Understanding Nursing Informatics and Large Language Models in Healthcare

Tamara Macieira (00:00.77)

Tamara Macieira.

Heather Herdman (00:04.894)

Hello and welcome to the NANDA cast. I'm your host, Dr. Heather Herdman, Chief Executive Officer of NANDA International. And today I have the personal pleasure of speaking with Dr. Tamara Massera.

I met you when you were an undergrad in Brazil. And I remember having a conversation with your advisor at the time, Dr. Tana Schianca. And we both felt you were something really special. And we agreed we had to keep an eye on you in the future and kind of see where your curiosity would take you. And I don't know that I would have ever thought we'd be sitting here, this many years later having this conversation, but kind of funny how life is. And I'm really glad that it brought you.

Full circle back to having a conversation with me. So for those who don't know you, Dr. Masciera finished her bachelor of science in nursing at the Federal University of Minas Gerais in Brazil. And then she received her doctorate and a postdoctoral fellowship at the University of Florida. She is an expert in nursing informatics, standardized nursing terminologies, and data science, including the application of machine learning and large language model techniques.

She's applied these methods specifically to examine the impact of palliative nursing care on hospitalized patient outcomes for cognitively impaired older adults using the analysis of nursing data retrieved from electronic health records from diverse health settings. She's been dedicated to nursing and informatics research focused on populations with life-limiting chronic conditions and the development of technology to improve patient care.

Dr. Maciere's program of research is funded by the United States National Institutes of Health, National Institute on Aging, and focuses on the implementation for the first time, that's pretty exciting, of the research data infrastructure that will generate and integrate standardized nursing data for over two million adult patients in the statewide One Florida Plus clinical research network. So with all that,

Heather Herdman (02:10.27)

Tamarara, welcome to Nandacast.

Tamara Macieira (02:13.036)

Thank you so much, Heather. And I just want to say how much I love the name of this podcast. NandaCast, really is very catchy. So I'm really happy to be here.

Heather Herdman (02:21.332)

Good! Well, we're happy to have you. So let's start kind of with something that's probably pretty basic for you, but maybe not so much for all of us.

Let's talk about what a large language model actually is. You we hear a lot about this in the literature and certainly in conferences over the past few years. But for those of us who aren't informaticists, can you kind of do a dummies guide to large language models type of explanation?

Tamara Macieira (02:52.578)

Yes, absolutely. I can definitely try to do that. And this question is very timing because just right now I'm revising a paper that I submitted to a nursing journal, described a little bit the work that we've done with large language models. And the paper was reviewed by different reviewers. And the main comment was, you need to explain this better. Like people don't know what these things are. So just by mentioning it, it doesn't make sense we get lost. And I'm like, that's

That makes a lot of sense because I didn't really understand at first when I started researching it and using it. But if I were to try to explain simply what it is, and I don't fully understand everything that goes behind it, I'm not a computer scientist, but the way that I like to picture a large language model is that basically it is a system, an AI, an algorithm, or whatever you want to call it.

that predicts what the next word or couple of words will be in a sentence. And that's basically what it does. it's almost like it's an autocomplete of steroids. You can think it that way. Because it is basically that. So the difference of large language models to the autocomplete that we have on our phones when we're trying to text something and suggest the next word that we should use is that

Heather Herdman (04:00.36)

So like we have on our phones?

Heather Herdman (04:06.819)

Hahaha

Tamara Macieira (04:20.962)

Those models have been heavily trained on everything that is out there in the internet. So basically imagine someone being able to read all of the books that have been published over the years that are available in the internet, all of the papers, all blogs that were written, anything on Twitter. anything or X, anything that is, it's in the internet that was publicly available. These large language models have been trained on. So basically they read it all.

Heather Herdman (04:40.645)

Yeah

Tamara Macieira (04:50.848)

and they read, they memorize, and what they learn how to do or what they were programmed to do is identify patterns within all of that information. So what a pattern would be is that let's say that for 90 % of all the text that is out there, 90 % of the time the words patient has is followed by no pain. So patient has no pain. Let's say that that's what we see, especially when we think of nursing.

and health care. So I say that that's what you find in the most of the literature or what has been documented for patients. The large language model learns to recognize that every time you have the couple of words patient has, it's followed by no pain with 90 % of the time. So based on the probability of that happening, that is the suggestion that the large language model is going to do to you or make for you if you have

a prompt or a question where you ask what are the most common signs and symptoms that a patient might have or something like that. It will bring to you the answer that it sees most often in everything that it has learned. Does that make sense Heather so far?

Heather Herdman (06:03.828)

It makes a lot of sense, but it also makes me think, well, so how do we know that what it's pulling is accurate?

Tamara Macieira (06:16.172)

Well, we know that by employing our own knowledge, right? Because the large language models, they don't know things. So they don't know what the meaning of patient has no pain is. They don't know, well, it doesn't know because it's not even a person, right? The model doesn't know why it's suggesting that. It just know that that's the pattern. So we, with our own knowledge or experience or expertise and what we have studied and what we know about that phenomenon.

we are in charge of determining if that information is accurate or not. So if you would think of the application, let's say that we are nurses and we work at a place that is saying that they're going to use large language model for us to help us in some way in the documentation. The large language model might make suggestions for you of specific

things that should be documented or something like that for your patient based on all of the information that has been documented for hundreds and thousands of patients in the past. But that does not mean that that is going to be the reality of your patient. So we as humans in this loop of using large language models, we still have the role of being the ones validating, verifying that what the model is suggesting to us is actually

the reality or the truth of the phenomenon that we're seeing.

