Latest / The JEMS Report / Predicting Time‑Critical Interventions and Smarter Dispatch

Transcript

Predicting Time‑Critical Interventions and Smarter Dispatch · The JEMS Report

• speaker-0: Everybody, welcome back to the Jumps Report Podcast. Excited to be here with a couple of new friends of mine today to unpack a little bit of evidence that ⁓ that was sent my way by Dr. Remily Crow. So she is she's on our advisory board. She made this post on LinkedIn and was one of the co authors in this paper and was able to connect me with ⁓ primary author. We had a great conversation and I wanted to continue that here. So I'll bounce it over to both of you so that you can introduce yourselves and then we'll talk about ⁓ we'll talk about the work that you're doing. It's really exciting. speaker-1: Sure, yeah. ⁓ Aaron Aaron Weedman. ⁓ I'm a PhD originally in social sciences, but now working in emergency medicine ⁓ and a data scientist. So, you know, I lead the technical analytics side of our machine learning projects here in the emergency medicine department at the University of Pittsburgh School of Medicine. and ⁓ that's what brought me to this work that we'll talk about today. speaker-2: And ⁓ I'm one of Aaron's colleagues. My name ⁓ Lenny Weiss. ⁓ I'm a medical doctor, emergency physician is ⁓ where I have my formal trade. ⁓ but I enjoy working outside of the hospital rather than in. So I ⁓ am an EMS physician. ⁓ most of my time is spent in medical direction of ⁓ air and ground medical services. ⁓ and in addition to that. ⁓ I work with some of our law enforcement colleagues ⁓ for tactical medicine and ⁓ medical direction for our county nine one one center. So a lot of this ⁓ work that we do as a team in our department is you know trying to use technology, use data to save lives in real time. And we're really excited to to partner. speaker-0: Awesome. Yeah. And and I'm excited to learn more about it and be able to share that with with our audience. So Dr. Weedman, I had ⁓ you know, the conversation that we had was on your original paper that's already been published. And I'd like to talk about that a little bit more, kind of get into the machine learning side of what we can ⁓ what we can pull and parse out of ⁓ of patient complaints, right? Or or primary ⁓ primary observations from the provider. ⁓ so can we kind of talk about that a little bit? speaker-1: Yeah, definitely. So I think the the core idea that came out of you know that led to this work and has come out of our research group in the past couple of years now is that the words that pre-hospital providers say or write down ⁓ can be useful sources of data ⁓ for driving patient care out of the hospital, as Lenny was saying. ⁓ and you know A big reason to look at natural language is that ⁓ as as you both know, the out-of-hospital environment can be relatively data poor ⁓ and sparse compared to an emergency room or another facility-based location. So you're scrounging around for any source of information that you can come by to give you a sense of what's wrong with the patient, what kind of ⁓ event led to their injury or illness, you know, what kind of intervention they need. And so we landed on natural language. It's a relatively underutilized data source in pre-hospital care. Some systematic reviews have have highlighted that. But we thought one that would be promising, especially because there is a lot of work showing that clinicians' gestalt impressions of what's wrong with the patient can actually be very useful and predictive. And And what is a gestalt impression but sort of a a jumble of words in your head or maybe that you even note down ⁓ as to what's going on. And so that was kind of the original idea behind this work is like let's look at words and use that as a predictive data source for important clinical outcomes that we want to explain. speaker-0: Yeah. speaker-2: It's it's interesting, you know, just to kind of highlight the the problems we see as Aaron is mentioning, when you look at a chart, a medical record, there is so much prose in there that is just what the human says and describes. They try to translate what the patient is saying, they make maybe some biased interpretations, they might have heard it wrong. Or they heard it all fine, but the way they describe it is not a tabular, ⁓ you know, form filled out way. ⁓ there are some fields that are, you know, tabular data and we can look at those, but ⁓ the juice is, you know, in the emotion. And ⁓ and Aaron is very talented that, you know, he's able to start to to pull that ⁓ from the records, which really has not been well done before. speaker-0: Yeah, I thought it was very interesting. Some of the you know, some of the the conclusions and observations that did come from the paper, right? That you were able to identify ⁓ you know, some complaints where there was a higher likelihood of a time critical intervention, right? ⁓ and that's you know, that's obviously