Dr. Nadir Yehya - Disparities in Pediatric Critical Care: Research and Methodological Considerations

Thank you everybody for having me. I really appreciate this opportunity and to meet all of you and speak with all of you and and just jumping into it. So why is a success ARDS, picky doctor talking to you guys about disparities research and like we'll see we'll see how much that comes up in the in the talk. So From my objective, so The reason I became interested in this area is because I started reading like some literature in the area and like a big portion of my like academic portfolio is some version of you're doing it wrong. And so this was this This was a very natural like direction from it goes like I'm finding like and worried about misleading inferences from observational studies and even from clinical trials. And As I started doing more epidemiologic work, I started thinking more about like the causal framework in which we ask questions versus like, you know, like other ways that people have constructed multi variable models. And so this led me very naturally in thinking that like, oh, the consequences of this and disparities research is actually pretty profound. And so we'll talk a little bit about that. We'll talk a little bit about selection bias and collider bias. And I promise you, this won't be like just Screen full of like stats, mumbo jumbo as much as possible. And then I want to talk a little bit about some of the research we're doing at So a few things. So what does what does race actually mean in disparity studies? So And the the epidemiologic concern here is that you're talking about exposure in this classification Like in other studies of like where you could have somebody who's looking at like the association between like Dexamethasone exposure in the bypass room with like a mortality or time on the ventilator after bypass run or something like that right you're looking at like your concern could be that like You could call patients who Have got no dexamethasone in intraoperatively as no dexamethasone but they may have gotten something immediately as soon as they landed in the ICU And so that would be an exposure in this classification. You're like you're you're misclassifying patients who you call no dexamethasone as having dexamethasone as actually having dexamethasone And so this leads to no results or bias because of because of problems like that And so with race that that question then becomes like what do we talk about here? Like what did when I say that black race is associated with Like worse outcomes higher mortality greater rates of admission to the ICU. What are we talking about? And it's unlikely that we're talking about genomic ancestry It is improbable that we're actually talking about differences in like haplotypes Okay, because the inter Race what we call race variation. Okay is like as wide as the intra race And so that makes it like if you're talking about A haplotype you should call it a haplotype you should actually talk genetics as opposed to like talking about race So race as a social construct Okay, is probably not what's being reflected in most studies but talk about exposure of black race or another race or ethnicity without So what are we talking about? We're probably talking about racism. Okay, which is the lived experience of being black or being a different race Okay, which is typically referenced nonco to to you know like non-Hispanic white as is most commonly done in studies in the United States as well as like in in England and in a lot of Europe and a lot of European countries who may do this But not not universally other countries that do disparities research also typically reference it against the majority of population Okay, and so in this country like very often that ends up being non-Hispanic white If we think of races this okay as like we're actually talking about the social the effect of racism on outcome All right, that's probably a more accurate lens and a less likely to be misclassified exposure So the IOM in 2002 came up with this definition of of disparities in healthcare which I absolutely hate and so describe it as Due to racial ethnic differences in quality of care that are not due to access related factors preferences and appropriateness of intervention Which to me is like and you'll see why this ends up being a problem once we actually gets all the dags and the collider bias But like but that that's a usually problematic definition because so much of disparities in healthcare due to these three factors And so the CDC definition I think is like much cleaner and and because it like coincides with my biases. I like it more and so I'm sorry I didn't mean to advance preventable differences in the burden of disease injury violence or opportunities to achieve optimal health that are experienced by socially disadvantaged populations So why does this matter in research? Well in observational studies I'm going to talk about like like how this leads to selection bias and collider bias and we're going to spend some time on collider bias And in clinical trials like this ends up causing a selection bias which ends up like affecting the inferences that you get from studies like this And so misleading inferences from observational studies and trials so I want to run through very quickly some COVID trials and this is a very crude mortality rate of Different races that missities over the eight not over the time period of like the dotted line being COVID onset and then seeing like you know relative to non-Hispanic white Okay, what the different what the different mortality rates are and you can see above one Okay, that Non-Hispanic black had this huge spike at the onset of COVID Okay, it stays relatively elevated okay, it was elevated before but it's only it's still a little bit higher than what it was in pre-COVID levels before and then for non-Hispanic Indian or Alaska native Okay, then what you find is like a spike that hasn't really returned to a pre-COVID baseline So this was a paper that came out early in the pandemic like summer of summer 2020 still and it looked at like a It was a retrospective cohort study that was centered in the auctioner help in Louisiana It described COVID positive subjects were hospitalized All right, and then it looked at different models for mortality And it found no association in several models, okay, like The model one was the unadjusted model but model two and Model three it actually found like higher rates of hospitalization among patients who tested