the Datathon back in New York, and I represent Team 1. I know it says Team 8 on the program, but I think it's Team 1. Anyways, we won the patient safety category, and I want to talk about how we examined the disparities in glucose measurements. So our particular research question was, what are the key factors influencing glucose variability in point of care and serum measurements? Now, before I get into the methodology, we used to study this, I want to provide some background for this type of research and how we wanted to approach this. So recently, there's been more research revolving around the concept of digital determinants of health. Now, we know that social determinants of health are the social and environmental factors outside of the medical factors that influence the health of the patients, right? In digital determinants of health, we talk about the inequalities and the differences in patient outcomes that are not due to the social factors in themselves, but more due to how the digital technologies that we use interact with those social factors. So a lot of research, a lot of brilliant research, has come out recently that talks about identifying these digital technologies that demonstrate an evidence of bias. And such pitfalls generally arise from insufficient consideration of patient diversity. Now, we see examples of this in pulse oximetry, where the measurements differ based on the skin color of the patient. And we see this for EEG, where the conductivity quality measures are different based on the type of hair on the patient. And I can't speak for the entire team, but I know that this is definitely the overall motivation for our research, is to look at some of these factors. Because differences in glucose has been studied for 20, 13 years, in terms of how we measure that initial glucose measurement with a finger stick at the point of care, versus the more so, but more accurate, results we get from the serum blood tests from the lab. So we look at when those two measurements differ in a short span of time, what causes those errors. Now, we know that there are errors, right? Because there are measurement errors in the finger stick. And the directionality of those errors, and how it affects people in general, based on the timeline, has been studied. But we wanted to see if we could find other key factors that influence this variance. So we use the EICU database, which is a really extensive, multi-center ICU database. And of course, we want the measurements that we look at to be as close to each other in time. So that the point of care finger stick glucose measurement is taken really close to the actual blood sample that's sent to the lab. So the way we curated our data set was we selected a time span of 15 minutes before or after the different types of sampling. So we aligned these pairs that we have of glucose measurements with a lot of different other variables regarding the patient, such as the demographic variables, of course, since this is our main focus. And also treatment variables, such as insulin, steroids, vasopressors, mechanical ventilation, et cetera. Everything that also could affect this variance in some way. Now, we started out with 139,000 patients. And we excluded 93,000 or 73,000 to begin with that had no pairs within this time span. Or no pairs at all, sorry. And then we excluded 7,000 pairs more because we wanted these 15 minutes. And then we also excluded 11,000 pairs from hospitals with less than 1,000 patients. So that left us with 46,000 patients and 190,000 pairs of glucose measurements. Now, when we're investigating bias in general and how different errors might disproportionately affect certain parts of the population, we also are really considerate of how that data got into the data set to begin with. So a really important thing is to look for patterns of missingness in the data. And this is what's called an offset plot, where we can see the different combinations of missing variables and how they're distributed across our cohort. So here we can see that we had a lot of missingness for all the insulin, vasopressin, lactate, and then the different combinations of missingness. Now, we also looked at how these were missing. And we found that they were missing at random across the demographics, which is also really important to consider when looking at sampling bias. Now, we found out after selecting these pairs that the general error was really nicely and normally distributed. So there was no directionality in the way these measurements were wrongfully taken, if that makes sense. And we can see that the distribution based on ethnicity is also there are no obvious differences here. So what we do is we specify an outcome called occult hypoglycemia. That is, we're interested in patients where the point of care glucose show above 70 milligrams per deciliter. And then the serum glucose shows below that. So we don't want, I mean, we accept some error in these finger stick measurements, right? But we don't want them to cause any harm disproportionately to a certain part of a patient demographic because they're affected by the directionality of this error. So how can we model this? Well, we fit a patient logistic regression model for each hospital to look at if we can predict this occult hypoglycemia. So who are measured as being higher than 70 milligrams per deciliter from the finger stick? And then when the serum value comes back, we see that that was actually much lower. So when we look at the posterior estimates here, we can see that each line represents a hospital. And if you're familiar with Bayesian statistics, these are basically the parameter estimates of the effect on ethnicity for these things. So instead of getting one parameter in typical frequency statistics, we can account for the uncertainty of that parameter estimation, basically. So we see a lot of hospitals where there's no effect on race, ethnicity, on these occult hypoglycemia harmful outcome. But then we also see a subset of hospitals with a really high probability of unequal harm. Now, it's hard to interpret this in the way that we should go out and point to these hospitals and say you're disproportionately harming certain patients because, of course, there are a lot of things that can explain this variability. But at least it's a start for us. We were also having a busy time during the Datathon in those two days, right? So we don't want to draw any conclusions too fast. But it's definitely some interesting preliminary results that we have here. Of course, those hospitals that were flaring up in terms of the effect parameters could have certain devices. They could all share some protocols or practices that are interesting to look into. And we can see, well, if these devices are consistently giving error in a certain direction, then, well, we can do something about it, right? And this is also the benefit of working with these large data sets. So for future work and diving deeper into this project, we really want to look at newer data, look at these particular hospitals, model them in more sophisticated ways, and then also adjust for some of the co-founders we collected also from the EICU. So thank you. This was Team Sugar Data. And I really want to thank all of the members on the team, Andrew, Enrique, Joao, Rishika, Vishwa Rao. Yeah, this was far from a solo effort. I mean, everybody did a lot. And I think if you have any questions from the clinical aspect, I'm a data science student. You should definitely ask Dr. Barrows down here, because I think he was really annoyed with me in the end of the data thing, because I was asking questions like 24-7. What does this value mean? What is that? What is this? Yeah, so yeah, thank you. So thank you so much. Thank you.

glucose variability

patient safety

EICU database

measurement errors

digital determinants