what we're what we're trying to get to and you know what we want to try to navigate to as quickly as possible just because it is time critical. So can you it kind of explain that a little bit more? speaker-1: Yeah, so the good news is you know, where the juice is, like Lenny was saying in the the chart, that unstructured text, we're lucky that there are very, you know, well developed ⁓ machine learning and natural language processing techniques to take free text and basically turn it into into numeric data ⁓ that can be used to make modeling predictions. And so that's what we did in this case. We had We had two cohorts from the ESO data set twenty twenty three and twenty twenty four. Each of them had close to ten million records. And we took the chief complaint field. speaker-0: Could you have got a bigger sample size? speaker-1: you know I'll try better next time. Sorry, Mike. ⁓ actually you can direct that you can direct that complaint to to the ESO headquarters. ⁓ we'll we'll throw it back to them. But yeah, it was it we were lucky, you know, you don't always get to have an eight digit number to work with is your sample size, so we'll we'll take it. So the chief complaints though, as big as the data was, the chief complaints were equally as small. So we're talking two or three words ⁓ on average. Respiratory distress. That was a really common. lift assist. You know, these are really succinct statements about what's wrong with the patient, sometimes longer, but often very short. ⁓ And so we just we use natural language processing technique to take those words, basically crunch it into frequency indices that look at, you know, which words are present in a chief complaint, how rare or common those words are across the data set. And like you alluded to Mike, output a prediction ⁓ for each patient of whether or not they receive not only any time-sensitive critical care intervention, but which one. So it could be airway management, could be hemorrhage control, ⁓ blood transfusion, ⁓ cardiovascular, so CPR. And we found that for most categories that the models were pretty accurate. We we could do a pretty good job by machine learning statistical standards of ⁓ predicting. And identifying which patients received a certain critical care intervention based on just that that very brief chief complaint text. ⁓ And so doing so in a very large ground-based EMS population, ⁓ you know, that's representative ⁓ of how EMS is conducted in this in the United States was a big feather in our capus. speaker-0: well. Yeah, absolutely. And I think this is just kind of starting to to pave the way, right? For where some of this some of this machine learning can be can be utilized to ⁓ not make decisions, but assist in decision making for the for the provider. You know, ultimately allow them to perhaps be able to weigh a greater number of data points and get to that decision faster. not to make the decision for them because I think that's a that's a place where we have to be a little bit careful at this point, right? Because of hallucinations and other ⁓ you know other issues that we could see within ⁓ AI use in in patient care. So we've got to be careful in how this how this occurs. But, you know, I I think it's exciting, especially to see the the volume of the work that you did and to to think about, you know, as I'm as I'm charting and I start to enter a ⁓ a patient complaint into my my ⁓ ESO EPCR, you know, then how can that assist in perhaps some cues or suggestions based upon what I've put in there? So that's mixing my gestalt with what the with what the patient reports and then kinda helps to get to those time critical ⁓ interventions a little bit faster. Wonder what you ⁓ what both of you think of that. speaker-2: Yeah, the the implications of this work really are are various and broad. anywhere from improvement in the charting, because we all know as as field providers and clinicians, you spend much of your time not taking care of people, but just charting, and that's frustrating for everybody. And it's sometimes, you know, not just the time of it, but the accuracy is a problem. ⁓ you have to recall what happened and when it happened. ⁓ the details of you know some of the the drugs or the dosings, you know, may escape you. and having tools to help you in the charting, you know, these are things that, you know, if you have a if you have a prediction, maybe the the chart downstream can follow the expected pathway. And if it detects an error, it can warn you of that and then you can go back and correct it. ⁓ a lot of the thing that we're interested in addition to that is in the real time. So we're using retrospective data, which always has limitations, but it's what we have right now and a lot of it. We're using that data to learn what happened these millions of times over and over again. And are those patterns recognizable? Can we apply them to the future? And then in real time, can we assist the