Heather Herdman (07:45.082)

so it's a tool.

Tamara Macieira (07:47.104)

It is a tool. Yes, it is a tool created to assist humans in tasks such as summarization, write ideas. Like if you want an idea for something, it will give you ideas based on the hundreds and thousands of other ideas that people have put out there. It's able to help you improve a paragraph that you have written.

Heather Herdman (07:50.354)

Thank

Tamara Macieira (08:13.262)

or give you bullet points for a specific topic that you want to talk about. But all of those suggestions or the output, the answers that the model is giving you, it's based on what has been seen in the literature, seen in the text and all of that, but really nothing other than that. So the model can't even assess if that is really pertinent to you or not. You need to be able to do that on our end. So our tools to assist us, but not to replace us.

or anything like that, at least not currently. I would not say the future, but yes.

Heather Herdman (08:44.244)

Okay. Probably a good thing. All right. So now that we kind of have an operational definition, tell me what it is about these things that intrigues you so much and why are you using them and how do you think they can help us in practice?

Tamara Macieira (09:03.416)

Yes, those are great questions. So what intrigues me the most about using this large language models or other type of machine learning or artificial intelligence models is the potential that it has to decrease the cognitive load that we have currently on different tasks that we're doing. So really the ability that it has to help us perform specific tasks.

and free up our time to focus on other things that only we can do and we need to dedicate more time to. So let me give an example when we think of healthcare. So I work with a lot of data that nurses generate, specifically care plans, which are my love and passion since I met Heather years ago and working with Tanya throughout my undergrad. So I've always recognized that care plans really show what the...

Heather Herdman (09:46.484)

Hmm.

Tamara Macieira (10:00.056)

The critical thinking of nurses are what the bulk of the care is and really what was the plan that the nurses had in place when they're providing care to a patient. So what happens nowadays with documentation in electronic health records or electronic medical records that some people know them about is that

There is a burden of the type of information we need to input, right? So nurses are required to complete these enormous tasks and click and click and document. And then depending if you're an ICU nurse, you are expected to round on your patients every hour and write down all this information, put all of this information on the computer. So a lot of that really takes up the nurse's time, like 30 % of the time, nurses nowadays.

have within their 12 hour shifts is in front of computers for about three hours, which is awful, right? It is. And it's a lot of time, but the amount of time that nurses are spending does not reflect in good data or good information that is going to the EHR. So nurses spend all this time and it's still when you look at the data and you're trying to understand what the nursing care provider was, it's not an accurate representation.

Heather Herdman (10:50.866)

Wow, that's ridiculous.

Tamara Macieira (11:15.202)

Because the systems are not designed to help us, right? They're not designed to really help with our thought process or documentation. So nurses are scrambling to just enter something because they have to, and that doesn't really show the true care that they provide, because we know they're providing care that is improving the patients. Otherwise, the patients wouldn't be discharged, wouldn't go home, but that is not going to the EHR, to the electronic health record. And that's one of the ways that I think large language models can help us. What they can do when

Heather Herdman (11:17.876)

Right.

Heather Herdman (11:32.977)

Exactly.

Tamara Macieira (11:45.506)

being used with other AI tools and other ways of capturing information is that they can help making suggestions of care plans or of data points that nurses could enter in the EHR. Well, that could be entered in the EHR after the nurse is validated. So with some of the work that we have been doing experimenting large language models for nursing documentation specifically with the care plans is that we've learned that when you give

the model, a narrative, giving an overview of what your patient condition is, almost like if you were giving a handoff to another nurse, but you're that handoff to the model, for example. The model recognizes some of the patterns in what you're saying, goes back to its alcoholic memory to see, for all of the patients that have these different things, these were the nursing diagnosis suggested or something like that. It brings back suggestions of care plans.

And those are very complete suggestions, depending on how you write your guiding question to the model. But it actually brings to the nurse important aspects of the care that the nurses are knowing that they're doing, but they never document. So some of the nursing diagnosis that we do a lot with our patients, we recognize like family issues or wellbeing issues with our patients or other psychosocial problems that we've identified. We talk to our patients about, we treat those.

Heather Herdman (13:02.515)

you

Tamara Macieira (13:13.41)

They don't go into the record because we don't remember what that Nanda is for that term, right? We don't know all of the Nanda's by heart. So we don't know what it is. And then I don't have time to type down the list and trying to find. So it doesn't get documented because the first terms in the list that I have to go through are those obvious acute pain, risk for falls, and all of that, and all of the patients have.

But the model can bring those to your attention right away. And then with one click or one command, you can get those things documented, reflecting what the reality is for the care that you're providing. But another way that I like to use large language models is that it came out of a large necessity in our project, where we are, as part of this project, to bring nursing data.

to this clinical repository, the one for the plus, one of the tasks that we had to do was to map local terms that nurses use in electronic health records to document care plans. Unfortunately, the hospitals are getting data from them. They don't use any terminology, not NAND, NIC, or NAC or anything else. And that could be a whole new podcast on that. so they use their own curated list of terms they put together.

