're all controlled for lack of a better word. You may have a single ventricle patient and you expect their oxygen saturations to be an 80% range. In this Bayesian model, we'll look at the most recent numbers and weigh them a bit more and calculate IDO2. If all of a sudden it's 100% because you've got an efficient membrane, the algorithm doesn't really know what to do with that. Almost always, I should say while they're on circulatory support, the IDO2 will be low. What I could envision doing as a study potentially is as you start to wean off of mechanical circulatory support, you can look at changes in IDO2. We do do that at the bedside. When you clamp the ECMO circuit, for example, we could see some changes that would indicate whether the patient is doing well or not. But it's always, again, these models are based on what the patient's more proximal physiologic state was. If you've been on ECMO and then you come off, you're going to have a rise in IDO2. I haven't found it to be a huge help in that sense, but it's something we could potentially study. At the least, it's a great platform to show different providers what happened when you clamped. I think that is very useful, because then we don't have to go by memory. I think the heart rate went up, or this do, or the NIRS went down, the CVP went up. You have everything in one place, and it's recorded, so we can show it to each other and try to figure it out, bring other attendings into the fold after the clamp. It's like, hey, what do you think about this? What do you think about that? Should we do this? Should we do that? What do we try next? It's a great tool for that, but it's still, from that standpoint, it just becomes a way to have a flow sheet that has vital signs, data, and ventilator data in one place every few seconds, which is extremely useful, but it's not an algorithm. It's not a computer helping you make better decisions. It's just that it becomes an amazing display. Yeah, I completely agree with that. It's a little harder to make stuff up when all the numbers are there in a five-second review, and you can fold other expertise into the mix if they're not there in real time, so totally agree with that. Here's another question pertaining to efficiency. Do you perceive a cost impact in regard to earlier vasoactive weaning? We did some cost analysis, and again, it didn't because the length of stay wasn't impacted, which is really challenging. We had patients stay in the ICU simply because there's a bottleneck on the step-down unit, for example. That's 80 percent of our bottleneck occurs downstream, so the short answer is no. We did not show a significant cost increase in our preliminary analysis from these algorithms or these decision score tools related to weaning vasoactive drugs. I think we need to do a little bit more work and look at the overall hospitalization costs and see where we end up, but not immediately, no. I think you will need a much higher because the power that you need to show that the ventilator time or the hospital length of stay are shorter is completely different than, let's say, the patient was discharged from the ICU, the lines were taken out earlier, and you have less CLABSI because CLABSI doesn't happen, knock on wood, very frequently, so some of the things that affect the cost and affect the patient directly or more directly or more severely are rare events, but they are, in my opinion, they are preventable if we can get the patient out of the ICU, so from that standpoint, the power will be different to show something like that, but it has potential if we get the patient out without central lines, without cardiac lines, etc. So I have hope, but no proof whatsoever. I do just have a couple more questions. This one for Dr. Borastino. For those in the NICU, there's some variation in developing lungs. How do you incorporate AI and algorithms for these tiny patients? I'm not an expert in premature babies, but I think that the power of comparing the baby to itself with or without some level of support that neonatologists are experts on will think it's appropriate, could give you information on how the patient is or how ready the patient is for extubation, and this is basically what we did. Is it going to be the same variables? Probably not, but does it have physiologically the same potential of choosing the right variables that apply to premature babies and then use the platform to help you calculate a probability of success? Yeah, I think this, and the same for adults. The variables might not be the same ones, but the concept is very much the same for all patients. When you do a spontaneous breathing trial, you're taking a big chunk of the help that you provide with the ventilator to the patient, and how the patient reacts to that change is what we are measuring, and the variables that we chose were important to us, but for different populations, you might choose different variables, but you might get the same success potentially. And then finally, for both of you, how do you like using the platform personally? I don't know if it's a fair question for me, because I'm an early adopter in most things, and we developed the platform here, so it's been almost, like I said, almost 15 years, 12, 13 years, I guess, of really interfacing with it. It's part of my workflow on every patient every day. I'm one of the guys that walks around with my iPad with the T3 platform loaded on that. The fellows know that they shouldn't talk to me unless they have T3 pulled up, and we can have a conversation. I use it for remote monitoring of patients at home when I'm backing up the junior faculty, for example. I don't know that I'm a fair person to answer that, but it's part of literally every patient workflow that I encounter. I'm not as early as a doctor as Josh is, but we've been using it now for six, seven years. And before I go, I have to discuss, I don't get, I'm not in staff, I don't get paid, I don't get any benefit for liking the platform. I just like the platform. That's my bias. I like the platform, and I like using it, and I use it a lot, just like Josh said, but this is not paid advertising. This is advertising from somebody who likes it. That's it. Wonderful. That concludes our question and answer session today. Thank you so much, Dr. Sullivan and Dr. Borristino, and thank you to the audience for attending. Again, this webcast is being recorded. The recording will be available to registered attendees within five to seven business days. Log into mysccm.org and navigate to the My Learning tab to access the recording. That concludes our presentation today. Thank you, everyone. Thank you.

risk analytics algorithms

ICU

weaning protocols

patient outcomes

Etiometry

real-time physiological data

clinical decision-making

IDO2 algorithm

IVCO2 algorithm

healthcare providers