medics or the nurses in making or the doctors in making their decisions when the patient's in front of them? And the societal tolerance for that will take time, as we all experience with with AI. But ⁓ having a human in the loop won't go away anytime soon. We we dream in our department of autonomous systems of care and we are confident that they are gonna happen and and we're proud to you know to have some participation in that. But having a human in the loop is ⁓ is appropriate, it's responsible, and the right way. So you could get decision support and then decide whether to take it or throw it away. ⁓ and And we practice that ⁓ right now as we're evaluating these things. speaker-1: Yeah. And just and and you know, to to sort of say how we are trying to start to turn that dream into a reality, you know, one ongoing project that we're working on that that's sort of started after we did some of this machine learning modeling work, ⁓ is we're s starting to kind of work out the details of how we would ⁓ apply this research in the field. So with ⁓ ⁓ in that c in this case, flight paramedics. So our our colleague Frank Ayet is the medical director for Stat Medivac, an air medical transport service in the in the in this region. And so with ⁓ with flight crew paramedics and flight nurses from that organization, we are doing some pilot work ⁓ where we're testing out ⁓ recording audio ⁓ as they engage in their patient care. and you know, transcribing that audio and summarizing it in meaningful ways ⁓ you know, on a on a on a computing device. And then, you know, we're gonna be building out ⁓ an ability to convey an alert, like a smartphone alert, for instance, as to like a that a patient might need a c critical care intervention. So we're out in the field, ⁓ in this case with the air medical crew starting to pilot And think about how this type of research could be applied prospectively, you know, where you're collecting audio and then making ⁓ clinical decision support recommendations back to those humans who are not only in the loop but ultimately making the call, because they're the ones who who have patients' lives in their hands. speaker-0: Yeah, and I think this is gr obviously ⁓ we've had ⁓ conversations prior to this, so I think this is awesome what you're doing for the for the EMS providers and you know, getting all of this information and starting to do this great work. I wanted to drill down a little bit more on the documentation side of it because that's an area that gets a little bit of backlash, right, on social media when you start talking about ⁓ AI, you know, AI generated or AI assisted ⁓ narratives and things like that. and you know, where that can, you know, potentially weaken the the patient care provider. And I I'm not saying I agree with that stance, but it is a it it is a it's a valid point of opposition to it. And I wonder what you both ⁓ think of that. And then I'll kinda ⁓ share some of share some of my thoughts as well. speaker-2: It's interesting, you know, and I I'm a clinician and Aaron is a statistician plus. ⁓ he's beyond statistician, but ⁓ you could probably look at it in in practical ways, which maybe I, you know, see and then Aaron has objective ways to maybe categorize some of these things, which is important to put together. We're already used to pulling in nursing notes into our notes. Like let's just say I'm working in the hospital. ⁓ the nurses write their report, they take vital signs, and in our in our records, they pull that in, and that way you don't have to type them. So, in a way, that's an automated process. It's pretty discrete factual data, and you can review it pretty quickly. But we do those things. We also have little phrases or templates. This is everywhere. You know, your physical exam pops up. You change the abnormal things. And also, like you said, Mike, I'm not saying good or bad. I this is just kind of the state of the art here. so automation is in our lives always, and we're used to it. It helps our jobs go a little more efficiently and faster. But you have to be careful because you don't want it to make you lazy or forget something or skip something. And That's where a lot of the danger is. Even when you have templates, we see folks put their template in and it says a normal heart exam when it was abnormal and you didn't change that part. So those are those are important human factors that we have to use these tools responsibly. And then once you come up to some of the AI tools that help chart for you, that could be, you know, using the same language processing and ⁓ you have to be careful of the assumptions that are made or the mishearing that the microphone may have picked up. And the responsible job is you review the work that comes before you and then you get into who authored this. Is it really the the voice and the opinion of the provider, which it needs to be in the end. So the efficiencies are great. And I think, you