Welcome to our session. It's a joint session between the European Society of Intensive Care Medicine and the Society of Critical Care Medicine at our 51st Congress. I'm Sandra Cangill, a professor at the University of Pittsburgh in the School of Pharmacy and the president for the Society of Critical Care Medicine. Today we'll be discussing how data science will affect you and your patients. We're going to begin by having Dr. Ciccone and Dr. Martin speak, and then we'll have a panel discussion. So it's my pleasure to introduce Dr. Maurizio Ciccone, who is the head of the Anesthesia and Intensive Care Department at Humanities Research Hospital, and also the president of the European Society of Intensive Care Medicine. Welcome Dr. Ciccone. Well, thank you, Sandy, for this kind introduction, and it's great to be here together in this joint session between SCCM and ESSIC to talk about data science. This was indeed one of the major focuses of our societies before COVID. It stayed there even during COVID times, but I know that we will invest more and more resources on this in the future. And why data science? Why artificial intelligence? Why data in intensive care? Well, we always had a lot of data in intensive care, even well before we had possibilities to work with computers to try to see whether we can apply artificial intelligence algorithms to something like this. And as we know, the intensive care is a very data-rich environment where we can really get physiological data continuously. We see that on our dashboards, we see this in our control stations, in our units, but we can also bring this pretty much everywhere now. So we have so much data in intensive care, and really we have so much information, but it's almost now a paradox that we have so much data and so much information, which actually brings a problem. And the problem is that we sometimes lose the information in the noise. And I think what we really try to do with our expertise, with our experience, and now more and more with the help of computers, is actually to try to find what is the real signal among all the noise. Now, if I see these trends here between the green line and the red line, I would be very happy if I'm looking at this after I decided to do a research study. And if I decide to do a research study and I see this, I see this is a perfect correlation. Maybe somebody would already jump and say it's a perfect association here. Maybe there is also causality in this. But in order to see this, I cannot just find the two variables that are correlated with each other. What I have to see is also if there is a plausibility for the association that I see. And that may not just be enough. And this is real data that I got actually from statistical data from the US. And if you look at what these two lines are, however, these are the number of civil engineering doctorals awarded in the US and the per capita consumption of mozzarella cheese. Now, somebody may argue that if you eat more cheese, maybe when you study, you're more likely to get your doctoral degree in civil engineering or maybe the civil engineers, they really like to eat mozzarella. But clearly, this is just a correlation that you find just because you are doing a fishing expedition with all the data that you have. There is no plausibility here. And that is very important when we look at this in our intensive care units, in our databases. It's so easy now just to get variables and just to see whether we can find an association. But sometimes this association can just be random. Or in other words, as Ronald Casey, a former Nobel Prize in economics, was saying, if you torture the data long enough, it will confess to anything. So to me, it's very important that now that it becomes easier just to get these databases and use computers to explore the data, we don't lose some of the key principles of good research, looking at primary research questions, looking at really at what is that is interesting us, interested for us in terms of improving patient outcomes and the experience for the families that are in the intensive care. If I have to talk about the first example of data science, maybe the first example of epidemiology, what comes to my mind is the story of Jon Snow. Jon Snow was one of the first anesthesiologists and also one of the first epidemiologists. And when I give lectures about data science to my medical school students, when I talk about Jon Snow, they tell me that there is a character from a TV series, which I didn't know, but apparently is quite a famous one. But the Jon Snow that I really like to refer to is the one that was instrumental in finding a solution to one of the worst cholera outbreaks in London. In 1854, there was a cholera outbreak in the Soho region of London, in Broad Street. And what Jon Snow did was getting a map and plotting on this map all the deaths from the households. And he noticed that they were concentrated in an area very close to a water pump. Now, I remember the background at the time. We didn't know all the things that we know nowadays about transmissions of bugs, bacteria. We didn't really know the theories about the vibrio of cholera and so on. People were talking about miasmas, which was the bad air, the smell that could transmit something. But they really didn't understand yet the link between water, contaminated water and cholera. And that's what he basically did just by looking at where the concentration of events was. And he noticed that the majority of households that were going to that water