Good morning, good afternoon, good evening, everyone. I'm very proud to co-chair this joint SCCM ESICM session. I'm Dr. Elie Azoulay from the Paris University. I'm the president-elect of the European Society of Intensive Care Medicine. It's my great honor to introduce Dr. Greg Martin, professor at Emory University School of Medicine and research director for the Emory Critical Care and immediate past president of the Society of Critical Care Medicine. Welcome, Dr. Martin. Thank you for that kind introduction. It's my pleasure to be here today. I'm really looking forward to adding now and talking about where data science is going to fit, particularly at the interface of clinical care. I'm Greg Martin, the past president for the Society of Critical Care Medicine and a pulmonary and critical care physician at Emory University in Atlanta. I don't have any conflicts of interest with this talk, and what I'm really going to spend my time talking about for the next few minutes is first some foundational things about data science, the way we might apply data science, thinking about artificial intelligence and machine learning, and then finally spend most of our time talking about where it's going to interface and how it might improve patient care. So let's start off with some foundations, make sure we're all on the same page. So thinking about data science particularly, and Maurizio really has done a nice job describing this, thinking about it as the discipline that combines several things like mathematics and science, statistics, informatics together, but the key is that they're using computers to gain insights from data, and particularly as we apply it to healthcare, it's using those insights for patient care and to improve healthcare systems. The components of data science that we'll really talk about, two things today, artificial intelligence and machine learning. Artificial intelligence, we talk about a lot, we hear about a lot, but really the key to this is the idea that you can autonomously or separately recreate intelligence using computers to create intelligence as you would perhaps from human beings. And then machine learning is then a branch of artificial intelligence, and the key here that I think of is the ability for developing algorithms that can continuously learn, continuously evolve, because they're learning from the data that they're taking in, and they become more free from that human intervention. Now if we pivot just a little bit, the third concept I'll give you is really about something that William Edwards Deming talked about, and you probably know him as one of the quality gurus from many, many years ago, but the Deming principle was the idea that you could adopt principles of management and that organizations can both increase quality and reduce cost at the same time, and the process here was one of the things he championed, which was the PDSA cycle, the Plan, Do, Study, Act, that we often use in healthcare as well. And the concept here is that you can plan these small changes, you can put those into place and see them change, and you begin to go through this cycle where you're making changes to improve care. But the interface here for data science is that without data, you are just another person with an opinion. So data science can provide us a lot of the data, and without that data, we really just are relying on anecdotes and maybe even prior clinical experience, but it's not the foundation that we need to really drive the best patient care that we would like. If you put all those things together in a more global sense, the idea of integrating data science into healthcare can lead us to what we talk about as a learning healthcare system, a system in which you're taking all the data together and things like science, informatics, and even the culture of an institution and integrating those to create the best practices. And the idea of a learning healthcare system is that you have input data that's coming in and continuously being integrated and analyzed, and output data that's continually being used to optimize patient care. And it's that same cycle, like a PDSA cycle, that's making it happen. Machine learning can be the foundation for that, and one of the things we've talked about is that if you're taking that afferent arm of data science, where the data are coming in, and the efferent arm, where it's then flowing back out to influence patient care, is the foundation for a learning healthcare system. And by using data science, we can facilitate the continuous collection, integration, analysis of all these data pieces that will allow us to perform better patient care. Now beginning to think about how this is going to impact what we do in the ICU, you can easily imagine that you could develop algorithms that could improve inpatient care, ICU care in many ways. For instance, making accurate medical diagnoses, particularly in a timely manner, early identification of new conditions, or even potential complications or deterioration with conditions, ensuring the highest quality care, what we talk about for SCCM is the right care, right now strategy. Transitions of care that happen in every patient every day, preventing things like readmissions, and even individualizing care, so long-term care needs after critical illness, and even things like the appropriate and most optimal use of palliative care. So what are the current clinical challenges? Well, what happens when we as clinicians are seeing a patient change? Patients are dynamic, and the trajectory changes from time to time. What happens when the care providers change? I mentioned that you would transition from one environment to another, or even there's a handoff from one provider to another. The patient type can change if you're familiar with or most used to taking care of