know, in in our s scheme, We want to help workflow, but we're also thinking if you have twenty soldiers down on the battlefield and two medics, there's no way those two medics are going to feasibly ⁓ save all twenty soldiers. There's not the resource, there's not the time. So if we could find a responsible way to make that efficient ⁓ practice pattern where we're talking about the minutes and the seconds that matter. to triage someone. and it's not just as a scapegoat to get your work done. You know, we're we're trying to scale the bites that we're taking in progress. and and I think if you do that to save lives faster, you know, maybe people would be more satisfied with that. But we we heed and appreciate all the warnings and and we try to make them, you know, ⁓ appropriate use. speaker-1: And one distinction I think that's important to make, you know, in terms of heeding the warnings, and Mike, you mentioned backlash, is between the clinical and the operational tasks. So I mentioned we're doing you know, field development of how to implement this work in real time. So a lot of it is we do qualitative interviews with, in this case, flight paramedics and flight nurses. And so the distinction we hear from them ⁓ between clinical and operational. On the clinical side, we do get a lot of, you know, this sense of, you know, we need a human in the loop. Like I, the flight medic or paramedic need to still be, you know, making the final call, like I'm the expert. But on the operational side, the the sentiment is if you can get the AI to like fill out my chart to plop some time stamps on the interventions and medications that I give, like two thumbs up, you know, like like save me the work. There's no push back in terms of like, I need to be the human in my loop, in the loop on my own charting. You know, it's like get take care of that. Let me, the human, be the clinician. That's the sentiment we hear. And so I think that's that distinction is really important to keep in mind when we're developing tools like this. speaker-2: If if you're driving your vehicle on the highway, you know, your check engine light comes on, y your oil pressure. I mean, these are these are signals, they're alarms that take a lot of different data from the the vehicle's onboard computers now to give the driver just some idea. And when you're busy driving, you don't want that to be a distraction, but you want it to, you know, give you meaningful information so you don't take the eyes off the road. You can speak And listen without you know, keeping your hand on the wheel, all these tools. And we know that clinicians are so busy and sometimes ⁓ they forget things or miss some of the signs that are coming to be. And you know, if if the computer can aid in saying, Hey, beware this patient's trends have been this pattern, we've seen this before. You know, they might stop breathing at some point or they might drop their blood pressure. You know, we we want to take every signal and every word that's available to us to just give some of those warnings so the medic can in an informed manner again decide to take it or leave it, ⁓ just like you would your oil pressure. Just a little longer. I'll keep going, but you know. speaker-0: Yeah, and so that's you're the first person that I've ever heard use that analogy for ⁓ explaining, you know, the the use of AI or the use of of technology and decision support. I think that that's a that's really a perfect way to look at it. ⁓ you know, and I I I agree completely. ⁓ and you know, a caveat to that is that this is not new, right? ⁓ for for many years that I've been going to my ⁓ my my yearly visit at my GP. I I go in, I sit down with the physician, and they're going to explain to me at the beginning of that office visit that they're using AI to record the visit and to take care of a lot of their dictation and ⁓ and record keeping. So this is this is not new. and you know, I think that having these conversations is great for the field personnel. to hear so that they see the level of care and concern that you're taking with developing these systems and the the the significant steps that you're taking to prevent errors from being made as we start to explore how we can better utilize these these systems. So I I applaud you and the work that you're doing and I'm ⁓ you know I'm I'm r as I said really happy that we're having this conversation. I hope that it can start to bring some some providers a little bit more comfort as we're as we're exploring this. ⁓ so unless you guys had anything else to talk about on that, I'd like to ⁓ kind of talk about some of the work that you're doing right now and what's on the what's what's at that next horizon for for the work that that you've done already. speaker-1: So I'll give a little intro and Lenny jump in, provide context. So the so you know the the nature of our working relationship is, you know, I come like groveling to Lenny with grade school