positive for COVID So black race was associated with increased hospitalization However in patients hospitalized for COVID in all of their models they found no association between race and mortality So what they concluded from this was that black race was not associated with higher and hospital mortality than white race Which was true as far as their study was concerned This was a jama paper which came out it looked at a Similar study in Detroit and looked at people presenting the Henry Ford Health Center And the entire cohort was COVID positive patients Presenting to care. So think about what you had to do to get into these last two cohorts. You had to Test for COVID and you had to test positive for COVID and then you had to like actually present for care So in the subgroup of patients who are admitted they then did okay They looked at in the entire cohort they looked at ICU admission for the hospital score and then mechanical ventilation for the entire cohort All right And so black race in this in this study was not associated Okay, with need for ICU or ICU admission And black race was also not associated with need for mechanical ventilation And so they concluded that Black I'm sorry they didn't show a graph of this but in in text they showed that like black race was also not associated with mortality All right, so they also concluded at the end of this study that it was not as raves was not associated with with ICU admission need for invasive mechanical ventilation And then later on mortality they found other risk factors which were associated with those things And lastly, this is another jamma study okay, which was looked at multiple hospitals over 12 states And they specifically aimed to test the association between race and mortality. It's straight up what they said they were wanted to do And in all of their models okay of hospitalized patients they did not find an association between race and mortality If ending black race was the odds ratio churned it toward protective if anything And so they also concluded okay that there's no association between race and mortality so What do we what are we seeing here? Okay, we're seeing race is not associated with worse outcomes and hospitalized patients however This is the percent of block that in these three studies 71 percent of anguiziana 70 percent Detroit and 37 percent multi-center massively disproportionate to the actual demographic representation of the black race in these three communities including the multi-center multi-state one So that requires us to kind of look at this Take a step back from a more causal framework. So causal framework and directed acyclic graph So like kind of like they've been around in the epi literature for a long time They've only recently made it Into the biomedical literature as a more common way of talking about things like this And so causal inference is not prediction and whereas a predictive model may Prioritize generalizability discrimination calibration areas under curve And you need like a derivation cohort and a validation cohort And like a lot of people publish their derivation cohort which is massively overfit has areas and there's perva point nine and greater and never reproduces anywhere else We all got to eat we all got to publish I get that This prioritizes useful variables that improve prediction that is that is fundamentally what it's trying to do because so if you're variable Like whether or not it's necessarily significant and if you're like you know based on your power or the sample size that you have Although they very typically these variables will be significant out of p less than 0.5 The point you're actually trying to do with the predictive model is predict So if your variable improves prediction It's actually irrelevant actually what it's what it's Individual statistical p-value and not model is By contrast causal models require a causal structure you have to you are effectively making a causal argument of this causes this And we'll go through a few examples of how that like and that kind ends up Requiring a different construct to think about what else you put in the model because that affects what question you actually are asking And it's important to answer the question you think you're asking So this is association between copy and death okay, and let's say That this has an odds ratio of two Which is frustrating because I drink approximately 120 ounces of coffee a day So if you ran an underjusted model of this okay, this is this could be the answer you end up with Unfortunately, okay Coffee drinking is co-linear with smoking and so patients who smoke tend to drink coffee with it all right We'll just go with this for now like this mirror I'm not making a social epidemiologic argument about what people do when they smoke but If this were the case and you actually adjust for smoking then this odds ratio goes away So this is what's called the confounder and you should adjust for this because this is compounding and so Because it's green okay like you should adjust for it in my slideshow and I apologize if you read green colorblind because that's going to come up in this lead so If you have an exposure and an outcome okay Losely you can think of three relationships that another variable can have Okay, with the exposure in the outcome one of them is a confounder and because it's green you should adjust for this That is to say it's something which is upstream and affects the exposure and affects the outcome Smoking affects your probability of drink coffee smoking affects your mortality So you should adjust for this if you are looking for the association between coffee and mortality however Variables can have other relationships with your exposure and outcome one of them could be as mediator So mediator something which is globally called on the calls pathway So you should not adjust for let's say aridmia If you were looking for the association between coffee and mortality because plausibly Coffee could cause mortality via aridmia so you should not adjust for that if you did that if you actually adjusted for aridmia Then what you have done is you have introduced Okay, a mediator and you have inappropriately you've adjusted away the association between coffee and mortality And finally the last thing and this is a harder one to wrap your head around is a collider a collider is something which is on the causal pathway for both the exposure and the outcome or is the descendant of them and we'll talk about what that means in like in in a