All right, thank you for this opportunity. My name is Sriram Gadre, and I'm a hospitalist at the University of Virginia. I have no disclosures to report, no conflict of interest to report, but I would like to thank the Manning Fund for supporting this research. So as the title of my study suggests, in this study I examined the current methods of measuring overt and occult hypoxemia, and I tried to point in the direction of a possible better way to measure hypoxemia. So let's start with overt hypoxemia. To illustrate some problems that occur commonly, I'll use sepsis-3, because frankly that's where my interest in this area started. So as you know, sepsis-3 came out in 2016. It represented an exciting new paradigm in sepsis, and so I was eager to sort of implement the criteria locally at UVA and see how it impacts our sepsis surveillance. And almost immediately, it occurred to me that the respiratory SOFA score was going to be the hardest to implement, because that requires the PF ratio, and the PF ratio requires an ABG, which is almost never available outside the ICU. I tried to look for solutions to bypass this limitation, and I can categorize the things I found into four groups. One is to assume normal when missing. This can work for some components, like the hepatic component. If nobody thought to check a bilirubin level throughout the hospitalization, you can be fairly confident that liver disease wasn't an active clinical concern. But that doesn't apply to respiratory failure. Even when respiratory failure is the main problem, many patients will not have an ABG. So this isn't a safe way to go about things. I saw some ways of statistical imputation, median, or multiple imputation. This may work for retrospective research, perhaps, but even there, when the proportions of missing data are too high, this can be quite unreliable. Then the most common solution is to substitute the PF ratio with other markers of hypoxemia. The SF ratio is the closest thing to the PF ratio, and that's what I'll consider here. But this may work, but you immediately see something wrong with this. You're substituting the PaO2 with the SpO2, even though you know from biochemistry that that's not a linear relationship. So the intuition would be that you're setting yourself up for a downgraded construct validity. And I'll illustrate this point in a couple of slides. And then the final option I saw, which I thought was the best option, is what I might call physiological imputation. So basically, you take the SpO2 from the EMR, you apply a model of the oxygen dissociation curve to it, and you estimate the PF ratio that way, non-invasively. This seemed like the best way, except that when I used the classical models in my hospital population, I found that they greatly overestimate the PaO2 in hospitalized patients. So none of the options seemed quite right. So we started to look into possibly creating a better option. So the first thing we did was we trained a new model for the oxygen dissociation curve, which is shown in this equation over here. We saw that the older models were trained on data from a handful of young, healthy males. And so we trained our model on a more representative population of hospitalized patients, which is typically much older, sicker, and more diverse. And as you can see in the table over here, the imputation results are fairly clinically sensible. For example, an SpO2 of 90 would be imputed as an estimated PaO2 of about 58 or so. Next, a fairly standard approach to estimating the FiO2. I used commonly available EMR variables, such as the flow rate and the device FiO2 settings, and applied a fairly standard formula to it to get the estimated FiO2. I won't delve into this too much for this presentation, except to point to some examples where the results were clinically sensible. Putting these two together, we came up with this measure that we are calling EPFR, or Estimated PF Ratio, the key advantage being that it's just as readily available as something like an SF Ratio, but it has better construct validity. And I'll illustrate this point with an example. So imagine this is the EMR data from one patient over time. They come in saturating 98% on room air, but over time, they start to desaturate, first to 92, then to 85. Once the SpO2 drops below the threshold of 90, the nurse initiates supplemental oxygen, and then the SpO2 rises to 91. If you were to examine this trend in the EMR and try to determine the first sign that something was going on, I suspect you'll agree with me that it's here when the patient first desaturated from 98 to 92. That's the earliest sign that something was going on. Now let's see if the SF Ratio can catch that deterioration. As it turns out, if you use the SF Ratio, even until the patient desaturates to 85%, the SF Ratio remains quite high. It's only after the FiO2 is documented as being elevated that the SF Ratio catches the deterioration, and that's the downgraded construct validity that I was referring to. The EPFR, in contrast, correctly identifies the time point of the initial deterioration. So I'll point you to the figure one in my paper to see multiple such comparisons, but the key take-home point here is that EPFR has better construct validity than other markers of overt hypoxemia. And now let's talk about occult hypoxemia. This is the scenario where your hemoglobin desaturates, but the pulse oximeter fails to detect that desaturation. This occurs more frequently in darker-skinned individuals, and a skin color-related bias in the pulse oximeter is thought to be the culprit. This disparity has been recognized for decades. So this reference by Gibran and Tobin was talking about it in the 1990s, and nothing's been done for these three decades, but luckily with the COVID pandemic, this problem has come into focus again, these three references from six to eight, or three recent major studies looking at it. I'll point out that the eighth reference here is actually in neonates, so it's truly a cradle-to-grave kind of a problem. So it's a great thing that more people want to study occult hypoxemia, but the problem we thought about was that the study design still relies on simultaneous measurement of ABGs and SpO2s. This excludes the vast majority of patients who don't get an ABG, and also those patients who did get an ABG but didn't have a simultaneous SpO2. So the question was, could we go to a non-invasive, truly population-wide surveillance study design? Could we, for example, compare the population-wide distributions of SpO2 by race? The answer to that is no, because SpO2 isn't truly a random variable. If you think about the clinicians who monitor and record the SpO2, they aren't simply passively sampling the distribution of SpO2, rather they're actively adjusting the supplemental oxygen levels to actively regulate this distribution within a pre-specified range. And that brings me back to EPFR. EPFR, by accounting not only for any bias in oximetry, but also the clinician's response to it, is