level EMS questions. And so, you know, when we we did some of this work with n language modeling and we wanted to apply it to you know to one of our our data from our local EMS agency here in Pittsburgh. And so You know, we got to talking and and as as I mentioned and Mike as you mentioned in our last work we looked at critical care interventions as the outcome we were predicting. But you know, when we started talking, you know, Lenny, you sort of made the point that dispatch ⁓ is a a very relevant outcome. Do you wanna say a little more about that and how we arrived at, you know, dispatch outcomes? speaker-2: Yeah, you know, when when you call nine one one, there's a lot of layers that are a challenge in the United States and great variability in different EMS agencies, whether you're urban or rural, who's available. So all of our EMS colleagues and nine one one telecommunications professionals too will appreciate this. You know, it it it just depends on where you're at to know how many resources you have, how busy you are, etcetera. But what we know is you get a phone call, people are upset, it's the worst day of their life, everyone's stressed out, the telecommunicator has to kind of feed through that information to get the the location, the instructions, the right resources to dispatch. And part of the challenge too is there's a library of different determinant codes that you have to punch in. To then release who's gonna respond. And computers help with that. It used to be flip cards, ⁓ but now we have computers and that's great. And the telecommunicators are pro are trained to work through these algorithms. We're curious if we could shave minutes. Again, you make a 911 call, someone's drowning, someone is burning, they have ⁓ cardiac arrest. ⁓ Every second matters and we need more ADs out there, we need more CPR, we know that. But the fact is, you have to generate a 911 call, they have to get the information, find the location, dispatch. We're talking two to four minutes already. The brain dies in four to six minutes. ⁓ then the the EMS crew has to get in their vehicle, find their way and drive four, twelve. Fourteen minutes, it's tough in the United States. In some of the rural counties, forty five minutes to get help. So we just need to get the ball rolling fast. And we could go all day about what tools we can then plug into these, but on the dispatch side, can we start as the call is happening, using words, using information to start churning wheels and even help telecommunicators, dispatchers? and EMS providers know what they need now ahead of the time that we would do it to date. ⁓ there's a lot of opportunity and we're we're pleased to be able to partner with the public safety professionals in creating these things and our talented ⁓ ML experts. speaker-0: Yeah, I mean I I don't think there's any I I don't think there's any way to argue that the faster you mobilize the needed resources, the better that outcome is is likely to be. So ⁓ yeah. Sorry, Doctor Weedman. I d go ahead. speaker-1: No, no problem. I I I was just gonna ⁓ say that of you know, of all the moving parts to dispatch and all the decisions that are being made that that Lenny was just articulating, in in the work that we're doing now, we tried to simplify and hone in on one critical distinction, and that's whether ⁓ an advanced life support ALS or basic life support BLS unit is dispatched to to Treat a patient. And one reason why we landed on that is there are some really striking statistics out there about what a high percent of 911 calls ⁓ are from a patient in a situation where they don't need any critical medical intervention. You know, all all sort there are all sorts of reasons why folks call 911 that are very important and and legitimate and But but they don't require critical care medical intervention. And so the research question that we started with was just can we use this AI and natural language-based modeling to get some more insight into whether a a basic life support, you know, a lower resource BLS unit is ⁓ dispatched to treat a patient. And so in that, there were kind of two sub questions. One sub question was similar to our last work where we were predicting receipt of a critical care intervention. Here, can we just predict whether a BLS unit was dispatched to treat a patient? and then the second question was can we, in the cases where a ⁓ the unit that was dispatched is actually different from the type of unit that was recommended by the agency protocol. Can the model actually provide some insight into what the protocol recommendation would have been, if that makes sense. speaker-0: Yeah, totally makes sense. And you know, the I think that's where ⁓ you know that's where we can really start to impact some things right away, right? ⁓ you know, because the this this initial paper looked at patient complaints and the the primary chief complaint entered by the by the provider. So someone had