pump were basically where there was a high concentration of deaths, with an exception of monks working in a brewery. Why is that? Because those that were not drinking water from the pump, they were actually drinking beer. You may argue that the lesson from the story is that beer is better than water. Or the fact that just by looking at data in a very methodical way, you can start to see whether there is an association between one variable and the other. Now, to prove that an association is also a link with causality, then you need to do an intervention. And by doing that intervention, you have to see if then something changes. The intervention was to stop people drinking from that water. And indeed, that's what happened in the cholera outbreak resolved. And if you go to London, I lived there for 14 years. And if you go to Soho, you can still find the water, the original water pump of the cholera outbreak of 1854. And as you see, the water handle was removed. And in the background, you see a pub which is named after John Snow. This to me, it's a story from many years ago, but it's everything about data. It's everything about looking at data science and understanding if we can spot associations, if we can spot causality, and if we can really use this to change the outcomes for our patients. And of course, to do and to to study and to do statistics and to to do research questions on data science, the first thing that we need to have is data. And I think more and more we need to understand that data is for everyone to use and to and for everyone to explore. And in this sense, I really believe that what happened in in America and the US was really fantastic. And many years ago, the Physionet and the MIMIC databases at Harvard, the collaboration from MIT and the intensive care units at the medical school and these databases that were put for free for everyone, this was really a pioneer approach in this sense. And by doing this, we are basically giving the possibility to researchers everywhere around the world to explore data and to see if we can come with research questions with an interesting idea that then we can explore maybe in larger databases. And the good thing is that from that pioneeristic approach of more than 20 years ago, suddenly we are seeing more collaborations also in other parts of the world. Amsterdam ICU with Paul Elbers and others, they were able to put together a database to be used again by anyone who wants to do research into this. This is the Amsterdam database that we published in Critical Care Medicine in 2021. But there are more and more coming around. And I really hope that we will come to a point where we understand that we have to share in a fully anonymized way the information that we have from our patients. That's the only way that we can advance. And I think even during this pandemic, we have learned how important it is to share information, how important it is to put the information together so that we can get very fast research done on the topics that are important for us. Different ways for which we can explore the data. We can do a lot of analysis now. We can do applied techniques. They are not as complicated as they seem. Actually, you realize more and more that they are available for many research groups using supervised learning, unsupervised learning, reinforcement learning, just to name a few. I'll show you some of the things that we've done with my group just to exercise. This was actually done on some of the patients from the MIMIC database. What you see in white here is written in Italian, but you can understand it's norepinephrine dosage and pressione arteriosa is the arterial blood pressure. And what you see, the white line is what the artificial intelligence algorithm is predicting will be the dose that then the clinician, which is the following blue, light blue line that follows, will give. And as you see, the arterial blood pressure is the one that you can see in the bottom part. And in practice, the algorithm is able to predict even before seeing the arterial blood pressure signal, what would be the dose that the clinician then gives. And now, if you work on a very restricted data sets and you put this data into the algorithms, they become very good in spotting this. And I can show that this works for every single patient in the database. However, what the system is doing here is not really creating a way to predict very well the unknown. It's just basically mapping the whole database and finding ways to fit the data in this way. So this is something that then if you apply the same algorithm to an independent data set, it may fail dramatically. This is what we call basically overfitting. So overfitting means that we are not very good at finding the separation of the points based on something that will work also independent data that you will see in other databases. In order to see if what we find with one algorithm will work also with others, that's exactly what we have to do. We have to use training course and testing course. And the best thing is then to test the algorithm to a data set that is being collected completely independently from the first one. You may all have seen this study that was published in Natural Medicine in 2018 from the group of Komoroski, Omar Badawi, Anthony Gordon and so on. And as you know, they were able to see that an artificial intelligence clinician suggestion for amount of fluids or amount of vasopressor was basically able to suggest fluids and vasopressor in such a way that when the