second but This is an example this is a graphical example of collider On the excesses is attractiveness on the y-axis is acting ability of the general population and the entire general population There's no relationship thankfully between attractiveness and acting ability. I don't even know what scale attractiveness is on we're just gonna go with it Okay, it's on the x-axis unit list All right, I guess acting ability for that matter too But if you condition on successful Hollywood actors if you actually Ask the question of successful Hollywood actors who are in red you find a negative artificial biased relationship Okay, between acting ability and attractiveness which in this case happens to be negative All right, so if you were to ask this question of successful Hollywood actors This might be an accurate dog's ratio if you were to extrapolate that or if you were to ask this question and say there is a negative association between Acting ability and attractiveness you would be you'd be lying. This is a this is a misleading inference because you have conditioned on a subset Okay That's associated with both the exposure and the outcome So Graphically this looks like this for ratio for race studies in the in the covid studies This is what's happening Black race has some non-zero association with that that odds ratio is something okay But if you condition on hospitalization and by selecting patients who are hospitalized or even test positive for covid This is what you've done All right, and if black races also associated with a greater severity of illness of presentations Substituted those patients are more likely to like actually be hospitalized Okay, then what you have done is you conditioned on the collider of hospitalization And so that will give you a biased in this case negative answer of the association between black race and death This isn't just in the covid studies this happens in biomarker studies that we talked about so a lot of my work is actually like You know risk prediction for a rds and sub phenotyping a rds and subsis and so like this is a prognostic model that we came up with using biomarkers that predicted rds mortality and using classification regression tree Using four by three biomarkers in age. We're able to identify a low medium and high risk of mortality nodes. Okay But Assuming adult rds studies have like have had a lot of literature Okay, which is like some of it's been done next door Okay, in which People have taken like a re-analysis of the artisanate work trials and they've looked at The biomarker biobanks that they've had and they've re-analyzed whether or not they can identify subtypes within Those trials and whether or not nominally negative trials for the entire cohort may have been positive Within a subtype and they found consistently two subtypes of a rds Okay, locally called hypodephyperinplamatory And in a observational biobank okay like people have identified like different rates of consent. Okay for biobanking So now these two studies in which we look at like bio banks are now biased by who actually participates in these studies So this is a pediatric biobank in this example, but this is true of adult biobank studies as well In which like none in this case the odds ratios are greater than one because they're showing the The odds were declining consent for participation in the biobank so if we know that Different races have different rates of consenting for observational and clinical trials Okay, and we do and we'll get to more of that data Then the higher rates of declining consent then end up biasing your cohort so the inference is you're making about the subtypes of rds or the mortality prediction of rds or subsets are then now conditioned on who consented for your study So graphically that ends up looking like this So you're now conditioning on consent so any association between subtype and therapy response Like you know corticosteroids only work and I use a lot of steroids so like steroids come up in a lot of my examples for some reason but If there's some association between like steroids and and a therapeutic response, but it's only active in one subtype so that subtype has an association between With with a therapeutic response. Well your study is conditioned on consent And so you've now introduced a selection bias. Okay, which is a little easier to wrap your head around than a collateral bias But you've introduced a selection bias for who consents and that ends up impacting the Generalizability of your of your study And that's true of trials as well not just observational biomarker biobics. Okay, so this is the effect of dexamethasone on invasive mechanical ventilation option only and no option received in one the earlier dexamethasone versus Ceva trials that came out of COVID and so you on balance and the entire population dexamethasone was beneficial And it seemed to have a stronger effect size Okay, that is to say a lower in this case risk ratio of poor mortality for patients who were sicker Okay, like option therapy you add a little bit better odds ratio or a risk ratio and then if you're invasive in mechanical ventilated you add a better risk ratio but COVID trials enrolled 50% of the patients Relative to their racial representation in the ICU's of black and Hispanic subjects and if we know that black and Hispanic subjects are more likely to be six Then we actually don't know The actual effect of dexamethasone on mortality Relative to the actual patients that we use in the ICU even though those results from those trials were likely generalized to every patient in the ICU Okay, and so the effects as could be stronger Because like if those patients representing sicker they would have made up a larger proportion of the people who are in basically a mechanical Ventilator or on oxygen or it could be different. It could be it could be less. It could be attenuated You don't actually know and so the built-in selection bias of who's consenting for your studies Okay, for it's a problem if it's non-representive uh Enrollment So what do I conclude from this so collateral bias and study design affects conclusions and this is unfortunately a design problem This is like you can't fix this with stats as I'm trying to show like no amount of math will fix Is the fact that you conditioned on a quieter or you have inherent selection bias and how do you picture co-work? So this is a design problem you have to ask her at this design stage of your study Whether you've accidentally introduced a variable or conditioned