patients of a certain type, but now a patient of a different type comes in from your CVICU, and suddenly you have a trauma patient or a sepsis patient. And then finally, maybe the biggest challenge we all deal with already every day is what happens as the evidence changes? So think about this. What if we imagine that the foundation of clinical care was the current state of evidence, and it was all objectively graded and individually optimized? With that in mind, then the application of artificial intelligence and machine learning to critical illness could look like this. Your electronic health record system, or some electronic system, EHR or otherwise, will read the current literature, be completely informed by it, will integrate all that with the prior evidence, the things like systematic reviews and guidelines that we know. It'll know the relevant local, regional, global impact of what's happening, for instance, your local antibiogram, or even travel patterns, local population differences, things that are going to influence what happens in your hospital and with your patient. And then finally, it's going to capture those patient-specific pieces that we may not always be aware of, things that come from the ambulance, people that, or sorry, information that even comes from a pre-hospital environment, like someone wearing their watch or their Fitbit or their actigraphy information from their smartphone. All of that can be integrated to the local environment, and now it becomes available to you to take care of your patient. So if you imagine data science at the interface of your patient, let's talk about a case. So a 68-year-old African-American woman with poorly controlled hypertension and combined systolic and diastolic dysfunction comes in 20 days after her pneumonectomy. She presents to the emergency department with acute encephalopathy and dyspnea. Her vital signs you see here, she has a low temperature, she's tachycardic and tachypneic, her blood pressure is low, her oxygen saturation is low, and her Glasgow coma score is low as well. So one of the things that we'll think about is where does data science help to augment patient care? How can it be applied? So one would be the diagnosis. What's the diagnosis for this patient? Is there an infection present? Does this represent sepsis? Is it septic shock? There's several different potential clinical scenarios that we may be looking at. What's the treatment for this patient? What are the things that we need to do, and particularly in priority order? What are the things that are most life-threatening? What needs to happen now, and what are the things that we can spend more time trying to decide is the most optimal strategy? And then finally, what's the optimal supportive care? And thinking back to the learning healthcare environment, taking all that data together, knowing how the health system works, knowing your individual patient, and what needs to happen for them to deliver the best care for their recovery. One of the things that we always remember, and it's one of the things we always need to remember, is that ICU care, we realize it's time-sensitive, but we often don't think about the stakes of the decision and the fact that the amount of information that comes in can be particularly challenging for any individual provider or even an ICU team to integrate and analyze and use efficiently. And so that decision fatigue that can come along with trying to manage all that knowledge, do it in a time-sensitive manner, and make clinical decisions is one of the challenges we face as well. So where might we start if we're thinking about how does data science affect my patient? So giving you the example that I'm following, what if you said, okay, the foundation for evidence in this case for a patient with potential sepsis is a surviving sepsis campaign, which was published in the last few months. We have 81 pages and 653 articles cited as primary literature. So to be able to keep up with that or individually integrate that and decide how to apply it to your patient is always a challenge, but it talks about things like cultures and antibiotics, the use of intravenous fluids, vasopressors, measuring lactate, measuring perfusion. And there's a variety of questions that are going to apply in this patient. What volume of fluid? What type of fluid? What timing of that fluid? What are we going to measure to try and assess response to this? What antibiotic do we choose? And if you're thinking of all this information, try and make sure that you're thinking about it in terms of not just the guideline and the evidence that are cited within it, but also all the continually evolving evidence that goes with that. So I'll give you a couple of examples here. So one is choosing vasopressors. So our patient appears to have septic shock, they're hypotensive, what's your blood pressure target? Traditionally, and according to the guidelines, we would apply a mean arterial pressure target of about 65 millimeters of mercury. That's based on this study in the New England Journal, which showed that there was no difference in major outcomes, at least in survival, between the lower target and the higher target mean arterial pressure group. But if you're carefully interpreting that study, you will remember that there was a stratum, there was a subgroup that differed than that. And the patients who had chronic hypertension actually had more acute kidney injury and need for renal replacement therapy if they were at the lower blood pressure target. So it suggests perhaps in our patient, if you recall, has poorly controlled longstanding hypertension, maybe the MAP target should be higher. This is something that clearly falls within the guideline, but may be lost to the individual practitioner. It may not be intuitively obvious to them