Heather Herdman (14:25.204)

you

Tamara Macieira (14:34.222)

And the different hospitals use different lists of terms. So if I were to look at the data or the information about a patient in nursing care coming from a single hospital, I could look at the way that it is. I can look at the list of terms that nurses use. I could try to understand that, and that would be the end of it. But when you try to get data or information from more than one hospital health system that has wildly different lists of terms,

You need to have a common language in place, right? Because you need to be able to see, well, the computer needs to be able to recognize a new as well when a patient has acute pain. And that then you should be represented in the same way, using the same words, using the same concept. So that's where the terminology is coming to place. So as part of our work, we had to map each one of those local terms to a nursing diagnosis from Nanda.

a nursing outcome from NOC and a nurse intervention from NIC. And our list of local terms, they were in the thousands. So there were like 2000 interventions that we had to map. And it's not like the nurses are doing 2000 different things, but when you don't have a terminology in place, you have the tendency or this health system has the tendency of creating variations of the same intervention.

Heather Herdman (15:41.456)

Heather Herdman (15:47.538)

Right.

Heather Herdman (15:57.031)

Yeah.

Tamara Macieira (15:57.432)

So the same trivance documented in different ways. So we had to map all of that. In the first two years of our project, the large language models were not publicly available yet. So we were like, OK, we're going to have to do this manually. We did try some natural language processing that was available at the time. And there were software such as MetaMap, which is actually out of the National Library of Medicine. And there's also one that's called Clamp.

which are natural language processing systems, which in itself is a type of large language models, but they weren't as advanced as what we have now. And when we tested those out, they didn't really work for us because they were built and trained on medical data, not nursing data. And those are very different. So the phenomenon and the things that nurses document, they didn't have any concepts that could be mapped to those.

Heather Herdman (16:49.289)

Thank

Tamara Macieira (16:56.182)

And also you could try to put the terminology there within their system to be used, but it didn't really work out for us. So we're like, okay, we have this milestone we need to meet. Let's just manually map these terms. And it took us a year. It took us a couple of years to map. And this is like weekly work that we're trying to do because the terms are so different. First of all, you need more than one person doing that task because I have my own bias and my own interpretation of.

Heather Herdman (17:10.706)

I can't imagine.

Heather Herdman (17:22.836)

you

Tamara Macieira (17:25.814)

term, which might be very different from my colleague. It might be really different from how the nurses are actually using it. So you need a team to do that work. So it's very costly and it takes a lot of time as it took us a lot of time. But then the large language models came into place and we were like, well, can they just do this for us? Could chat GPT just map all of the terms and we wouldn't have to do any of that. So we started doing some experimentation with that.

not only the GPT models, which are large language models, but also combined with other AI methods, such as text embedding, some other different kinds of AI to see if that would work for us. And what our results showed, which is in this paper that I'm revising, is that the AI is great in suggesting to you what the main, let's say NAND does.

for a specific local problem are. So these are, it can suggest how many you want. We experimented with the AI giving us the top five NANDs that are more closely related to this local problem or the top 10 or the top 15. But then when you give the GPT model or the large language model, the prompt or the question to choose the best one out of those five or 10, it doesn't perform that well.

Heather Herdman (18:52.702)

Yeah.

Tamara Macieira (18:53.062)

And it doesn't perform that well that it can just be employed by itself. And it's not necessarily the fault of the model, but it's the nature of the task in itself. It's very subjective in some way because you could have different interpretations for a local problem, especially because those don't have definitions. They don't have anything to them. So I could interpret, there is a term that we had to map. And I think this was a...

a goal actually, which was mapped to a knock that it was called device complication. What is that? Right. And these are terms that nurses are using device complication. like, okay, what is this device? What is this complication? It could be different number of things. Right. And then when you talk with the nurses, they're a part of our team. They're from different units in the med search unit. A device could be a walker in an ICU. The device is usually a pacemaker.

Heather Herdman (19:31.38)

Thank

Tamara Macieira (19:53.388)

in the labor and delivery unit, the device is actually one of the balls that they use during labor. So it's different things. And even with those, are different interpretations. So the model would select different things, which is understandable because of the nature of the task. So we couldn't employ the model, the AI to do this by itself. But what we did is that we used the AI to give us what those top 10

Heather Herdman (19:59.156)

Very different.

Tamara Macieira (20:22.602)

Nandas were not knocks and nicks. And then we would have our team members choose from those. So we'd have three people select trying to select the same Nand or the same knock and so on, independently and then through a consensus process to get to the best match. But just by having the AI suggest the best 10 Nandas, it already decreased the cognitive load and the time that it would take us to map those terms in more than half.

Heather Herdman (20:40.084)

.

Tamara Macieira (20:52.908)

Because before what we have to do, we get the local problem. We look over the 240 something NANDIS and let's see which one would be the best one or could go, let's look at a class first and see. It would take us a lot of time. So it speed up our process. So what took us initially two years to complete with the help of the large language models, we were able to complete that work in six weeks. So yeah, it is.

Heather Herdman (21:06.803)

Mm-hmm.

Heather Herdman (21:17.554)

Wow, that's a substantial difference.

Tamara Macieira (21:22.862)

It is. So it was able to help us in that main task. And now we have mapped over, I think, close to 5,000 terms across the 12 hospitals we work with as part of this pipeline to make the data available.

Heather Herdman (21:39.506)

That's amazing. So two things came up while you were talking for me. One is the environment in which you are, I think you use the word training, the models. You do this in a private server. You're not doing this on the main internet, right? I think that's important to say.

Tamara Macieira (21:49.739)

yes.

Tamara Macieira (21:56.958)

No, yes. Yes. And that's a very important question because you don't want to just put anything within, just give it to the mod or anything like that. Yes. So what we've created is what we called a retrieval argumented generation pipeline or reg for short. And what that pipeline does is that it puts the content off the terminal that you want to use in a local server.