on a variable That is actually introducing bias and leading gunnaweed you to a misleading inference And causal model and a causal modeling in general requires a causal structure Like you can't you can't throw everything in the model just to control for everything because that's not what's gonna end up happening You're gonna end up throwing in mediators and and moderators and colliders and that's that's gonna end up giving you a biased answer Particularly if you're looking for a causal structure and predictive models. You're doing something else. So good And your p-lesson-05 is more likely that your model is mystified Then something real and even if that's not true like it's good It's a good heuristic to kind of just think as if it's true because then you now have to prove it And now you have to convince me and reviewer twos like me That that's not true. Okay, and so like in general This is a good way of being like a little bit skeptical of your own data So conditioning is not always wrong. Sometimes the conditional effect is actually what you care about Okay, like so these are some examples of conditional questions that that could be interesting Given that a subject is hospitalized What is the effect of race on mortality is there persistent racism once or actually hospitalized? That's actually a closer to what these studies were asking Given that a subject survives 14 days on ethanol. Okay, what is now on day 15 your probability of mortality Those are conditional questions you're conditioning on surviving and asking so this the sounds more intuitive This is actually like fairly intuitive and qi work also This is this is the qi equivalent of people asking what's the correct denominator This does require clarity of your aims and presentation So finally some I want to talk I want to spend the rest of the time talking about some research or chop that we're doing Uh, that we're pretty excited about so pick you disparities then has like kind of a built-in problem that we're gonna talk about So there are well described disparities in outcomes. Where's outcomes for non-wide publicly-insured subjects Okay, that's been described in multiple diagnoses and adults and pediatrics and We've already decided that research conclusions depend on your study design and that collider biases pervasive cohort selection pervasive confounder selection is super important and The fragmentation of US health care system does lend itself to Trying to ask some of these questions in a geographic area single center studies may not generalize and that's okay Because the way racism operates and Philadelphia or Boston isn't necessarily the way it operates and like Texas for Louisiana So disparities have been well described in adult ICU of structures and NICU hasn't really been studied and picked you but One of my residents came to me and she's now a fellow at In England doing pediatric ICU, but she came to me and she wanted to like a disparity studies in the PICU and like we talked about this and One of the problems we ran into immediately is like you're gonna run into this exact issue You're gonna run into the fact that if there's some association between black race and mortality Okay, that if you condition on picky Then that's gonna be a problem because you're gonna have a built-in collider So we can't do that so we didn't do that so instead we looked at pediatric sepsis in a Hospitalized sepsis database and we figured out like okay if there is racism in the effect of black race and mortality Then Where would it be operating would it be operating at the healthcare system the community or in the PICU and we said it's okay It's possible that there's residual racism once you like get to the PICU maybe different people get to from therapies different but It seemed more possible that there was different diffuscument differential recognition of sepsis either in the community pediatricians Or on the floor that could lead to different rates of transfer to the ICU or different rates of severity of illness when they present and that carries over So we looked at the entirety of Using the kit 2016 database which was like a nationally representative database pediatric hospitalizations We looked at the association between race and mortality and we found increased odds of death for black race as well as like a Lower probability of being discharged alive. This is essentially hospital three days, but as a as a hazard ratio When we split it up Geographically kid has universal or near universal representation of the hospital system Primarily because it de-identifies the hospital so you can't actually tell which like you know this hospital in Massachusetts in Boston has the worst mortality like you can't actually do that with kids And so like the geographic regions are given grossly is northeast Midwest south and west That's one of the ways that it protects the hospitals and by protecting the hospital is they like get higher rates of hospital participation So that's the trade off So in the earlier versions of kid you could actually like they had more granular geographic data and you could actually tell like oh, yeah Man my hospital sucks and then hospitals stop participating and so they lost participation So they went through some measures to try to protect hospitals from figuring that out or at least investigators from figuring that out uh the effect of The association between black race and mortality is most pronounced in the south and the west and so this is geographically driven We did a similar study using kid 2016 and 2019 for bronchialitis and so this was done by like some mantis mine Gerkheim who gave him a faculty at shop and Stephanie want to make her okay, who's applying for picky right now Okay, we also found an association between black race and mortality which didn't reach statistical significant in the main overall multi-variable model We also found it for other uh for other race okay in addition to black race of higher rates and mortality Okay, but in the south Okay, we did end up finding okay like that that bronchialitis like something with like a less than one percent mortality rate with a large enough data set that It was actually powered to find this okay, we couldn't find in the overall cohort but in the south there was a statistically significant increase rate of mortality for black and other race subjects Okay, suggesting that some of these disparities even in a Absolutely survivable diagnosis like bronchialitis still has detectable effects of race We did this with