at the time that they're trying to make decisions for their patient. There's also other literature that might inform therapy. So this is a cancer patient, a patient with lung cancer is more likely to have a respiratory source of infection, that can inform the type of antibiotics we're going to choose for empiric therapy while we're trying to decide what's going to be the most effective strategy. There's also demographic factors that the data tells us will help us to define what's going to be the most likely infection. So a female with a respiratory infection is going to have more likely a gram-positive infection. And alternatively, a patient who is black and a respiratory infection is even more likely to have a gram-positive infection. So it begins to help inform a respiratory infection, but these demographics might help us to inform the optimal antibiotic strategy for that patient. How do we manage, for instance, respiratory failure in this patient, and how do we get to the individual strategy that's going to affect this patient? We often think about oxygen support strategies, non-invasive and invasive mechanical ventilation, but what about lung protective ventilation? In most cases, our respiratory therapists and our providers would say, this is a patient at risk for ARDS, they need low tidal volume lung protective ventilation at six cc per kilo. But remember, this is a patient with a pneumonectomy, and that's not intuitively obvious when people are in a time-sensitive crunch and they're trying to make decisions quickly. So you have to remember that this patient's six cc per kilo would not necessarily be the appropriate strategy. We also remember that women are less likely to receive lung protective ventilation than men, so we need to be attentive to their body size and the appropriate cc's tidal volume for them. And we also know that chest x-ray interpretation, defining ARDS, is one of the challenges, it's one of the biggest disparities in identifying ARDS as being present or not, but we also realize that chest x-ray interpretation can be automated and that we can make that a digital transformation so that we can identify ARDS electronically. A couple other pieces I want to mention. So thinking about cardiovascular management, again, a patient who has systolic and diastolic dysfunction, if you're relying on data science and the reading of the literature electronically so it's all brought to you so that you can make decisions, this is a study that came out a couple of years ago now that showed that in patients without ARDS, the use of intermediate tidal volume decreased left ventricular, right ventricular systolic function, but it didn't have an effect on diastolic dysfunction. And this could be potentially important in our patient where you're potentially going to be looking at invasive mechanical ventilation and you're trying to manage their heart failure at the same time. And stepping even one further step forward, now we're really talking about the cutting edge in shock management, and this is an article that literally just came out and it tells us here that early initiation of norepinephrine might be favorable, that it's associated with a longer survival time, more rapid resolution of shock, shorter duration of invasive mechanical ventilation, and may even help to reverse the incipient or early organ failure that we see with sepsis. This is something that would not be in a guideline, may not even have shown up in the table of contents for you to look at yet, but it's the current state of evidence. And as we think about data science, it can be more proactive in addressing and integrating these into our practice. And then the last piece that I'll tell you is think about the potential therapeutics. So we've got a person who came in hypothermic, we know that hypothermia occurs in a subset of sepsis patients. And when it does, their mortality is much higher, but it's also associated with a certain cytokine expression panel. And you can look at things like TNF, IL-6, prostaglandin, but even things like soluble RAGE can be used as a potential association for ARDS. So one of the things that you might see in this is identifying the patients who are most likely to respond to a new therapy by measuring things like soluble RAGE and TNF to identify the best optimal therapy for your patient. All of this is trying to get us from this lower hand corner, the descriptive analytics piece, where we know what happened, but we want to move along that spectrum. So we're going through the diagnostic piece. So we know why something happened to the predictive piece, where we really are now, where we know something will happen, and we may be able to predict when it will happen. But what we're trying to get to is prescriptive analytics, meaning how can we make something happen? We want to choose the optimal therapy, and we want to do it at the right time for the right patient so we can control what happens to that patient. And data science can help us to get there. So the last thing what I'll conclude with is that data science, and specifically artificial intelligence and machine learning, can and will play key roles in the future of critical care. The integration of all that data and the way we put it into patient care is the basis for a learning healthcare system, and that's where we really like to be able to get. And then finally, all of that integration together will allow us the most effective tools, and it's working together between the clinicians, the entire multi-professional team in the ICU, and the data scientists who have their expertise to work together to help us all get there. So with that, I will stop, and I look forward to our discussion. Thank you very much.

data science

artificial intelligence

machine learning

patient care

healthcare systems