We use another AI method, not GPT or anything like that, to bring back from this local server the terms we're interested in. And this model is only run locally as well in our computers. And then only the top five terms or something like that are given to the large language model to look at. So that way we're ensuring the privacy and protecting intellectual property of the terminologies themselves as well.

And we're not giving anything directly to large language model or open AI or anything like that.

Heather Herdman (22:57.82)

Yeah. So the other thing that strikes me and I've been playing around with some of this language stuff online, just trying to figure out what can I do with this? And it was very quick that I started to realize the questions have to be very precise.

Tamara Macieira (23:10.71)

Hehehehe

Heather Herdman (23:18.8)

So are there guidelines or how did you, do you just keep iterating the question until you get to a question that gives you an answer you're really looking for or are there tricks that you have along the way picked up?

Tamara Macieira (23:32.206)

Yes, that's a great question. So I'm very lucky to be able to work in this project with different students from computer science. They're very smart. I want to give a shout out to a scene. The student that worked with me before that now works for Google with Gemini AI. And now our research assistant is the done with who is graduating this spring as well. So I'm excited to see him go. But they have been doing

Heather Herdman (23:58.11)

Thank

Tamara Macieira (24:01.41)

They both did a wonderful job with what we called employing different prompting techniques to give us what the best question is. So we started with a very basic question. And we kind of like, OK, let's see what the model gives us. So we start with a very basic question that was not yielding the good results we wanted. It was not giving us mapping between the local terms and the terminologies. And then we proceeded to employing different prompting techniques.

which are basically different ways or specific ways for you to write your question. So we've tested out few short learning, chain of thought, and self-refinement as the three main prompting techniques. We wanted to see which one of those would give us the best results. So few short learning is a method where you write your specific question, what you want the model to do.

But then as part of the prompt, you give examples of what is the expected answer. So what we would do is that we would give the specific task. I want you to find the closest contextual matches or the closest terms in similarity. And this is the local term. These are the top 10 I want you to pick from. But these are examples of what those matches should look like.

So you give an example of this local problem, it was mapped to this Nanda. This one was mapped to this other Nanda. And the reason why you give those examples of that is that the model learns from it to recognize the patterns in those not so obvious matches. Because we gave really hard ones for the model to kind of learn. It's like you're not only looking for the local problem and the Nanda having the same words within the label or the terms. It goes way beyond that.

Heather Herdman (25:42.482)

Interesting.

Heather Herdman (25:52.851)

Thank

Tamara Macieira (25:54.606)

So you of give those instructions for the model to learn and kind of to reinforce what the behavior should be. With the self-refinement, and we're working on another paper on that, need to get the papers out. Like, oh my God. By the way, the other one with the self-refinement is that you give what your main question is, what you want the model to do, but then you have the model double check what their output is.

So it's like the model is going to give you an output, but then within the prompt, you're telling the model, OK, look again at that output. Double check if that's the best one or not. And then check another time before you give the answer. So that makes the model run through that task at least three times before giving you an output. And then with the chain of thought, it's kind of similar to that. You give more detail, and you break down the instructions. And you give as many detail as possible to the model for it to truly follow that.

specific behavior that you're looking for. So that's basically just adjusting what your question is in different ways. You don't even need to have, be a programmer to be able to use that or employ those techniques. It's just different ways of refining your questions and giving, making sure that the instructions are clear. And each one of those ways has a name. guess people like naming things. that's one of the ways that we...

Heather Herdman (27:15.42)

Hmm

Tamara Macieira (27:21.73)

We try to write down our questions, but it takes multiple rounds of iterations. It does.

Heather Herdman (27:26.312)

I'm sure. Are you able to see the, does the model give you a printout or review of how it went through its thinking, how it got to the answer?

Tamara Macieira (27:41.804)

That's a great question. I think it can probably give you that. What we have been able to get from the models is that not necessarily a printout of all of the different steps it went through, because a lot of those, it's like a black box, right? So you don't really know exactly what the model is doing. But.

For some of the models we do, so when we're trying to get similarity checks between two different terms, the model gives it a score of similarity between the local term and the standardized term. So you can get what those scores are. So you can see, those scores range from zero to one, with one being the closest one. So basically, you have the same words across the two different things you're looking at.

But it can give us those scores. So we use those scores as well as a way of understanding why these were the top 10 chosen and why this was top one versus top two. So you know it's giving a score. The score is usually based on vector distance, which is a whole nother thing, right? So yeah, basically the computer transforming the words into numbers and the numbers work.

Heather Herdman (28:56.882)

Hahaha!

Tamara Macieira (29:05.87)

almost like it's coordinate. If you were to think of the vector space being a map we see on the GPS, the computer gives numbers to specific words and places them on a map, which is the vector space. And then this local term is placed here. This nanda is placed here. The distance between those two that it's calculated gives you how close, how similar or not those terms are. So we get this course.

Heather Herdman (29:13.044)

you

Tamara Macieira (29:34.882)

from that too. But I won't go any deeper into that because honestly, I don't know any deeper than that. I'm to explain this.

Heather Herdman (29:37.684)

That's fascinating. It really is fascinating to think about. So you've talked a little bit about this, but you recently gave a keynote at the Ascendio Conference in Rotterdam, which was fabulous. And for those who weren't there, tell us a little bit more about your research. And I think it would be really good for people, nurses sometimes think all this documentation that we're doing.

