Thanks, everybody. Thank you, Tomas, and to the committee for putting together this exciting platform to discuss innovation. In particular, kudos to you for including a talk about software, because I feel that's an often neglected and under-emphasized aspect of innovation. A few disclosures. I am the founder of a company called Swirl Technologies, but we're largely a nascent company at this stage. We haven't done any fundraising, et cetera. I'm an inventor of a couple of patents and have the funding disclosed there, some of which has funded the work I'm about to describe. But really, my objective today is not to convince you to buy any particular product, but really to convince you that innovating in the IT space is just as important as it is in the drug and device development space. So I'm going to talk about three things today, three ways that I think information technology can make our lives better and enhance what we do every day. One of them is enhancing the efficiency with which we at the bedside can consume patient data. Two is improve care in new ways that are currently really not possible. And a third, at the end, just for fun, is about streamlining documentation. I know we all came here because we want to know, can we finally stop writing notes? So we'll try to answer that. OK. So this is how every day starts in the Children's Hospital. This is a shout out to our amazing nurse practitioner team, who are really the bedrock of our unit as they are many units around the country and around the world. And so we all show up with smiles on our faces. We care about each other as humans, catching up on family life, et cetera. And then we quickly turn our attention to our patients, and very quickly our faces turn to this. So we have to pre-round. Pre-round. So as if rounding were not painful enough, we have to do it once before we round. So why do we do that? So at least at our institution and those that I visited, we pre-round because we have to synthesize data. So I need to get the intake and output from one place and the labs from another place, and I want them looking in these little stick figures. And then I have to write medications, not in alphabetical order, but really organized in a certain way that we all think about them. And so that's why we pre-round, OK? And then we round. So when we round, rounds at our place look a lot like this. So one thing that I hope you notice is that there's a lot of people and a lot of computers, but every computer and every person is in their own sort of silo. And the reason for that is written on the computers, because every computer screen contains, even if they're within one system, like a source EHR, contains data sorted by data type. So meds are in the MAR and the orders. Notes are in a separate section. Orders, labs are in a separate section. And then if you want continuous vital signs, maybe you have etiometry or sick bay. Maybe you have a PACS viewer that you do during rounds. And that splits up our cognition so that the synthesis of all of those data elements are done in the ether. They're done in between all of those computers and in our minds. And that creates a major cognitive burden. And why is medicine so complex? Well, because we have to answer a complex question that require the integration of multiple different data types. So take this simple example. Is this patient with sepsis responding to antibiotics? Well, you need the antibiotics from the MAR. You need the culture results from the microbiology viewer, hemodynamics from the monitor, et You get the point. And that just looks like this. It's a big mess. OK. So this infrastructure has significant consequences. It's not just annoying. It is consequential and expensive. So hospital errors, the majority of them are caused by a failed communication and recognition. And separating data elements creates this opportunity, the Swiss cheese model, to make it harder to recognize something. We all know that we spend more time with our computers than we do with our patients. We waste a lot of money. And I like that there's a report card for EHRs. And many of them, not all of them, but many of them get an F. So we're trying to innovate in this space. And I want to convince you that we should all be innovating in this space. It takes all of us to do this. So TWIST and SHOUT are a platform. I apologize for the name. I'll explain. So TWIST is an acronym for the way it should think. And then SHOUT sort of just happened naturally. SHOUT is the mobile app. And TWIST is the web interface. So but there is skin that goes around current systems. And so at our institution, we have a very fickle administration that likes to switch vendors often. So we have both EPIC. So we had Cerner. Now we're switching to EPIC. We had Philips. Now we're switching to Nihon, Coden, et cetera. So we have experience interfacing with lots of different vendors. We integrate data from PACS, the communication system called Vocera, and middleware devices that interface with devices such as ventilator and ECMO circuits, and several different third-party applications. So what do we do with this data? So I'm going to give you a couple of screenshots and examples of what our system looks like inside. This is TWIST. So this is essentially the top part of the screen that's like a navigation bar to the patient. And this is a respiratory support score that we described a few years ago, which essentially goes up and down over time and allows us to see changes in respiratory support. So whether the PIP or the FIO2 is higher, the score is higher or lower. And then you can see this line here represents mechanical ventilation. So if you're below this line, like in these spaces here, this is when the patient is on non-invasive ventilatory support. This colored bar up at the top shows the patient's location. So if they're green, they're at home. If they're red, they're in the ICU. And yellow, they're on a ward or in the emergency department. And you can see superimposition here of procedures that have been done. So it gives context to where the patient is in their course. Around medications, we do a lot. So we summarize orders by these icons over to the left. And