to obviously already be there to collect that. and and enter it in. So if we can start looking at this data and make real predictions that are that are grounded in evidence for what type of what type of resource is likely to be needed and mobilizing that resource more quickly, yeah, I think that that stands to really be where ⁓ where where some big impact can be made. speaker-1: Yeah, and so so you know, along that impact line, ⁓ when we were choosing our data source to use, the language data source, ⁓ we moved out of the chart, like the chief complaint that gets put in the chart, and for this work we looked at the the notes that the 911 dispatcher writes down. So again, free text longer than chief complaints, you know, they were ⁓ a few sentences on average. ⁓ for patients who had 911 call notes. But we we wanted to move, you know, again a little closer to a prospective type design, even though this was data that was already collected where we're looking at 911 call notes that are jotted down in advance of the dispatch decision of whether to send ALS or BLS. And you know, Lenny, you can say more about these cohorts ⁓ but we use we used two cohorts from Pittsburgh EMS. They both had about twenty to twenty five thousand patients. And I don't know if you want to speak to kind of the agency and the types of calls we s we see. speaker-2: You know, we have ⁓ so for example we have fourteen to sixteen ALS transport units up at any given time, plus some rescue trucks and supervisor vehicles that are circulating. Those are paramedic units, plus we have anywhere from four or five BLS transport units. But Again, if you look across the country, so much variability and getting a paramedic these days is hard to come by. ⁓ it is a field that is a challenge to maintain. ⁓ school's expensive, it's a lot of hours to go through, and and having a longevity in this field where it's exhausting and the hours are tough, trying to get folks to commit to that long term. especially looking at other allied health opportunities or other school opportunities that may they might want to pursue, we feel the challenges in staffing across the country. And that translates into many locations. How many paramedics can you actually put up on the shift today? ⁓ do we have to run two BLS trucks instead or and and knowing what resource you can deploy ⁓ you know, fortunately we've been we've been strong in in Pittsburgh to date and you know we've got good staffing, but we've seen our roller coasters. and taking these cohorts of all of our calls, ⁓ looking through the the chart and the dispatch data, and knowing what our administration decided, all of these 911 determinants should be an ALS unit, all of these determinants should be a BLS unit. All of these should get two units to help each other. You know, we go through that exercise, but then we see what actually happened and we want to compare those differences, protocol versus reality. And then the way the dispatch happens sometimes influences things. So ⁓ nothing turns out perfectly. But why why wait to analyze these things? We'll We've waited years because we have all the data to analyze it, but if you're an administrator or a person in real time, you shouldn't have to wait. And we want to give you some side info to say, hey, this call sounds like it actually might need a paramedic, even though it sounds not that bad. Or it sounds like the way they're describing this, most of these turn out to actually be okay. You could probably save that resource. And send a more basic unit and slow down your approach. That would be very helpful for EMS agencies nationwide. speaker-0: ⁓ yeah. With without a doubt. And anytime you're dealing with a finite resource, if you can better ⁓ you know, if you can better ⁓ you know have have a better response then that's going to that's gonna pay dividends in the end. speaker-1: Yeah, it ⁓ and just to give kind of a concrete example of one thing we've found in the work so far that speaks to what Lenny was just saying, and this is maybe my favorite finding ⁓ b because of the potential implications. So imagine two groups of calls. One group of calls are patients ⁓ who receive ALS. ALS is actually dispatched. And the agency protocol recommends that ALS was dispatched based on the severity of illness or injury that was reported. For those cases, our models, based on those 911 call notes, our machine learning models, they made really strong recommendations that that indeed ALS is the right unit to dispatch. BLS would not be a good idea. Now imagine another group of patients. that also received ALS. ALS was actually dispatched. However, the nature and/or lower severity of these patients' condition was such that protocol would have recommended BLS. For those for those patients, our models actually were more in line with protocol. Our models actually intuited the the lower degree of severity and Swung back toward recommending BLS dispatch to a