Tamara Macieira (29:55.0)

Thank you.

Heather Herdman (30:06.82)

is it just disappears. And why do I have to do this anyway? Because unless joint commission is coming, who cares, right? Because it's going to go away. So I think if nurses understood the way that their data can actually be used to improve patient care, could be really helpful for them to understand why it's important. So tell us a little bit about your project.

Tamara Macieira (30:09.389)

Peace.

Tamara Macieira (30:32.066)

Yeah, absolutely. So this project, I like to call it my first baby. So I don't have any children, but this project has been my baby for four years now. And because really is.

Heather Herdman (30:42.802)

you

Tamara Macieira (30:48.256)

It has been the biggest thing that I've done so far in my career when it comes to what drives me in research, which is to bring visibility to the nursing profession. So since I was an undergrad during clinicals and all of that, it always bothered me that nurses were doing so much work, but nobody was really recognizing or that we would document all these different things and I didn't really see that people were using it.

or I would come on the next shift and then nobody had looked at what I had written or something like that. So I wanted to kind of to change that. I wanted to be able to use all of that information that we as nurses already required to document, I'll put it that way, to really be able to bring visibility to the work we're doing. And what you can do with data that is entered is that when you analyze it in a larger scale,

you can actually see what are the different things that nurses are doing. So what are the different interventions that nurses have done or what are the unique signs and symptoms or something about the patient that the nurses have recognized and treated or referred to another healthcare provider in a multidisciplinary team. But you can pinpoint those things and you can analyze how much that is impacting a patient.

So you can see if the nurses does interventions A, B, and C, it actually reflects on the patients staying, I don't know, 24 hours less in the hospital, something like that. It's related to the patient being discharged earlier. Or if you see that the nurses recognize that the patients have the nursing diagnosis A, B, and C at admission, and then those get treated and you have these different goals for the care. And then if those goals are met,

that the patient actually does not get back to the hospital, doesn't have an implantary admission. So you can quantitatively measure the contribution of nurses to care. So the work that we do and the care we provide is no longer invisible. It's not like, I know that the nurses are doing something, but I can't really say what is it that they are doing. But with the data, you can. You can absolutely show what nurses are doing.

Heather Herdman (33:02.004)

But I'm not sure what. Yeah. Well, it also sounds like you could begin, especially with a couple million records, you could begin to say, you know what, when you do this intervention D, it doesn't make a difference. So stop doing it. Like why waste our time doing this thing that we all learn to do in school?

Tamara Macieira (33:23.534)

Thank

Heather Herdman (33:29.586)

when it really isn't contributing to the output and that in itself could decrease nursing time.

Tamara Macieira (33:35.534)

Yes, exactly. All these different things from the task, task list or different interventions that nurses are directed they have to do because it meets whatever standard for documentation or care that it has been published years ago and all of that, that really doesn't do anything for the patient. And that makes me think of the work that my current PhD student is doing. His name is Sami and he's doing

great work with missed nursing care. And we were trying to propose that for his dissertation, he work on missed nursing care as it comes from nursing documentation. But also we need to differentiate what missed nursing care is and what missed nursing documentation is, right? But what has shocked me the most so far on what we have found from the literature

Heather Herdman (34:06.132)

Mm-hmm.

Heather Herdman (34:23.666)

Right.

Tamara Macieira (34:32.02)

is that throughout the years we have measured missed nursing care using this static list of different things that nurses are expected to do. They go from walking the patient, they go from turning the patient, which are main things, right? But some of the other things are like brush the patient's teeth and all of that, and expected to do that every shift and everything, which are very important things that nurses do. But truly those

are not necessarily missed nursing care for all of your patients. Because the patient's needs are different and what is expected and what is prioritized depends on what is happening with the patient for that moment. So if you are not able to do one of those tasks throughout that shift, it doesn't mean that it was missed nursing care. It was just not the priority of the care at that moment. So nurses, because we're telling them that these are the different things they need to do at every shift and then they can't do them,

because of the demand and what the priorities for the care are, they feel like they're not fulfilling what nursing care is about. And they're feeling like they're failing their patients in some way. But in reality, the way that I see it, it's not missed nursing care, it's just reprioritizing the care, but not necessarily you're missing the care. we can look at, I think I said all of that to get to the point that.

The way to truly measure what Miss Nursing Care is or what is the care the nurses are providing that is working but that is not working is through data. Data can tell us that now at larger scale. Instead of us just use not just instead of us using these tools and interviewing the nurses, which is great work as well, but using that by itself to really measure what is missing, what is not for the patient. How about look at the data and really see what was it that the nurses have done.

because they might have not had the time to brush their patient's teeth throughout that shift or to, I don't know, do another task for the patient, but that really didn't impact the patient's progress. The patient was still able to progress and get discharged. So truly was not misnursing care in that way. I don't know if that made sense. I feel like I went on a tangent, but.

Heather Herdman (36:44.788)

Yeah. No, it does make sense because I think a lot of times there's a lot of guilt sometimes within nursing that, didn't get to that, whatever that is. And really in the scheme of things, that wasn't that important. It wasn't, I mean, yes, it's good to do. Yes, it's something we want to do on a regular basis, but keeping the breathing going was much more important than whatever that thing was. And yet there seems to be a lot of guilt because I think

Tamara Macieira (36:52.29)

Yes.

Tamara Macieira (37:00.065)

Yes.

