Full Interview: An Automated System to Track Patients in Real Time With Sarah Collins Rossetti and Kenrick Cato
You both are RN, PhDs. What does having that degree mean in terms of the work you do everyday? Is it purely research/teaching? Is there any clinical work?
Rossetti: My background is as a critical care nurse. I worked in many different intensive care units over the course of my clinical career. I do not practice clinically anymore. What I do now is full-time clinical informatics research. I received my PhD at Columbia from the nursing school in nursing informatics, and I completed a postdoc in biomedical informatics also at Columbia. Much of my work is in the applied space—informatics brings information science and computer science into the clinical sciences. I work a lot with electronic health records focused on patient safety, studying how systems can really come together and make patient care safer and more efficient.
Let’s back up to the pre-COVID days. You 2 are the principal investigators, or PIs, of a study called the CONCERN study. I’ve now done enough interviews to know that scientists love acronyms. Can you talk about CONCERN?
Rossetti: It stands for “Communicating Narrative Concerns from RNs.” Essentially, we know that nurses are at the bedside in the hospital with patients observing them the most of any clinicians in the hospital. They observe when patients are not doing well, they observe when patients have signs and symptoms before vital signs change that indicate that they are going to have a potentially bad event in the hospital, such as a cardiac arrest or death.
We are actually able to detect when nurses have this level of concern about a patient because we look at how they record information about the patient. They record information more frequently, because they’re going into the patient’s room more frequently, and they’re paying close attention to certain things. So we can see that pattern in the electronic health record (EHR) and we’re able to use different data science techniques to identify early when a patient’s at risk for deterioration for a bad event.
We can use that information in a prediction model so that the entire care team can be better aware earlier that the patient is at risk. So this really is a way to improve communication among the care team; the nurse is worried about the patient and is noticing these things, and we want to escalate it and make communication really clear that this is an at risk patient. Clinicians, the whole team, they’re very busy and they’re caring for multiple patients at once. So we surface that information to say, “Hey, you need to look at this patient more closely.”
How did you both get together and come up with the idea for this study?
Cato: Dr. Rossetti had started the preliminary work on the CONCERN study, and we were both in graduate school at the time at Columbia. She reached out to me to do some natural language processing work on the study. We started trying to find various signals in the data to predict patient degeneration. It was through those analyses that we started realizing that for our first paper, there was qualitative work that had been done, where nurses had indicated their concern about a patient and the increase of surveillance that they placed a comment in a nursing flowsheet. But it was really the quantitative analysis that showed that signal for deterioration. There was a signal for increased nursing surveillance that had an association with deterioration of patients. That analysis led to deeper analyses and other data mining methods. It all stacked up on each other and here we find ourselves now 8-9 years later.
What are the problems with the existing electronic health record system?
Cato: I wouldn’t exactly phrase it as problems, but with any system or interface, you have constraints because of the way it’s built. Before we had electronic health records, it was easier for clinicians—nurses, doctors, social workers—to instantly get certain information. It was easy for them to know if there was a really sick patient or if there was a patient who had been in the hospital with lots of interactions because their paper chart would be thicker. Also, information could be highlighted very quickly, like pulling the paper out of the folder a bit so it stuck out or putting it on top of a section so you’d know it was the most important thing. All of those kinds of ways of communicating information got lost with the transition to electronic health records.
Our project is less about prediction and more about fostering communication between clinicians. In the EHR, if nurses are concerned about a patient, they can’t just put a note in the front of the folder. That’s what our work is really about; doing data mining and analytics to pull that information out, looking at what the nurse has done and saying, “this nurse is more concerned about this patient,” based on what that nurse is doing and metaphorically pulling that patient note to the front of the chart so that other people can see it.
I’m sure that communication was strained once the pandemic came around. That leads me to your COVID-19 project, which is called “Scaling up for Surge Capacity and COVID-19 Patient tracking in the Electronic Health Records: Leveraging Healthcare process modeling”. Can you break that down?
Cato: The idea behind this project is that during a community-wide emergency like COVID-19 or 9/11, where you have lots of people using hospital resources at the same time, there are a couple things that happen. One is that the hospital has to change how it’s configured physically where it has to change what types of beds are being used. The EHR is configured to work during normal times. Let’s say we have a bed that usually has only a medical/surgical patient, but you have a lot of intensive care, very sick patients coming in. If you change that type of bed, the EHR still thinks that bed is a medical/surgical bed. So someone has to physically go into the EHR and say, “No, this bed is now being used for really sick patients.”
