Hi everyone, thanks for attending this talk. I will be talking about using AI to guide critical care decision making by predicting morbidity and mortality. Here is the outline of my talk. First, I will have a brief introduction to this topic. Then I will talk a little bit about how we can use AI in these kind of settings for predicting acuity, mortality, trajectory predictions, and what kind of changes do we need to make to existing architectures. And then I will talk a little bit also about the existing challenges and some of the future directions that we can take in order to address the shortcomings of existing models. So let's start with a brief introduction to this topic. As you have seen in the other talks today, there are no really shortage of examples for applications of AI in medicine. It could be examples from clinical decision support system, diagnosis aid, prognosis prediction, robotic surgery, mHealth. And then more specifically in acute care settings and in critical care settings, the use of AI is rapidly expanding and we can see many examples of applications ranging from using AI for early prediction of sepsis based on structured data such as the data that is available in electronic health record system and also combined with unstructured data in terms of the clinical notes. There are some recent works for using AI for improving kidney care, for using AI to guide some of the intervention based on reinforcement learning. And one thing that I want to make clear here is that those systems are AI augmented healthcare system. And I think they will be in the long term also AI augmented healthcare system. And they might turn into something that we call as the conditional autonomous healthcare system, which basically is the case where some of the low level decisions will be autonomously made, but most of the critical decisions will be still made by humans. So still we will have humans in the loop and they will not be high automation or full automation. And those concepts are based on this diagram that you see here, a paper by Eric Topol that compares the stages of medical automation to self-driving cars and autonomous driving from no automation, driver assistance, partial automation to conditional automation, high automation and full automation. I think it's important for all of us to have this in perspective where we are going to be in the near term and how does it look like also in the long term. Now, I want to talk a little bit about some of the projects that we have been working on along with our colleagues, Dr. Asher Pihorek, Dr. Tyler Loftus and a number of our students and postdocs. Really team effort and we are grateful to also have been supported by NSF and NIH in all of those projects. The first one that I want to talk about is regarding acuity assessment and our motivation for this project was, can we predict a real time acuity score for each ICU patient? And we know that those acuity scores in general, the scores like this can be very helpful, especially if they are precise for patient trajectory monitoring to see whether the patient is improving, whether they are deteriorating, and also for clinical resource allocation, as we have seen during the pandemic, that can be really critical. And as you know, there are a number of scores that are already being used in clinical practice, including something like sequential organ failure assessment, SOFA score, or some of the other scores such as the Apache score. And those are really helpful in practice and they have been used for quite some time. But as I will talk about that in my next slide, still they have some shortcomings that we need to keep in mind. Now, let's take a look at the SOFA score. As we know, SOFA score is based on six organ system and 13 variables and those range from respiration calculation all the way to renal and liver system. And each one also has a number of variables that are associated. And based on those variables and based on the subscores for each one of those organ system, then a total score can be obtained. Let me show that to you here in this next slide. And the total score that is obtained based on the scores for each one of those variables and based on simple thresholding for a single measurement in 24 hours, then we can have the serial measurement. The serial measurements basically tell us that for two consecutive measurements, if it is increasing or unchanged or decreasing, what are the chances of mortality? So helpful in practice, because again, it gives us a tool to understand the trajectory of the patient. However, as you can imagine, it's just a single score for 24 hours. So basically, it does not capture the fluctuations of patients in the 24 hours. And also, it's not based on all of the data that is being captured already in the ICU. We have a wealth of data that every minute is collected from those ICU patients. But scores such as SOFA or some of the other scores, they are typically based on the worst measurement, based on a single measurement, understandably because of the resources. But nowadays that we have those machine learning methods, those powerful AI techniques, probably we can do better. So we developed a model based on the same variables that are available in the clinical SOFA score. And we use the values for the same variables and we provided that to deep learning models that we developed. The deep learning model also has a self-attention mechanism and that helps us to identify those variables and also in terms of the temporal measures, the regions, the timestamps that are important in terms of the decision making process of our model. We developed this model on a cohort of 36,000 encounters from University of Florida. And then we also externally validated that on MIMIC dataset. And we did also the opposite experiment. So we also trained on MIMIC and then externally validated on the UF Health cohort. And in both settings, actually, we realized that as one would expect, once we took one model and we applied that in another setting, once we tested that in another setting, there is a decrease in performance. It's not very significant, but still you can see a decrease in performance, which kind of also tells us that if you are developing models as a community, we should work with each other in order to develop better models that are applicable not just to one single institution, but also across institutions. And here you can see also the model in action as it is highlighting some of the areas that are important in terms of decision making. We also applied a similar approach, but using a more recent transformer model, which you might have heard about them. They are all over the place in many different domains. For example, in national language processing, you will see them in terms of the large language model. We used a long-former architecture in this work, and we used that for predicting multiple different tests. So instead of just predicting one outcome, we used that in order to predict multiple outcomes at the same time simultaneously. And this architecture is again using an attention mechanism. It's slightly different. Now we have local attention and global attention, and they help us to, first of all, do a better job in predicting, and also they allow us to highlight again the variables and the timestamps that are important to the prediction task. And this was based on our UF cohort, about 73,000 distinct ICU stays. And as we expected, the transformer model, it provided the best performance because not surprisingly, it was able to leverage all the information in a more complex model. And also, we found out that when we have multiple tasks, that helps with the prediction model because then the parameters that are learned for one task, they can also help the prediction of our other tasks. In a slightly different direction, we