we've really tried to iconize the MAR. So that allows us to consume and see over a long period of time, even months, if you're looking at a patient who you're trying to sedate and see what is their sedation history. This allows consumption of a lot of different data elements much more quickly. So you can see this here is the dexmanetomidine infusion going up and down. These are IV doses of ketamine, for example. These are infusions of hydromorphone and boluses of hydromorphone, et cetera. This allows us to time align the drug that we are monitoring as well as effects. So like up here is the Watt score looking at withdrawal assessment, et cetera. And this manifests in many places, anticoagulation monitoring, et cetera. This is a screenshot that one of my colleagues sent me in the EP division. They were having a nerd fest like they oftentimes do, trying to debate whether it was avabradine or procainamide that made this patient's tachycardia better. And he was thanking me because he was right that it's easy to see the correlation between doses of avabradine here and the changes in heart rate. So this is just one of the many examples of the power of juxtaposing different data elements together on the same screen and in the same system. Here's another example of a patient with a ductal-dependent lesion, ductal-dependent systemic perfusion in which we increased the dose of prostaglandin and saw an increase in their systemic blood pressure. We can also see x-rays rapidly over time. We ingest the images from PACS. So what you're seeing here is multiple x-rays over a period of weeks that we can just see quickly. And this is helpful to see evolution in lung changes over time, for example. OK. So back to this question, is this patient responding to antibiotics? Let's see how we answer this question. So in TWIST, you can see the infectious disease notes are these little icons. You can click on them and read the commentary. But you can also see the antibiotic history, what they're currently ordered for, their temperature curves. These are the cultures. Red ones are positive. And then you can also see hemodynamics. This patient is weaning on their vasopressin and has stable hemodynamics, for example. OK. So then there is a mobile application called Shout. This is much beloved. I apologize. A lot of it's going to be grayed out because of PHI reasons. But hopefully, you'll get the basic idea. So this is the census page. These numbers over here are the MRN. But they're fake. So don't try to look them up. You can. But nothing will come up. Sorry. I got distracted on that one. You can call any member of the team directly from your phone. You have a summary of the patient. As you scroll down, you can see labs as a table or as a graph. You can see all of the current medications. You can see the patient's lines, their intake and output on a granular level. You can see their x-rays. You can zoom in and out over time. You can see their current monitor. You can see data, vital signs, history with 24 or 72-hour resolution, reports, EKGs. All of the notes are manifest in this application as well. So does this actually do anything, or is it a cute toy? Well, we've started to try to understand that in a little bit more detail. This is work by one of my colleagues, Taylor Smith, who did a side-by-side comparison of asking eight basic questions. And they're listed here. But they're pretty essential questions. How many patients are on your unit? How many of them are intubated? What are your patient's current vent settings, vasoactive infusions, sedative infusions, et cetera? And what we found was a little bit astounding, that our current infrastructure does not allow providers to answer some of these questions in 120 seconds. Now remember, these are questions that we might answer multiple times a day on multiple patients. So these inefficiencies really should be multiplied by a couple orders of magnitude. And it really starts to make it obvious that the way that we process information really needs to improve. And our patient will be making more informed decisions, and our patient's care will be better. So let's now talk about innovation in the way that we provide care to patients. Now many people in this room and in this society have built very important statistical models for major outcomes. They're all shown here, mortality, cost, length of stay, hospital complications, cardiac arrest, readmission to the ICU, and many others. And I take nothing away from those. They're all extremely important models. But there's also an opportunity here for us to create models that do some of the things that we don't really, really want to do. So we ask a lot of mundane, but in the end, consequential questions and make decisions like that every day. And my least favorite one may be, at least at 1.30 in the morning, is, do you want a chem tomorrow morning? So this is a question that I think many of us would be kind of happy to turf off to a computer system. So we tried to build a model for this. So we said, OK, let's take the last decade of serum chemistries at our institution. And we looked at 100,000 of them. We thought that was enough power. 8,000 patients, we put in all of the predictors that you would want. So how much exogenous potassium were they getting? How much diuretic were they getting? Their dosing, their most recent potassium measurements, renal function, urine output, age, et cetera. And you can see here, this is the leftward skew of our neonatal population. We waste a lot of potassium because we're a cardiac ICU. We give a lot of diuretics, et cetera. So you can see kind of the peak is leftward skewed. And what we found was that we can actually create a model that is highly sensitive and specific such that we could decrease the number of potassiums that we would need without missing any critical values. So meaning that if the model suggested that the potassium would be normal, that we would skip it. So that would occur in 27%. And you would have no critical misses. Well, that was great. And we actually thought about incorporating that into TWIST. But then we realized something really important around software innovation. One of them