Tamara Macieira (37:09.367)

Yes.

Heather Herdman (37:13.864)

For some reason, nurses tend to have this belief that they have to do it all and that they should do it all. When in reality, patients are getting sicker and sicker and it's just not possible sometimes to do every single thing. So we have to be able to prioritize. Does that make sense? So today we're seeing this huge, I think, explosion of interest in nursing about

Tamara Macieira (37:19.283)

Yes. Yes.

Tamara Macieira (37:29.728)

Exactly.

Tamara Macieira (37:40.034)

No.

Heather Herdman (37:40.166)

informatics and where informatics can go and certainly a lot of buzz about artificial intelligence and big data and how we can use these things that most of us don't really understand but we are aware they're there. How do we use those to improve nursing visibility to improve patient care, to support clinical reasoning and definitely to decrease documentation burden which is always a number one complaint of nurses.

So I had a conversation not too long ago with your colleague Dr. Gail Keenan and she mentioned that you had this prediction that you would, nurses, informaticists would be able over time to reduce nursing documentation substantially. And so I'd be interested in hearing what you think is going to happen.

Tamara Macieira (38:27.618)

Yes.

Heather Herdman (38:32.072)

within the informatics world and where you think the field is going to move nursing in the next few years and what things can maybe nurses look forward to in terms of being able to refocus on patients, which is why we're there in the first place and focus less on charts.

Tamara Macieira (38:41.655)

Yeah.

Tamara Macieira (38:49.56)

Yes, absolutely. And I'm glad that you and Gail had that conversation. So this prediction that we have, it actually has something to back it up. It's not just a wild prediction or anything like that. We are working towards making that happen to nurses. And the way that we and our team envisions it and what we are working on is an innovation.

Heather Herdman (39:03.95)

Yeah.

Tamara Macieira (39:18.562)

that will help nurses document everything they need.

Almost like...

We weren't, it's like seamless-less. I think that's the word we use and I have a hard time saying that. But basically documentation is going to become part of everything that you're doing already and all the communications and conversations you have with patients, with family members, and also with other healthcare providers and so on. So the vision that we have, it's really leveraging ambient listening, which I know a lot of people are using already.

which is great that people are moving that direction. But what differs us from what has, what other companies are doing or other people are investing their time on is that we're going to use ambient listening to capture what nurses are saying. And that can be in a conversation with the patient. That can be something that they say to the system itself, but that audio gets converted into text, which is the premises of ambient listening.

Heather Herdman (40:28.82)

you

Tamara Macieira (40:28.93)

But within our innovation, that text gets transformed into all of the different data points that nurses are expected to document. So if I'm your nurse, Heather, and I'm telling you as part of our conversation, you ask me, OK, how am I doing? And then I explain to you how much you have improved.

how I just assess your pain and it's now a seven, but also I took your temperature or somebody else already came and then it's lower. We did all of these interventions for you. All of that, simple conversation that I'll have with you as my patient, which nurses already do nowadays, this is not something they would have to start doing, they already do that, would be captured and transformed into what the care plan is for that, what it was told there. So what is it that the nurses have done, what the care plan is from that, but also,

Heather Herdman (41:08.392)

Great.

Tamara Macieira (41:19.692)

I said all these vital signs, right? So all of that is going to be captured and transformed and sent to the right flow sheets in the EHR. And then your response to what I'm saying might be transformed into a note that I describe what the patient's response was, and that goes into the progress note. So basically, as part of the care that the nurses are doing already at the bedside, all of that will get transformed into documentation that can happen anywhere, anytime, hands-free.

So imagine, I know, I know it does. And that's what one of our team members, Karen Pryola always says. She's like, we need to bring Star Trek into this. People will relate. And I was like, Karen, I think I'm from a different generation. I don't know Star Trek. haven't watched it, but I'll try.

Heather Herdman (41:49.554)

This sounds like Star Trek.

Heather Herdman (42:06.868)

You need to watch it.

Tamara Macieira (42:12.534)

I know, but it's basically what it is. And honestly, we're not far off of it because it's not like we're going to have to invent all of those different things. We already have the pieces. We already have the ambient listening. Like, there are all these products that you use to capture audio and get transformed into text. We do that when we dictate things into our phone, right? So that can be done already. And we have large language models that can help us with processing all of that text.

Heather Herdman (42:24.34)

Mm.

Great.

Heather Herdman (42:34.26)

Right.

Tamara Macieira (42:40.79)

And then we have machine learning that could help us with some predicting coming from that data. We have all these different connectors that can help us connect to the EHR. But really nobody's bringing all of that together in something for nurses, which is what we're doing, because what happens is that usually healthcare innovations go to physicians first because they're the ones that bring money, right? That's how healthcare systems.

Heather Herdman (43:00.372)

you

Tamara Macieira (43:11.224)

seed them. So innovations go to them first and then potentially they try to adapt something that was initially developed for physicianship nurses. It fails because what workflows are very different here is very different. It's not going to work. So our team is, we are nurses developing something for nurses. So I think that's why I'm confident that we'll be successful.

Heather Herdman (43:11.597)

Yeah.

Heather Herdman (43:25.704)

Very.

Tamara Macieira (43:37.664)

after we get millions and millions of dollars to develop it.

Heather Herdman (43:37.78)

I think this is fascinating, right? I mean, it's very exciting because, I mean, it's true, I've been involved in many electronic health record purchase decisions over the years, and I've always been so frustrated that A, the first question is, will it do what finance needs? It's always number one.