also have used another type of data in order to collect more information on patients that can be helpful for predicting their trajectory and ultimately for predicting whether they can be discharged from ICU, whether there is a risk of mortality, whether there is a good chance of recovery. And that is based on our observation that there are many indices in critical care that are repetitively assessed by overburdened nurses, and that can include something like physical function. As we all know, physical function can be a very good indicator of recovery of patients in the ICU. But right now, if it is being assessed, it will be assessed manually by nurses using manual scores that they have to observe the patients, whether they are able to sit upright in the bed or whether they are able to have enough mobility. The same goes for some of the other indices, such as pain, that if we have nonverbal patients, then we have to use nonverbal pain scores. And there is quite some limitation to those assessments in the sense that those are manual. It needs to be done by overburdened nurses. Those can be subjective, depending on who is observing, who is being observed, there could be considerable variability. And those can be also non-continuous measures. There is no high granularity because the nurse will be in the room for five minutes, and once they are not in the room, there will be no indices that are being observed anymore. And we also realize that there are also many other types of information that are not captured at all, and that includes information on sleep disturbance factors, such as the bright light, the loud background noise, the excessive visitation throughout the night. And those all, as we know, can contribute to conditions such as delirium. It's important to capture them, but those are not being captured. And ultimately, all of that can contribute to the trajectory of patient recovery, but we don't have a very good way of quantifying such information. So in one of our projects, we have been collecting data from patients using over-pervasive sensing devices, and that can include sensors such as EMG sensors, accelerometry sensors, depth cameras, IoT sensors for collecting information on the light level in their room. And so far, we have collected data from about 190, I believe. We are up to 200 patients. And here you can see the system that we have used for collecting information from those patients. It is actually quite elaborate because we have to have a local recording system in the ICA room that has to be compact, that has to be usable, and then all the information ultimately has to be uploaded to a secure server to be processed, and that also has to be integrated with data from EPIC, for example. So we have a ground clinical truth for some of the events, such as the pain scores that have been obtained by the nurses, or if there is any information related to mobility. And this is a project that is still ongoing, and we have been able to develop a number of different machine learning modules that are able to identify patient mobility, their activities, and so on. Now, you may be asking how we can use those, and whether we can use such information also in conjunction with the clinical information, and can we use them, for example, for acuity prediction? And the answer is that yes, we actually did that in one of our recent papers. We tried to combine the information from EHR with the clinical information, with the pervasive sensing information, and we showed that once we incorporate the sensing information, then the performance of the model gets some boost. And that is not surprising, because in terms of patient acuity prediction, if we have information, for example, on patient mobility, we would expect that patient mobility can be a good indicator of their recovery, and that can be very helpful also for predicting their acuity. This is based on, again, our rather small cohort, and we are hoping that this can be extended and more validated. Now, let me talk also a little bit about some of the challenges and future directions. The clinical models that have been developed for predicting acuity, mortality, and similar other types of outcomes, although we have had a lot of progress still, there are a lot of challenges that also provide opportunities for making them better. And still, we have to address issues such as irregular measures that we have, how to incorporate all the data that are being collected at different intervals. Data sharing is still an issue, as we all know that because of the identification and patient privacy preservation, it's very difficult to share data from one institute with another institute and to kind of pool their data together. And it seems like some of the more recent approaches, such as federated learning, it might be a promising approach for addressing these type of problems, but we are still at the very early days of using federated learning in practice. And then in terms of interpretability, again, we have to work on that a lot more, explainability, interpretability. We still do not have a very good way of explaining our models. And in some way, some of those models might not actually have a very good, there might not be a very good way of explaining them. And we need to do also a lot of usability testing to better understand whether those interpretability models are actually useful for the end users and whether we have to develop different types of interpretability models for metals for different users. And, of course, the models that I mentioned, we're not using any clinical models, but I think the future is going to be more multimodal data, more multimodal models that are incorporating data not just from EHR, not just from EPIC, but also from pervasive sensing. And the data that we get also from the electronic health record system, it will include also clinical text. And the clinical narratives, they can contain a lot of information. And we have seen that in many of the recent publications, and hopefully we will have more of the multimodal models that can incorporate all these different sources of information. And one more thing that I want to mention is also reinforcement learning, that we have talked about predicting, predicting accuracy, predicting different types of complications. But we need to take that one step further, and I understand that's not going to be easy and there will be lots of complications and also liability issues as well. But at some point, we need to go beyond just predicting and start to think about how we can also provide intervention, how we can provide a solution based on reinforcement learning. Finally, I want to say that AI has a lot of potential for maximizing our opportunities to increase the healthcare access, minimize healthcare costs, to improve patient outcomes, and also to improve the efficiency of healthcare systems, and to help us in many tasks, including patient acuity prediction, mortality prediction. But in implementing an AI tool, we also need to have discussions on bias, potential harms, the eventual cause in terms of, for example, the carbon footprint of our large models that we are implementing. But still, I think that there will be a lot of opportunities, a lot of exciting opportunities for all of us to work on problems of future. And I want to thank everyone. This work is really team science, and I want to thank especially Dr. Azra Biharach, my close collaborator in the past seven years or so, as well as others such as Dr. Tyler Loftis, my former PhD student, Benjamin Schickel. Most of the work is based on his efforts, as well as others that have been very instrumental in all of that work. And that's all that I have. Thanks a lot. Any questions will be welcome.

AI

critical care

morbidity

mortality

clinical decision support systems

diagnosis aids