pneumonia using the fizz database. Okay, that prior study was also done in like in kid And so that's also hospitalized, you know, okay, this one fizz is also hospitalized So like none of these things are ICU specific studies Although you can imagine the most these patients are dying in the ICU But in none of these studies did we actually want a condition on being in the pick you So this was done by Garrett Kym and Cody Gathers. Cody was a resident here in pediatric. He's not like a Pick you fellow. He's not doing a cardiac ICU year Okay, and so we by this time we knew we knew that geography and age were major Like players in this risk of mortality in a lot of these diagnoses because like that was consistently coming up as a as a signal as well Geographies come up in all studies that I've shown you so far And so we had a large enough data set now that we after created a a four-part variable Okay, which is a joint exposure variable where we referenced everything off of white Less than one year from the northeast so like a baby in Boston a white baby in Boston It's like a reference population. It's your reference baby and every other Exposure okay is built off of that and we had a large enough data set to do that like to me This is actually like a cleaner way of doing this than looking at a bunch of interaction terms and things like that Most data sets just started large enough. We were fortunate to have one that's large enough to do this And what we found is actually that the largest impact Okay of Of mortality was actually on age so the patients less than one year We're actually the ones that were worse off and in that we actually did find differences in race as well Okay, in which we're finding differences in like the south and the northeast Okay for black okay in the in the northeast were Hispanic less than one year, but the major effect was very much that of age Closer to home we started asking what are some of the ways that racism could be playing into the actual pick you like okay Like you know a lot of these diagnoses pneumonia broncholitis substance They happen and the the worst of those kids and up in the ICU and that's probably where the majority of the mortality occurs But what about actual racism that's actually happening in the pique is that is that something we've matched your study So there I kind of thought about the COVID trials and like and how representation in trials and cohorts And why somebody choose to participate or not participate in research was an area in which you could actually potentially have some effective race On the outcomes that you care about within the picky constrained within the pique And the patient experience time at that side use of interpreters and experiences of racism That's a that's a relatively like understudy area within the ICU So this was a study which came out um like I forgot this was 2022 or 2023 and it looked at the association in uh Race with research consent across all pediatric studies or pediatric clinical trials over a decade And what it found is that relative to the US population Black race was overrepresented In clinical trials But this sophisticated audience by now and the number of times I've shown you this Okay, like has figured out the problem with that study Okay, because if you have an association between race and research you have to be eligible for the trial Which means that if you're more medicalized Related to the population if blackheads are overrepresented in your population base and are in your trial eligibility Okay, then they're going to be overrepresented relative population So This is to me a misleading inference from a study like this that the inclusive research practices that we do with black subjects This is the conclusion of the authors that the inclusive research practices Okay, it could be extended to other historically and currently just in franchise racial ethnic groups now with suggestions that most of things We have done to black patients. We should not extra to other racial and ethnic groups So So how do we get around this? So one of the ways that we thought about getting around this is like okay instead of referencing the The the relative rates of race versus the US population or the local population Like we should look at like who's eligible for a study and so a shop in the PQ we had detailed Screening consent logs for the last decade and this is primarily because the type A personality of most of people in my PQ Okay, my colleagues are all sorts of anal about why didn't you approach that kid? Okay, and so the research coordinators over a decade have been unfortunately like traumatizing condition to write down I approach them. I promise they weren't eligible for this this reason And so I have really detailed screening and consent logs So we used it so our inclusion criteria for this study was whether or not you were eligible for a study And we excluded you if we fundamentally could not figure out whether you were eligible or not for a study And so one of my residents who's not a fellow at Seattle children doing pediatric ICU looked at the association between four different exposures race preferred language religion and social deprivation index which we based off the zip code We added a bunch of confounders and we looked at three outcomes. We have did we approach you? Did we consent you among all eligible subjects and then did we consent you if we approach the final After looking at 35,000 screening screening entries we ended up with a final study population of a little above 3000 modest degree of missingness Okay, and these are the answers. Okay, so among approach Oh, there was supposed to be a circle but circles non so I'm gonna circle it manually all right among approach all non-white subjects all non-english languages Muslim religion and a small but significant social deprivation, we're so so deprivation index we're associated with lower odds of approach referencing consent rates among everybody who's eligible black race other race Arabic other language and Muslim had lower rates of consent among all eligible subjects and keep in mind This is the exact opposite conclusion of the prior study I showed you which showed black race was overrepresented But this is not conditioning on the US population. This is conditioning on whose eligible for the study And then if you look at the consent it approached Almost every disparity went away. Okay, black race remained significant to have lower rates of a per uh lower rates of consent Even if approached other