You can imagine in a time like COVID, when a lot of really sick people are coming in, the city, state, and federal government are trying to track these things: they’re asking, “how many really sick patients do you have?” So they created manual processes, which are very resource intensive. What we were trying to do was help the hospital track patients in a more automated way to understand how many really sick patients were on ventilators.
We used the same modeling methods that we used in our CONCERN project and applied that to tracking resources for COVID. The non-technical summary is that we can look at how a clinician is interacting with a patient and understand what type of patient it is. So when a clinician is interacting with a really sick patient, they write different types of notes, they take different types of measurements compared to when they’re interacting with a patient that isn’t as sick, that doesn’t need as high a level of care.
We applied data science methods—it’s actually really similar to when you go to Netflix and it says, “this movie’s a comedy,” as opposed to “this movie’s an action film.” Those methods aren’t based on someone actually going in and tagging it as that, because then you’ve led yourself to a human error.
What we wanted to achieve was two goals: one was to reduce that human error, but two, we wanted to automate it so we could have it in real time. The reason we wanted to do it in real time is because we also had a situation where it’s not just beds that turned into really sick patients’ beds; sometimes, a really sick patient would be in that bed and be treated as an ICU patient, and the next day, a not so sick patient would be in that bed. So you want, in real time, to understand what type of patient was being treated.
That’s basically what that project was about, because the research shows that resource allocation is actually one of the number one things that is associated with how well patients do. If you run out of trained people, ventilators, other resources, then the patient outcomes won’t be as good. Our goal was to help in that process to make sure we had the highest level of quality care as we should.
After you made the models, you evaluated their precision. How were your results?
Rossetti: They were very good. This was a classification; we weren’t predicting, but we were looking at the retrospective data to look at what the bed was and what the patient was. It was extremely accurate. It was a good process to be able to see how we could increase automation and validation of these manual configurations that have to occur in the EHR. Normally, they can occur manually and we have the time and resources for them to occur. With COVID, the resources were much more constrained and things had to happen quickly, and this was a way of ensuring accuracy.
Cato: From a data science perspective, our methods actually can perform really well because we’re not asking the program to learn things that are really complicated. We’re just determining whether this is a really sick patient type or a not so sick patient type, type of bed, type of person in the bed. We had really good results; our precision was about 99%, which means that for that type of situation, if you had 100 really sick patients, our method would know that almost 100 of those patients were really sick patients. The take home message from that is that human beings aren't right all the time.
What are your next steps in terms of COVID-19 research?
Cato: There are three things we’re doing now. One is the resource mapping that I was just talking about, the bed tracking. We’re going to take that and we’re going to look for funding to build our automated systems. The idea is that we want to be able to build applications that people can use throughout the country and world where this can happen automatically and a human being doesn’t have to go and count the resources every day. Once we can build that functionality, then we can also build dashboards or reporting on top of that so that if or when something like this happens again, we can get better reporting in real time. We’re going to look for funding to continue to do that and build those systems.
Two: Our work overlaps a lot with documentation, notes that clinicians are writing. We know that documentation is a big issue, especially in terms of documentation burden: how many notes a clinician usually has to write when they’re working. The EHR is great for some things, but one of the issues that has come along with the EHR is that clinicians have to write more notes than they used to, and it’s taking up a lot of time. During COVID, some of the rules around documentation are relaxed. It’s a great opportunity to see what documentation might be necessary and what might not be necessary. We’re going to do research to look at how clinicians are documenting during COVID and see what ways that we can reduce the documentation burden. That’s a grant proposal that we just submitted last week, so hopefully we get funded to do analysis for that.
Finally, we have lots of data that were produced during this epidemic, and there are lots of clinical questions that need to be answered. One of them, in the emergency department, is trying to understand which patients are going to have bad outcomes and which patients aren’t. There’s a lot of science around COVID that we don’t know yet, so we’re working on projects that help to predict; we’re applying the same methods that we applied with our other work of modeling how clinicians interact with patients, to produce prediction to say that this is a patient that, based on how everyone in the emergency department is interacting with them, might get much sicker in the next 24, 48, or 72 hours, and maybe you shouldn’t send them home.