is that this model was kind of specific to our microenvironment. And if we wanted to go to another unit, even another unit at the same institution, a lot of the founding premises that we had for that model would change, which would require us to revalidate and redesign the model. The second one is that when you're predicting something like a lab value, that's a regulated practice. And while it's not impossible, it certainly increased our barrier to entry. The third thing, which is really important to say, is that the dominant vessels, I have the last talk on my mind, the dominant variables in this model were all simple and intuitive. They were what were the most recent potassiums, how often are they abnormal, and are you replacing a lot of potassium. And also, potassium is really just one among many factors, of course, in a chemistry. So instead, we pivoted and built this tool, which is in beta testing right now. We hope to roll out in the near future at our institution, which is basically a communication tool in which we can, on rounds and then on overnight rounds, communicate with one another, yes, we want a chemistry, yes, we want a blood gas, yes, we want a morning X-ray. But we're doing so in the context of new information. So it summarizes how many chemistries were there, what were the results, how often are we giving serum potassium, et cetera. And that allows more informed decision making at the bedside. I'll keep going. Can I keep going? Next. All right. Then we read this beautiful paper that described the work of another group that essentially listed the volume of phlebotomy to the intake and output of their MAR. And what they showed, even with that just minor tweak and what providers were aware of, is that they significantly decreased the number of labs every day from 9.5 to 2.5 clinically and statistically significant. So we've added that into our platform where we're going to depict here how many MLs it will be to draw all of the labs that we would like. And maybe we'll start to think about it a little bit differently. So these minor tweaks, really just informational, but will provide new information and inform our decisions in a different way. Another question that I'd like to discuss, this is a little bit of an invitation to any people who are in the audience who would like to participate in a trial that we would like to design around sedation. So we first tried to build a model around making decisions about which sedative would be most effective. And we used a lot of morphine at our hospital, as you probably do as well. We looked at a lot of administrations of IV morphine. And we actually, not super surprisingly, found that only 27% of them result in a decrease in pain and sedation score. About half of them result in no change, and about the same number as decrease increase. And now there's a lot to say about this, but we did create a model that would predict unfavorable responses, meaning an increase with a lot of accuracy. And so we thought, well, maybe what we'll do is put this algorithm into twist, and we can let all of the people know, oh, the best sedative, eventually, the best sedative to use in this particular instance is X, Y, or Z. And what we learned is that that's also going to be very complicated, and adoption will become challenging because clinicians don't really like to be told what to do. So we've pivoted a little bit away from that, and we're thinking about sedation protocols. Now, we know that sedation protocols work. They've been well-studied, and I'm very excited by the work of this organization around ICU liberation. But even during the implementation phase, compliance with protocols is a problem. We heard a presentation yesterday where compliance, even during the implementation phase, was as low as 50%. Now, even 50% compliance with recommendations results in significant improvements, but there, of course, is attrition over time as clinicians' attention move on to the next initiative. Complexity is also a barrier to compliance. So this is our institution's current sedation protocol called Restore, and it's very complicated. And in particular, there are elements of this that are also very complicated. So let's look at this one, for example. So if a patient's SBS is more negative or as prescribed, so let's pause there, compliance with writing the prescribed SBS is quite low. It's also hard to find because it's embedded in the orders. Then we also need to know whether less than three non-procedural resque boluses of a sedative have been given in the last eight hours. So that's very difficult to see because where we record procedural versus non-procedural is in one place of the electronic health record. Then you need to count eight hours. It's just a lot of work, and not many people have the time or energy or impetus to do this. It might be a very great idea, but very difficult to implement. And so here are some ideas about what we could do in an interventional trial. One is actually annotate the number of non-procedural boluses that are effective or ineffective. So we can look at change in heart rate. We can look at change in SBS or change in other things and put a little X or different color shading on doses that are effective or ineffective or that meet certain criteria. Another one would be a sedation dashboard. So could we create an icon here that can show you all of the opiates that a patient has received and what their morphine equivalents are? You could look at cumulative opioid exposure, for example, and go through different elements of sedation in this way, creating a dashboard to raise provider awareness. Then could we embed protocols into the software? So instead of putting the impetus onto the clinician, could we put the impetus on the computer to say, okay, here's the phase of illness that we are in. Here's how many non-procedural boluses. This is what needs to be done according to the protocol. And could we push those to a mobile application or to a web application and prompt the healthcare team in that way? And then fourth, I'm really excited about creating new endpoints to quantify sedation. So many people are looking at SEDLINE, including Matt Lucchetti, one