language had lower rates of consent even if approach although this didn't reach statistical significance with very white confidence intervals probably because we didn't have that many other language And then Jewish religion remained significant for lower rates of consent if approached An informal mediation analysis. There's my circles informal mediation analysis black race had almost a 10% lower rate of approach But that mediated half 51% half of the association between lower rates of consent for black race Okay, if you approached black subjects Okay, you increased your odds of consent Seems obvious But now we put a number on it So the reasons for why different races were not available that side for non-Hispanic for Hispanic race and for um Non-Hispanic black was family unavailability was somewhat overrepresented for Hispanic There was also a perceived language barrier So for this study if you had In your eligibility criteria that we're only approaching English subjects then these patients weren't included That was like you were you were considered ineligible for the study because by definition you weren't eligible However, if your study IRB did not say that then these patients Were eligible and there was a choice somewhere along the way like either a choice on the part of the investigator to not Figure out how to include eligible subjects There was a choice by the research coordinator team about like Operationizing interpreter services or the availability of interpreter services But this actually was an effect that we wanted to capture and so we found perceived language barrier to be persistent for other race and for Hispanic Unsurprisingly for other languages for every non-English language perceived language barrier was a dominant reason For why they weren't approached either like and these were now patients were eligible like there somebody has said in their IRB like you can approach every every Every every language okay. We have a short form consent process. We have interpreter services Make it happen But we still ran into the actual logistical barriers of getting somebody within the enrollment window Okay, of getting interpreter in time finding an in person like somebody said like in person consent only and so you had to find in person interpreter things like that is like what we ended up finding all right So perceived your actual language barrier who ended up still being a pretty high reason for lack of for lack of approach So what do we conclude from this so we concluded that many eligible children are not approached and there are differences according to race language and religion But that these differences in consent rate are actually attenuated if you get around approaching them There are some demographics in which this remains persistently low So black children half of the association between racing consent is mediated by uh a lower approach race and To the extent that People aren't consenting Because they don't want to that's not that's not a problem To solve like you can do whatever you want. It's your kid like I'm not gonna you know, I'm not gonna bring it balls about that To the extent that like your answer of not participating in research is a reflection of mistrust in the medical system of be personally Okay, my research coordinators how we approach you when we approach you that is a problem to be solved And so this gives us some clues as to some areas to like actually delve into We have ongoing work on disparities um one of the one of them is a decline and decline is split up into a quantitative and qualitative portion The decline is determined caregiver logic and saying no to research which I have to say is one of my better acronyms um, so mixed method study At the point of every subject who is approached for research in the pick you if they decline One or more one or all studies. Okay, whether you can decline you can say yes to three of them. Let's say no to one of them um Our research coordinators are trained to ask we're always trying to get better at this Can you let us just know why you said no to this study or these studies? Okay, and then we have one of like you know 15 pre-filled answers and they check it off or put it in the other category and then Recently we introduced a A follow-up interview we then asked Okay, is it okay if somebody comes back and talks to you in a couple days not right now not while you're Clearly saying no to me about things, but in a couple days Is it okay if somebody comes back to you and talks to you about your reasons for saying yes or no to research Okay, most of us Like we should leave ask like probably like like 10 families that okay like we've we've been doing the quantitative portion for about a year and a half And we have like nearly 200 subjects but like we only recently trained up our coordinator to start asking for the qualitative portion And so we just rolled that out and like the the uh The no rates on that follow-up interview are quite high as you can imagine if you're already saying no to me You're probably going to keep saying no to me But we have a couple patients who say yes, and then we actually do think that the qualitative is actually like a pretty good way to tackle this part of things Because the granularity that you can get from the interviews is probably higher than you can get from pre-check boxes as to why you said no in the moment Too overwhelmed and no more procedures were the dominance signal that easily the dominance signal in the in the pre-limin analysis The first 170 patients We have ongoing work on patient experiences which is also a mixed method study And this is the point prevalence of parental presence in the piqui or the five p study This is being done by Michelle Grog check who's one of our fellows who's graduating this year looking for jobs This is a direct observation of percent time at bedside over 48 hour period and then from there in the in the first like um in the first like a quarter of patient she enrolled We identified the top quartile and the bottom quartile and then we performed interviews Based on which quartile you're in So this is what the study structure looks like so children have to be in the piqui for 48 hours and anticipated to be in the piqui for at least another 48 hours and then We recorded the caregiver time at bedside how would you do that you may ask we ask Our family research council our family research council is a group of parents who used to have kids in the piqui And we said we want to do this But we don't want to do