Rossetti: Another related issue that we were looking at is which of our patients are on ventilators. Ventilators were a huge issue during COVID. In the EHR, you can only identify the information that is recorded. The information about ventilators is recorded in some ways that makes it complicated to identify the precise time at which a patient started on a ventilator and stopped on a ventilator. We’re able to apply this healthcare process modeling technique in order to have more accurate timeframes for when patients are on ventilators.
These are really important things that can then be used in different types of research. For instance, to say, “let’s look at patients who were ventilated and what their outcomes were” or “for patients that were ventilated for this period of time, how well do they do versus the patients who were ventilated for a shorter period of time?” You can start to use that information to answer some valuable research questions. This is applied clinical informatics research that can benefit data science work in many broad areas.
For any students out there who want to pursue the career path you’re on, do you have any advice?
Cato: I think the major thing is that there are a lot of different pathways to learn data science right now. Data science is very “in.” What I would say is that when you’re pursuing that, it doesn’t matter if you’re in the healthcare field, finance, environmental science, criminology, social work, English, or history; it’s important, as you’re learning, to try to get experience in those fields as well. For example, if you’re interested in data science in criminology, trying to get experience in criminology to understand the actual field, the domain.
Right now, we still have a big issue with analysis. Dr. Rossetti and I are very lucky in that we have, over the years, developed domain expertise, both clinically and analytically. It’s very helpful if you start off trying to develop that experience and that knowledge in the area you’re interested in. It’ll help you understand how applicable the different types of analyses are and how you can use those different types of analyses, if they make sense for the questions and problems you’re dealing with.
For anyone interested in learning more about what you’re doing, where can they look?
Rossetti: We have a website: concernstudy.partners.org. Our study is a multisite study, but primarily at Columbia. People can also look at my profile, as well as Dr. Cato’s profile.
Last question: To end with some optimism, what gives you hope for the future? This can be related to your research, science more broadly, or life in general.
Rossetti: So many things have occurred in the last few months and weeks. I think we’re seeing a lot of change—needed change and good change that needs to happen in the world regarding Issues around racial disparities and systemic bias. These conversations need to happen and are happening, so I see that movement being escalated, and hopefully, will continue.
Related to the specific work that I do, if you look at what happened during the COVID response, we have amazing clinicians that were and are our heroes. We hope that we can begin to demonstrate that great work through the research that my team and I do. We actually can show from the health care records the decisions clinicians are making for their patients, and the expertise that they’re applying is really amazing. We want to continue our work and continue perfecting these data science techniques, these healthcare process models, to really show the great work that clinicians are doing.
Something that we’re also working on is the issue of documentation burden. We ask clinicians to document a lot in the record. But really, clinicians are the experts. They know what patients need. The work that we’re trying to do is to surface the great expertise and decisions that clinicians are making and know the patients need so that we can hopefully start to decrease the documentation requirements on clinicians, so that the information in the record is really essential information. That can really drive better patient outcomes so that clinicians aren’t as tied to manual data entry in the record but really are able to use the record to support decision making and be with patients more for direct care; to get the systems to work for them. That’s where I think we can go; I think there’s a lot of movement and recognition that we need to help our clinicians in that regard. I’m excited for that in the context of the specific applied clinical informatics work that we do.
Cato: A couple of things. I’m a fairly geeky person. I’m going to be 50 years old next year, so I’ve been around a while. When I was younger, we didn’t quite see the analytical things as much in the popular world. You didn’t see people from the NIH on the news as much. I think that there is a lot of talk about distrust in science, but I think that science is being more infused in popular culture. The term “flattening the curve” is an epidemiological concept, and it’s being talked about; kids are talking about it. It gives me hope because I think society moves forward when you can combine science and other fields. You can bring science into the mainstream and bring it into leadership and politics and other domains. Science shouldn’t just stay in the field of academia. I’m very encouraged when I see epidemiology, data science, statistics, machine learning, talked about on the nightly news and the internet. I feel very optimistic about that.
I also feel very optimistic about how, when you look at COVID, we’ve been able to get different scientific domains to ramp up very quickly and throw a lot of research and work at these problems relatively quickly, so hopefully we can utilize those workflows for other types of problems that come up, other questions that aren’t during an epidemic or a pandemic, but other intractable problems that we’ve had. Hopefully these teams and all these great scientists that have been working on this problem can be able to pivot and work on some of the other things that we thought we couldn’t fix or solve previously.