of our fellows at Boston Children's, using respiratory impedance to quantify movement or ventilator waveforms, and then even analyzing video in real time to look at movement as a way to look at pain and sedation scoring. Okay, last one. In a couple of minutes, let's talk about whether we can use a large language model to write our notes for us. So everybody here probably recognizes the collaboration that took place between OpenAI and Epic. I'm very excited to see what happens through that. I want to do a little bit of beta testing. Many people here are probably familiar with this paper, which is that the chatbot GPT is kinder and more empathetic than we are as clinicians. Okay, we'll keep going. So maybe note writing for informational purposes will be different. Of course, you know, the chatbot doesn't get tired and it's social, right? Like writing a nice note to a patient, I'm sorry you're not feeling well is different from synthesizing complex clinical information, especially so in the intensive care unit. So I created a fake patient here. Let's pretend that this is future facing. We don't really have these features yet, but let's pretend that there was a way for us to enter a physical exam using iconography and a future plan either by reading the orders or looking at an order set like this within twist, and we could populate backend tables. So this is really what API, that what the GPT or large language model API would look at. It would look at all of these data elements and then try to create a note. So here's a sample patient that I created. So here are the events. So this is a newborn child with hypoplastic left heart syndrome, partially restrictive atrial septum who was born on January 1st, came to the ICU and then went to the OR on the 4th for a stage one palliation, okay? You can see what their labs look like here. So this is the post-op period. They're very desaturated, have an SAO2 on the arterial gas in the 50s and a venous sat that's in the teens. These are all terrible if you don't know that. And then has a very high lactic acid. And so I've entered some of the physical exam findings and orders. So if you haven't played with GPT-4, I encourage you to, but this is basically what I did. I uploaded that spreadsheet of data. And I said, this is called a prompt. You're an attending physician working at the hospital, summarize what happened on January 4th in the context of the patient, basically saying, hey, can you write this note for me? And so you can see it's working. And in a moment, it will start producing text. Here you go, complex, critical day. So we're gonna look through this. I'm not gonna wait for this whole thing to play. I'm gonna show you some of the results here. Okay, so I like the fact that it wrote in human prose. It said complex and critical day at Boston Children's Hospital commenced. I like that. The day commenced with the patient undergoing stage one palliation, including a DKS, It included all the salient events of the day. So that was good. But then the good stopped because the patient did not get a five millimeter BTT shunt. The patient got a ringed sonoconduit. That's a very clinically important thing. So that's called a hallucination in tech terms where GPT will just sort of like make something up or insert something that's incorrect. And so it also inferred something that was not exactly right, that the patient was cannulated due to circulatory difficulties. I suppose you could consider profound hypoxemia as a circulatory difficulty, but I would have wanted it to be a little bit more specific. It really poorly synthesized these lab findings. And it also hallucinated that there was a concerning drop in platelet count. It totally made up all of these numbers. I have no idea where it came up with those. I didn't report a platelet count at all. So that's called a hallucination. It's a known problem with large language models. It did a good synthesis of physical exam findings, but then it also had some inaccurate recounting of a couple of different things. I won't get into it because it's really not critical. So can we automate note writing? The answer, at least in my opinion, is no, not yet. We still have that role, that onus is on us. But over time, this field is rapidly evolving. And so I do suspect that that will become more promising in the near future. One place that we do think that we're gonna be able to integrate GPT is for questioning and answering. So our hospital has signed a BAA with OpenAI, and we're in the process of standing up APIs for all of the patient data that is in the backend of TWIST so that clinicians can ask questions. So you can say things like, tell me about this patient, or why is this patient on antibiotics? And it will give an answer generally a little bit more performant than the one that I just gave, but it'll be really exciting to see how performant it actually is in the real world. Okay, I'm almost done. Challenging parts of innovating in IT. There is a significant barrier to entry creating the infrastructure. There are many fewer opportunities for funding in IT, especially to build the infrastructure than there are to build a new device. So I think that's a problem. At most institutions, IT is viewed as a commodity. And so typically under-resourced and overstretched. The customer, the clinicians, oftentimes love what we're doing, but the buyer says, well, what's in it for me? How's this gonna save me money? Those are typically the CFOs. And then changing a system is more difficult than providing a new product. But software can help us do a lot more than it currently does. It can certainly align with our cognition more in many ways, as I hope I've shown you. We can look, see more information with less of a cognitive burden. We can make more informed decisions and we can have more face time with our patients. I'm excited about embedding protocols. Please come and talk to me if you're interested in talking about that and creating new interventions. And have more fun. Medicine should be fun. Thanks very much for your time. Thank you.

innovation

healthcare IT

electronic health records

data integration

large language models

patient care