it with consent because if we ask parents for consent to do this Okay, then we're only going to get the parents who are like a at bedside in the first ways and the Like not gonna be not gonna be razzed by somebody asking them how much time they're spending at bedside and feel bad about it So we need an unbiased way to do this they said You can have the nurses do it If they get caught Immediately make it clear what you're doing and why and make one of yourselves available We told the rb we would do that and the rb said cool So that's how we did we had nurse confederates who are like actually recording time at bedside surreptitiously To get an accurate sense and unbiased sense without consent of how this this rattled and I actually think that the ethics of how we actually got there and the number of meetings We had with different representatives and stakeholders to make sure that we weren't creating a new problem Okay, it's actually like a really important thing in disparities research Like you don't want to create a new avenue for mistrust Okay, while you're trying to solve something else and I think that that has the potential to happen and research like this So caregiver time was recorded at bedside accurately And then in the we look at the media and the variance of that side presence in the quantitative analysis and Then in we looked at the association between specific exposures and outcome and we'll get to that part of the analysis And then we recruited caregivers with high and low Okay, time at bedside to like kind of talked about like What their opinions were about their time at bedside And that was our qualitative analysis So the quantitative analysis is set out very similarly to my other studies because like I run on like you know Pretty much one speed and I do the same take over and over but different exposures Regression confounders selected mccasal framework Okay, looking at the outcome of percent percent bedside presence It's actually I put logistic regression. That's yeah But it might end up being actually like a quantile regression Uh, time at bedside Okay, I had a very bi-modal distribution patient for either 0 to 10 or 90 to 100 Formally the median was 82% out over 48 hours of observation And the interquartile range okay was 17 to 99 or 8 to 48 hours median of around 40 hours So 48 to 48 in a 48 hour period Black race and multi variable analysis was associated with 17 36% lower or 17 hours lower time at bedside and public insurance was associated with 32% or 15 hours less time at bedside So So the qualitative study using a social ecological model we conducted 18 interviews with 20 caregivers Okay, eight were high eight were low and two were in the middle So this is some of the data that we're getting from this and this isn't the middle of being written up and analyzed Okay, so right now we we finished the quantitative portion now We're like like just kind of getting into the granularity of the qualitative portion putting that story together so At the individual level we found a lot of fear of the unknown when I'm away from the hospital I become so tense. I need to know how she's doing and everything The thing that I was the thing was the thing is that I was really tired My legs were telling me and my sister told me to go home and rest, but I was very worried Very good for obligation It's important all the time to be with her because I'm her mom so I need to be sure that she's fine All right, so this is at the individual or this is at the parent level at the interpersonal level. There's now like relationships I think it helped with the with the patient themselves I think it helped with the zoroah growth and progress to give him hope It helps him push through to show him that he could get support at home improving child's care He needs me to be there because I know exactly what he needs. I get a form quickly I'm totally paying attention all time Parents don't work as better than the doctors every week of new physician things get lost They're starting to bleed into mistrust All right communication the trust level. I have with a particular doctor or team is tied to their willingness to communicate with us Take time to answer questions give explanation for you to assist people when I when my trust level is a little bit lower I'm inclined to stay here longer because I feel I have to advocate for more strongly for him This is now a parent who would have gone home and taken it now, but she doesn't trust us So she feels like she can't Organization Job flexibility this this this and the being a pretty important thing as you can imagine I work remotely so that's the biggest thing my body just said I don't care spend as much time out of the offices you need to this some of the stuff that we saw during COVID as to what kind of jobs Were able to be done remotely versus not work. That's the only thing keeping me from being able to come up here. I try to come up on the weekends Community support structure. We co-own a two-family home. We live on one side my mom lives on the other They kind of hold things down at home. Okay resource normally Mirror my wife is staying in the hospital. So we call each other and she's going back home She's calling me and asking what her status is. I do the same thing two parent house Public policy FM LA with him being so sick I ran out of all my FM LA I had to go back to work accumulate more hours She had FM LA so she was there for a good while, but she did have to go back eventually So yeah for that short while that she had FM LA she was there every day I would be with my kid 24-7 if I could I just couldn't because of this thing that we all voted on and policies that you know, we don't really think about when we vote very much had an impact on my life at that moment So all these things okay like individual fears. Okay fears with relationship between the parents and the health care team the parent and their their interpersonal relationship with their child And how they felt that that was an important component of them being up that side their job flexibility their support structure and You know quasi legal structures like FM LA and policies public policies I guess is probably the most accurate way to say that all right these were the themes that came out So we conclude from this the black and public literature children spend less time at best side And that caregivers perceived the benefit of being present because they think that they're improving their child's experience or their child's actual care And they're