Use of Machine Learning to Predict Outcomes in Patients With Sickle Cell Disease Treated for Vaso-Occlusive Crisis
In this podcast, Nirmish Shah, MD, discusses using machine learning and leveraging data analysis to make patient predictions about pain and readmission, to use interventions to minimize complications. Dr Shah’s team presented research on this topic titled “Use of Machine Learning to Predict 30-Day Reutilization of Care for Patients with Sickle Cell Disease Treated for Vaso-Occlusive Crisis” at the American Society of Hematology’s 2023 annual meeting in San Diego, CA.
Additional Resources:
Use of machine learning to predict 30-day reutilization of care for patients with sickle cell disease treated for vaso-occlusive crisis. Paper presented at: American Society of Hematology's Annual Meeting & Exposition; December 9-12, 2023. Accessed December 19, 2023. https://ash.confex.com/ash/2023/webprogram/Paper188165.html
For more coverage of ASH 2023, visit the newsroom.
TRANSCRIPTION:
Jessica Bard: Hello everyone, and welcome to another installment of Podcast360, your go-to resource for medical education and clinical updates. I'm your moderator, Jessica Bard with Consultant360, a multidisciplinary medical information network. Dr Nirmish Shah is here to speak with us today about his team's research presented at ASH 2023 titled "Use of Machine Learning to Predict 30-Day Reutilization of Care for Patients with Sickle Cell Disease Treated for Vaso-Occlusive Crisis." Dr Shah is the Director of the Sickle Cell Transition Program and the Director of Clinical Research in benign hematology at Duke University School of Medicine in Durham, North Carolina. Thank you for joining us today. Please provide us with an overview of this research.
Dr Nirmish Shah: This is an exciting part of what I do nowadays, which is to work with a lot of super smart investigators that do machine learning and algorithm development based on big data. One aspect of what I have been focusing on is whether we can use physiologic data–and that can come from vital signs, it can come from your Apple Watch, they can come from multiple measures that show how your body's doing, so heart rate activity, sleep, etc., anything that you can get that is a reflection of your physiologic status–and have it predict an actual clinical endpoint.
And so in the past, leading up to this abstract, we've been focusing on pain itself. Can we predict pain? And so we have a number of publications and investigations to show that we do a fairly good job. We're actually kind of in a 70% range of predicting a pain score at any given moment, based on just physiologic data. But the next step is to say, well, that's pain in any given moment. Can I predict what's going to happen in the near future?
And so one really important part of taking care of patients with sickle cell is to recognize that pain is unpredictable. And many patients who come in and are hospitalized in pain, come back to the hospital within 30 days. And readmission rates are historically high. The range is somewhere between the 30% and 40% range, meaning you have a patient admitted, you get them better from pain, but there's a 30% to 40% chance they're going to come right back within 30 days to the hospital. And so it's an unfortunate situation for patients, but just as important, it's a difficult situation for the medical teams to figure out which patient's going to come back. How do I help them? What support can I give them?
So, the overview for this study was can I take medical record data, and in particular vital signs, that would be heart rate, blood pressure systolic and diastolic pulse ox, and whether their pain scores, and can I predict that they're going to come back within 30 days? And so we ran a number of machine learning algorithms–there are many different techniques now– but we ran a number of them on that data for 19 patients that had been admitted. And then we looked at a number of the patients and tried to determine, for those that were admitted, what's our ability to predict that they're going to come back within 30 days? What we try to do is calculate whether they would come back within 30 days. What was the date that they would come in? And would that day be less than 30 days or beyond 30 days?
What we found, which is I think exciting, is that just looking at medical records, and of course, we'll get into what the next steps are, but what we found was that we're just above 60% accuracy of predicting that the patient's coming back within 30 days just on vital signs just looking at their physiologic data.
Jessica Bard: What is next for research on this topic here?
Dr Shah: Yeah, so there are several different ways we can look at this going forward. The first is that, of course, we're really oversimplifying the dataset by just looking at vital signs. There are a lot of other aspects that are just as important, if not very relevant, to what's going to happen in the next 30 days. And that includes what medications they're on and their comorbidities. For example, if they have a history of avascular necrosis, which is necrosis of, let's say their hip, then that's a source of pain, maybe a reason for them to come back quickly. Do they have chronic pain? All these comorbidities need to be added. What is their historical readmission rate in the past? How far do they live? What is the temperature outside? So that actually is the next step for us is to layer all that data into this analysis, as well and then further refine it. So this was really just proof of concept. Can we do better than just random guessing? And can we build an initial algorithm based on very simple physiologic data? Because that's what we've been doing, and now that we have a little bit of a signal, we have a little bit of promising results, can we make it even better? And I think we can. Because as I mentioned, many things seem to be very contributory to readmissions, and I think we can add those types of measures and hopefully improve on our data.
Jessica Bard: Now, let's sum it up here. What are the overall take-home messages?
Dr Shah: The take-home messages, I think we can think about it in a couple of different ways. The first is that I think we have to recognize that there is data that we may not understand but is helpful for us to understand our patients. But we need help to do that. We need help to digest that data. And so I often get a very common comment, and that is we already have so many things to process as a medical provider, as a physician, I'm looking at labs and diagnoses and the patient history, and now you're asking me to now think about another data point. Well, my goal hopefully, again, coming from experience is that I think it's a relative risk that we're trying to build here, is that the best analogy I always give is when we build algorithms for the weather, when we have a weather prediction app, we don't always or very rarely say with 100% accuracy that this is going to be rain today.
But we often say 80% chance of rain or 70%. So if we can build those types of predictions into outcomes like clinical outcomes for pain, of course, but also readmission or for complications, I think that would be very meaningful. Where again, we have to put it all into context. But the more that needle moves towards 80% and 90%, that really, I think raises a lot of concerns to the medical team to say, well, what can I do to try and hopefully move that needle back down? And then the other considerations are that I think we need to think systematically. And so if we identify, for example, that the vital signs are an important part, but there are certain aspects of the vital signs that are most important. So, I'll bring up an example, you find out that the sleep quality is a huge aspect, or you find out that the heart rate and activity is a huge aspect. Well, are there interventions that improve each of those aspects? That might be a way to address what is eventually going to lead to higher readmission rates or bad outcomes in general.
Jessica Bard: Dr Shah, thank you so much. Is there anything else you'd like to add that you think that we missed?
Dr Shah: Yeah, I would just promote the concept that, I think, machine learning, AI, and leveraging the computer to do all this analysis is really where we're moving. I would love many investigators to work together because it actually takes a lot of data to try and make this kind of analysis work best. I think as investigators across the country, across different institutions, I think if we could put our heads together and put our data sets together, which is probably a whole other podcast of how to best do that, but getting data together I think is going to be a huge opportunity for us to learn more about our patients and also potentially about how to approach our patients and manage them.
Jessica Bard: Perfect. Well, thank you again for joining us on the podcast today.
Dr Shah: No, thank you again for the opportunity.