large degree. And so again, tying it back to what Lenny was saying, you could imagine when time is critical, that recommend the model-based recommender system starting to fire up and maybe leading the dispatch in the direction of BLS because of the protocol recommendation, even if ALS might have been the default assumption. speaker-0: No, that's amazing work and ⁓ excited to ⁓ read it. When do you think ⁓ when do you think that paper will get published? speaker-2: It's it's in good shape. We have some tune ups and you know speaker-0: Gotcha. speaker-1: I would I would say we're hoping to submit it, you know, in the next month or so, early fall. speaker-0: Gotcha. Nothing I'm gonna hold you to. I was just curious. Really, really excited to see the data. so we'll kinda round this out and I'll I'll pop it back to you both of you for any ⁓ any closing thoughts. But ⁓ yeah, I just wanna thank you for your time today and the time that you put in every day to to advancing this this profession. And as I said before, hopefully this can ⁓ if they're if ⁓ for the skeptics out there, if you can start to see the the care, consideration and concern that's being put into to developing ⁓ these systems, hopefully that can can bring a little bit of ease. But I'll let both of you ⁓ kind of close it out. speaker-2: We are so thrilled with the the collaboration is is how this happens. We intimately work with each other in our department and that has many layers of technicians and data scientists, clinicians. but our partners are those working in the field too. And you c you can't ⁓ make this happen without that that rich Collaboration. We work with our data colleagues at the 911 Center and our administration there and our EMS colleagues and the actual clinicians that work in the field so we can share our thoughts and how the realities are. So we're sensitive to tailoring these tools to what is helpful, what isn't helpful, what's feasible and not feasible. And cost, like you can't release a gizmo that's gonna cost an exorbitant amount of money. And we don't want to distract them away from the the bottom line, which is paying attention to your patient. So w we are sensitive to that. And we also want to have accurate information that that we provide. And we're we're really meticulous about pouring through our records, but we feel lucky to have access to this. And what's so important is that. You know, we save every bit of information to try to build robust models and predicting when folks are in danger. And then we would just we'll find ways to communicate that to paramedics in the field, EMTs in the field. And ⁓ whether they act on it is is likely gonna be their final decision. We're not here to substitute a profession and we we don't want to do that. We just want to enhance things that the five senses, however many senses we have, might not pick up in the heat of a moment, or when there's so many victims in an MCI, let's say, and and we think there's a lot of ⁓ help that could be done. speaker-1: Totally. And I I think to sort of add a slightly humorous, you know, take to it, I I think if when if you ever hear someone talking about AI and how it's gonna replace all of our jobs, they probably don't actually work with AI in in the real world, you know? And so just to echo what Lenny said, like co-development is a term that I think gets thrown around a lot and for good reason. Like without People like Lenny and and yourself, Mike, I know you're you know a paramedic. There there is no data. I probably don't even have a job. ⁓ so you know, we we ⁓ we do this only to return the favor, ⁓ and hopefully to the patients as well. speaker-0: Yeah, I agree. I ⁓ so there's a ⁓ there's an interesting phrase, I'm sure you both have heard it, but for anybody and I'm not gonna try to own it, but ⁓ AI is not gonna take your job, but somebody using AI will. speaker-2: That's true. Yeah. You know, if and if you're working in the field, we know when we get overwhelmed. Like I admit I am often overwhelmed at work. Whether I'm in the hospital or in an ambulance or in a helicopter, it's hard. And we we know that if you're one or two paramedic partners with one patient that's really bad, we can help you. And we can help you. assign your work order for what priorities have to happen. Or if you're a medic with multiple victims, we can help you there. But I get overwhelmed all the time and I know our colleagues do. So we we feel what it's like and where we just need to admit, hey, this is beyond our our human limitation. speaker-0: Yeah, absolutely. Well, I I think that's a that's a great way to end it. And ⁓ you know, it it's it's so amazing to see both of your passions for everything that that you're doing and the work that you're doing and hopefully everyone else has appreciated seeing that. And again, thank you. Thank you for all you do and I hope you have a great day. speaker-1: Same to you, Mike. Thank you. speaker-2: Thanks.