Tamara Macieira (43:49.752)

Yes.

Heather Herdman (43:57.426)

And I'm thinking a dollar is a dollar. You don't split it up into parts of a dollar. It's a dollar. It's the simplest thing. Why do we care? You go do your thing. Then it's the physicians, and then maybe somewhere down the line. And generally, what many of the EHR vendors do is they leave it for the hospitals to figure out themselves what they want or to spend a lot of money on consulting. And yet...

Tamara Macieira (44:04.226)

Yes.

Tamara Macieira (44:07.885)

Yeah.

Heather Herdman (44:22.386)

who documents the most, the nurse. So I think this is really exciting. It was interesting when you were, couple of the keynotes at Ascendio, I heard people saying, having conversation afterwards, and one participant said to me, does this worry you as the Nanda person? And I said, what do you mean? And she said, well,

Tamara Macieira (44:24.654)

Yes.

Tamara Macieira (44:29.324)

Exit.

Heather Herdman (44:49.652)

doesn't this kind of mean if we can do all this stuff with AI that standardized languages would go away? And I said, well, that's interesting. I thought it would actually strengthen standardized languages because to me, you have to have something to map to.

So I'd be interested in hearing what you think, how does this technology advance or not standardized nursing language and does the need for standardized languages eventually go away because of technology?

Tamara Macieira (45:24.034)

Well, I think my short answer to that would be no. It would not go away, at least not the way that we have AI now and things like that. But AI and large language models, they're really good to capturing what we're saying, to transform that into text and all of that. But when you think of getting all of that information represented in some way,

Heather Herdman (45:36.264)

Thank

Tamara Macieira (45:52.472)

within the HR, within a database or within a patient's chart, you need to have concepts there. You need to have something that in a few words represents what was done for the patient. And it's something that has been validated and really showcases what the magnitude of that care is. And that's what the terminologies are. So if you only have large language models showing all of the work for you, what is it that it's going to do?

it's going to get all this data into narrative format in the EHR. How is that going to help you? Well, then you're going to need another AI to summarize all of that narrative information and give you an output. What is it? But then the output that you're going to get from the summarization of something from your patient in the words that are being used or the words that are being used to represent pain will be different from how pain is being represented at a different place.

So you need the terminology to be the centerpiece or the language among the languages. So you can look at nursing and nursing care at a larger scale across different health systems represented the same way. And then you can use that to do all of the analysis you want. You can use that to even show the economic value of nursing, which is something we've been trying to do for so long, right? You need to have terminologies the same way you have ICD code to represent medical diagnosis.

Heather Herdman (47:15.348)

Right.

Tamara Macieira (47:21.784)

you need to have terminology represent nursing diagnosis so you can quantify those and attribute an economic or dollar value to those things. So I think large language models are here to help to definitely, I'll say increase our use of terminologies. I think it will come to that because then you don't really even know, don't, me as a nurse, I don't need to know what an end that is.

but the large language model will help me identify what the NANDAs are for my patient. And then because the NANDAs are already well-defined and all of that, I can read it and understand that that truly represents my patient. But I don't need to know that NANDAs have existed. Yeah, I don't need to know anything other than that. So I think AI is really here to argument and use after terminologies. And I think the magnitude of data that we're now

Heather Herdman (47:58.932)

Yeah, 248 of them.

Tamara Macieira (48:15.852)

Well, not now, but with that we are having access to in the last five or 10 years. But now that we're getting that access to the nursing piece will help us even more to use the terminologies because there is a real need for that. We need to have terminology in place. You look at data from different health systems. Otherwise, it just doesn't make sense.

Heather Herdman (48:37.076)

that makes sense to me. So I know you mentor students now, you have graduate students that you work with. As we're starting to see these newer people in the profession with this interest in informatics, how are you trying to steer them? What are the things that you're trying to get them to consider as being important for their curriculum?

Tamara Macieira (48:38.527)

Hahaha

Tamara Macieira (49:00.878)

That's a great question. So I start with them while they are undergrads. what I do is that, so I teach a couple of courses for the undergraduate nursing students here at the University of Florida. And I try to transmit throughout my lectures and all of that, the passion that I have for research and how much informatics is helping, even if they don't know, in all the different things they're doing.

and all the different things that nurses do as part of their clinical practice. So because of that, and I won't say that it's brainwashing, but because I strongly explain and try to have them understand how much informatics and data science is part of their day-to-day lives, some of those students come to me and they ask if they're in the opportunity to keep involved in research and all of that. So they work with me.

Heather Herdman (49:41.599)

Yeah

Tamara Macieira (50:00.234)

as volunteers on the projects, get authorship on papers, they do their own work as well. And I think that really helps them understand a little bit more how informatics can really guide their lives in their clinical practice. And what I say to both my undergraduate nursing students, but also my graduate students is that

As nurses, you don't need to be computer scientists. You don't need to be engineers or anything like that, or informaticians to be able to do informatics work. What we as nurses need to be able to do and know it's first understand how informatics can help us. So really understand what is it within informatics that we're working with that relates to what I'm trying to do. So be able to recognize the potential of informatics and of data science to understand.

what some of those things are, right? So to understand what our language, large language model is and what it does, you don't need to understand all the programming or anything like that. Because then when you're working in this multidisciplinary teams, you are really the person as we are in nursing practice that we're the glue to the healthcare team. Nurses are the glue to the research team as well to the multidisciplinary research team. So when we work with our peers, engineers, computer scientists, we need to be able to

guide them to the specific task or to the specific question we're trying to answer. They'll come in with all of the knowledge. Let's try out this model and all of that. But you are really going to be the person evaluating what that model is giving you and if that makes sense or not. Because they will know. They don't know what an end is, right? They will know what do you mean by environmental management, which I think is enough.