real clinician was diverging too much from these suggestions, in practice, there was an increased signal for mortality. Interestingly, as a U-shaped relationship, very clearly defined here for vasopressors. Now, this study was a milestone study. Everyone looked at it and many people were saying, why don't we start to use this algorithm in our daily practice? Well, I think we have to be careful with this. Even if results like these are extraordinary and we see some new possibilities, what we don't know is if it's something that you have tested, even simulating an intervention in an observational data set, will hold the truth than with a real intervention. And that is because maybe what we are recording on our data sets is not really the real variable that maybe needs to be manipulated. It could be that we are recording a lot of data, but not the real one that maybe matters for this. An example for this I really like to refer to is the story of Abraham Wald. Abraham Wald was a statistician from Europe, has to escape the Nazi regime during the Second World War and went to the U.S. And there he started to cooperate with the U.S. Army. And what was happening during the battles of the airplanes, when the airplanes were landing at the airport, they were inspected by engineers. And with the inspections that were made, they were finding where the bullet holes were. And the engineers were putting extra steel where they were seeing these holes, with the idea that the airplane would be more robust, more solid, more resilient to further shooting. Abraham Wald comes along and says, baby, you're doing everything wrong here. You're just making the airplanes heavier. What you are doing is actually just focusing on the known critical injuries, because all these airplanes that you're seeing are the ones that have not been damaged enough and they can make it to the airport. And the ones that are really having injuries in other parts of the world, they cannot make it. And usually those are the ones that are injured in the cockpit or in the engine. So it could be that what we are recording is not necessarily everything that is important in this sense. And I do believe that if we want to move forward, we really have to find the way to allow for big data to meet big trial. There has been an endless debate in our community about whether randomized controlled trials are useful or not. I really like this editorial of Derek Angozo a few years ago, where he was saying that basically we have to find smarter ways to do trials in ICU, where we really learn basically while doing and we use the data that we are collecting, even while we do the study, to adapt the study design. Interesting, this was the background for some of the studies that the research platforms that we have been able to use during this pandemic, like RemapCAP, in a way also Recovery. They were all based on this principle of trying to learn while you're doing, bringing observations, so large data sets and the trials at the same time. So to conclude, I think we are going to live in a very exciting time over the next 10 years with a lot of changes that are already occurring. One may argue that we may go to a full automation on everything that we do in intensive care, from collecting the data to algorithms that are already deciding on what to do for us. I don't think that's necessarily the best way forward, not at the beginning, at least. I think what we wanted to do is to have a more, let's say, augmentation approach, where our reality is augmented by the information so that we can really spot things that maybe are not easy to see immediately. And these algorithms, they can show the data in a way that we can be faster and better in what we are doing. I really like on this example, this study that was published in Critical Care Medicine, and it was about using an algorithm to identify patients with sepsis very early so that the providers could give antibiotics according to antibiotic stewardship. And what they found was that they were decreasing the time for the administration of antibiotics in patients with sepsis and they were able to have better outcomes with shorter length of stay. So I think we, in order to become more and more familiar with this, we also have to learn that we cannot do all of this just on our own. We need to basically bring more and more teams in our clinical teams and mix our teams. What we were doing before COVID-19, and hopefully we will restart again, was also doing datatones. Datatones are these events where we bring together clinicians and data scientists. We create a mixed team, they challenge each other, and they try to answer research questions. And it's really amazing what you can get when you have a team with different competencies. Some people may be more experienced about data handling, somebody more about clinical research questions, and I really think we have to move forward in that direction and do more of this. And to conclude, I am concluding by quoting Eric Topol. I invite you to read this book if you have not done it yet. Basically, one of the quotes from this book is that artificial intelligence hopefully will free time for clinicians so that they can spend more time at the bedside with patients and families. And I really think that is the key in what we are doing here, using technology to improve the experience of patients, families and health care workers. Thank you.

data science

intensive care units

research question

collaboration

anonymized patient data

artificial intelligence