much closer to a true normal distribution because it's less in the control of the bedside clinician. And so the hypothesis being that this might be the basis, comparing population-wide distributions of EPFR, may be the basis for population-wide surveillance for occult hypoxemia. So having hypothesized these benefits of the EPFR, I'll talk about the study that we performed to test these hypotheses. We did a multi-center retrospective cohort study in three sites across two academic centers, over 5,000 hospital encounters. The sites were chosen to maximize the diversity of the study population. So while University of Virginia has a majority white population, most of the patients at the MRE sites were black. The cohort characteristics were fairly typical for the sites that the patients were enrolled. The only difference between the sites, as I said before, was the racial mix of the population. The comorbidity burden, as well as the outcome rates in the patients, also typical for acute care COVID, and no major differences between the two sites that we studied. So let's talk about the analytic methods and the results that pertain to overt hypoxemia. We modeled the risk of clinical deterioration using logistic regression. EPFR was our primary predictor of interest, and we adjusted for age, sex, race, and comorbidity. We used the adjusted odds ratio and the ROC areas as our measures of predictive validity, and we compared the performance of EPFR against other hypoxemia measures and against multi-organ dysfunction models, such as NEWS and SOFA. Although I'm aware, especially with this audience, that SOFA is trained really for mortality prediction rather than deterioration, but we just wanted to compare how the EPFR does. So with overt hypoxemia, we found that low EPFRs were a strong predictor of adverse outcomes. So 100-point drop in the EPFR would be adjusted odds ratio of 2.7 at UVA and 1.7 at Emory. The model discrimination, as expected, based on the construct validity improvements, outperformed most hypoxemia measures. But interestingly, probably because this was COVID-19, it also outperformed multi-organ dysfunction models like NEWS and SOFA. And then let's talk about the analysis and results as they pertain to occult hypoxemia. So before I describe this analysis, I wanted to say a few words about why we chose this. We thought about what might you expect to see if there was occult hypoxemia going on in a group. So what you might see is a group that's affected by occult hypoxemia disproportionately might appear to be less sick, but actually the outcomes might be worse. That's what you might expect in occult hypoxemia, because it's a false reassurance of the bedside clinician leading to undertreatment of the hypoxemia. So the first analysis was to see, do they appear to be less sick based on hypoxemia? So we compared the population distribution of hypoxemia markers by race. I'll explain the Kolmogorov-Smirnov distance in a little bit, but that was our measure for the disparity. And then the second one, did the outcomes vary by race, the relationship between hypoxemia and outcomes? Was that influence? That was a standard logistic regression. So I'll try to illustrate this study design using this figure. And I'll use SpO2 as the example. So this is the figure comparing the population distribution of pulse oxygen saturation for all recorded values at the UVA cohort. So the color represents race, with red being black patients and blue being non-black patients. X-axis is your pulse oximetry data from 85 to 100%. And the Y-axis is the percent of all recordings in that group that were less than or equal to the corresponding X-axis value. So if you take a threshold of 95% of SpO2, in non-black patients, 35% of the patients were less than or equal to this level, whereas for black patients, it was 0.28, and so the disparity was 0.7. But as you can see, this disparity varies by the threshold, and so how do you get one summary statistic that describes the overall disparity? And this is where the Kolmogorov-Smirnov distance comes in. What you do is you measure the disparity at all possible thresholds, and whatever is the maximum disparity, that's your Ks distance. The good thing about this measure is that there exists a non-parametric method to test for statistical significance. So that's why we chose this as our measure of disparity. And then the second analysis, this is much more standard. The X-axis would be pulse oximetry, the Y-axis would be outcomes, and then the measure of worse outcomes in a group would be adjusted odds ratio and multivariable regression. So with that in mind, let's look at the results from pulse oximetry and see if you see any evidence of occult hypoxemia. As I mentioned before, you would expect a big right shift in the hypoxemia measure. You see some right shift, but it's not big, and especially in the SpO2 range of about 90 to 95, where you're more worried about occult hypoxemia, you really don't see much of a disparity, and the outcomes certainly don't diverge. So if you just looked at SpO2, you might conclude that there's no evidence for occult hypoxemia going on in this population. But if you used EPFR and repeated the same analysis, that conclusion would be overturned. Here you see what you might expect with occult hypoxemia, which is a big right shift that's uniform throughout the range of the EPFR, and that's the black patients appearing less sick, less hypoxemic than the non-black patients. And yet, if you look at the outcome curves for the same level of hypoxemia, the black patients were about twice as likely to clinically deteriorate within 24 hours. So this, a group of patients appearing to be less sick, but actually having worse outcomes, this we are proposing is the population-wide signature for occult hypoxemia. So what are the implications of this study? When it comes to overt hypoxemia, the key finding is that EPFR is a simple-to-implement and accurate measure of hypoxemia. It has at least one advantage over all other hypoxemia measures. So whether you're modeling the risks in a predominantly respiratory failure syndrome like COVID-19, or even a more multi-organ syndrome like sepsis, using EPFR as your hypoxemia measure might be valuable. And when it comes to occult hypoxemia, this study proposes an entirely new type of study design which may significantly increase the scale on which we are able to study occult hypoxemia. If a population-level signature of occult hypoxemia is truly validated in larger studies, then that could open up the doors for a rigorous post-marketing validation of pulse oximeters. So the overarching hope is that we create the market forces and the regulatory climate that's needed to put an end to this longstanding structural source of inequity in healthcare. That was my study. I'm actively seeking collaborations and mentorship to take this idea forward, so I hope you won't hesitate to reach out. Thank you.

hypoxemia

oxygen levels

overt hypoxemia

occult hypoxemia

EPFR