Heather Herdman (51:41.62)

Thank

Tamara Macieira (51:50.082)

they won't know what that means. They won't know if this is related to what you're seeing for the patient, but you as a nurse will be able to do that. But, but as part of the curriculum of the students that are working with me, I do advise them and as part of the plan we put together, they do take courses in data science. They take courses in translational bioinformatics. They take courses in, in like entry level programming and all of that to kind of understand more.

What is it that they're getting into? So what is it that I'm trying to help them do? But also the main thing that I want them to understand, it's really the data piece and what all of those different data elements represent because...

Heather Herdman (52:22.382)

Mm-hmm.

Tamara Macieira (52:39.734)

We nurses, I think, the best positioned people to understand all the data that is being generated through health care documentation, because you know exactly if a patient has that specific, I don't know, specific temperature or something like that, you can take a look at that data point and you can see if that makes sense or not or not for that patient. if you are in your EMR, EHR, and then you document X.

And then that goes into the data. You know, the reason why that was documented because somebody else can just look at the data, some analysis, and then the models don't make sense. They're not applicable to the reality of healthcare and in nursing clinical practice. So that's kind of how I try to mentor my students. And I'm not so sure if I'm succeeding because I had a pre-nursing student, Arlen, that was working with me.

Heather Herdman (53:14.58)

Mm-hmm.

Tamara Macieira (53:33.646)

And then he was in the pre-nursing track. then on his second year, he's just like, no, I'm not doing nursing anymore. I'm going to do data science. And I'm like, OK. I don't think I was a good mentor. I didn't want you to go just that way and not nursing. But he had a passion for data itself and data science. So he went that way. now he's working at a company that does.

Heather Herdman (53:41.123)

you

Tamara Macieira (54:02.23)

research on with cancer treatments and all of that. So he's not far off of of health care. But yeah, but that's kind of what I try to do with my students. And I just I really try to get their curiosity on any kind of research, even if it's not informatics related, then for them to really be able to understand and want to do some something like that, because

Heather Herdman (54:06.984)

Yes.

Tamara Macieira (54:28.398)

That's what happened to me, right, Heather? know, 2011, you were there in Brazil at Belo Horizonte giving a talk at my university. And I had been working with Tanya already for a couple of years as a researcher assistant. And the reason, I always tell Tanya, the reason that I knocked on her office one time asking if I could work with her in any projects that she had was that because she gave my class a lecture on care plans.

And she really was the first one of the first faculty that I've had, and that was already my second year of nursing, that really showed the value of what we do as nurses and all the critical thinking, clinical reasoning that goes behind it. And I was like, I want to work with this woman. I want to do some of this work that she's doing. And I'm hoping that I'm doing that for my students because I think you can.

If you open up the door and give them that opportunity, can really help shape their future and they will become better nurses because of it, even if they don't do research afterwards.

Heather Herdman (55:30.682)

Absolutely. And you were very blessed to have a wonderful mentor in Tanya and throughout your career, Gail, later on in your career. So I think mentorship is probably one of the most critical things we can do for the future of the profession, really. And I can't imagine that you aren't every bit. And you know what? Maybe he belonged to data science. That doesn't mean anything bad happened there.

Tamara Macieira (55:36.802)

And Gail too, yeah.

Tamara Macieira (55:57.26)

That is true. And I always like to highlight as well the time that I went to the University of Iowa and then I worked with Dr. Sue Morehead as well and Liz Wansum and the team there because I went to the birth of Nick, of knocking Nick, right? So that was a very life changing experience to me as well. And what inspired me to come back to the U.S. and do my PhD here too. So yes, I think that if I hadn't had all the

Heather Herdman (56:16.657)

Yeah.

Tamara Macieira (56:25.324)

last mentors that I've had in my life, you wrote my letter of recommendation for the PhD program. I wouldn't be here now, honestly, because it's very hard when you don't have people supporting you and helping open doors and get you there.

Heather Herdman (56:34.151)

and

Heather Herdman (56:38.32)

Absolutely. And I think it's important for us to remember it's, you know, like, that's our job. Our job is to do that for the next generation, right? So I can talk to you forever. But it's always a pleasure to speak with you. And it's been amazing to watch your career just blossoming. And I'm excited to see where you go next.

Tamara Macieira (56:45.366)

Yes. Yes.

Tamara Macieira (56:59.406)

Thank you.

Heather Herdman (57:00.052)

But if our guests have enjoyed the conversation, I hope you will give us a review on the platform you're using for your podcast. And we're really interested in topics you'd like to hear about as well as individuals you might want to hear from. So drop me a line at ceoamanda.org.

And I hope that you'll join us at our international conference in Lisbon, which is this coming June. It's all on our website at nanda.org. And while you're there, don't forget to sign up for the newsletter. Follow us on LinkedIn and Facebook. And you can also find links to purchase the NANDA International Nursing Diagnosis Definitions and Classifications text. So until then, let's keep defining nursing one concept at a time.

Heather Herdman (57:44.018)

And that's that.

Tamara Macieira (57:45.912)

Wow!

Heather Herdman (57:47.41)

That was really good.