probably right Okay, they often feel their advocacy is required for optimal care Which is a little bit of us airing our dirty laundry, but I suspect it's not unique to our institution And Facilitators of presence include job flexibility and family availability and generally just kind of support just like either You know organizational or institutional or like you know like community level support For communication promotes mistrust and anxiety about leaving the bedside so even parents who would go home and take a nap Don't always feel like they can because they feel like we're gonna mess up Next we want to do this on a larger scale and see if there's an association between time at bedside and more patient center Documents like either a delirium time in the ICU experience in the ICU Wrapping up now just a couple more studies We also thought one of the areas where we might find racial disparities within the pick you was in DNR status and so like And adults there's been some work in oncology There's been some more suggesting that there's differences according to race ethnicity and language for obtaining and acting on DNR versus not And so we looked at the co-primer exposure is a race ethnicity and preferred language and we looked at two outcomes probability of DNR and time to DNR And what we found jumping straight to results is that every non-English language had a higher odds of having a DNR Okay, and Black race had lower odds of having a DNR This adjusted for everything you can imagine Okay, the standard severity of illness and time in the ICU sort of confunders In time to DNR The data was a little bit less consistent We found that other non-English languages had a long a faster time to getting a DNR for those who had a DNR They had a faster time to getting a DNR and if your language was missing then you had a shorter you had a shorter time to getting a DNR Non-Hispanic black in this analysis also had for those black children who had a DNR they had a longer time to a DNR So it was a little bit unclear where exactly the higher the association between non-English language and higher rates of DNR were We thought it could be one of two things one as a referral center a lot of non-English non-English language patients are referrals And so they even though you are adjusting for severity of illness course You may not be fully capturing just how bad off there And so you're dealing with like a much more refractory population that has higher odds of death The other thing we thought could be going on is that a lot of times DNR discussions are a little bit tailored and and with English language Patients you may not introduce everything right away Okay, you may it may be more of a Conversation see where the parents are at see where you're at and something which you may See how they're doing at 6am 9am noon 3pm now you're talking a little bit more about like this isn't going to go well We should talk about the next steps that you you have the opportunity to do that a few times One of the things we're curious about is whether the pressure of having an interpreter Makes you feel like you need to get everything answered at that visit okay at that discussion at that encounter with the parent And so we're wondering whether that could be leading to Uh higher rates of DNR as well as like faster DNR like you just feel you need to get there Whereas like you may give the natural flow of this conversation a little bit more time And so this is an error where we're still investigating the association between black race and like lower rates of DNR and longer time to DNR That is consistent with like what's been seen in adults and in oncology Finally, we're doing some ongoing work on the patient experiences in the uh in the In the PICU and so we're doing a qualitative semi-structured interviews of black children in the PICU and just straight up asking them about experiences of racism And this is being carried forward by Vanessa Denny one of our current PICU fellows So My conclusion from this far the talk and like pretty much the whole talk at this point is like there We're trying to take a multimodal approach to disparities research a lot of large scale epidemiologic stuff some institutional stuff and we're at the level right now where we think actually personal stories and qualitative research Is actually like an appropriate step to gather the data to actually ask the right types of questions on a larger scale The folk the way I'm as at ARDSF's researcher I think that the way I'd like to merge these two portfolios is I spent a lot of time thinking about like about trial participation and and The way we're going to try to move this forward in the future is that given that anti-black racism is a major driver Forachimus in healthcare then representative research is necessary like I'm a scientist at heart and like and and the way I think about these things and the way I think about this problem and the way to solve this problem is going to be very much through the lens of like things I already know how to do and so this is how I'm choosing to interpret this area which I think is a problem So My overall aim is to improve the rates of participation of black children in research I have a lot of people think there's a PQD I committee all my mentees like yeah at this point like half of them are collaborators and they're not even mentees And they they I mean they come to my office but like but I Must say those days people know more than I do on both the stats as well as like the actual deeds of their study so Um with that I want to thank you for your time and I'll take any questions I Also really like this picture because my daughter truly truly hates what she's doing right now And she's still doing it. I just did it tracks you up so much Hey dear. Thank you so much for that fascinating talk Um, I have just one question and I'll see if anybody else has any questions the um A lot of those studies you showed you talk about black race One of the things that I wonder about is how socio-economic status plays into that And how you've considered that yeah um in your studies We get this question a lot actually so like so very often when I this is just a representative example of like how we set up a lot of these studies and We set up race and insurance type as and social deprivation index as separate exposure so Uncommonly do we put them in the same model and the reason for that is Annoyingly the arrows Here The arrows here start getting problematic If black race and the experience of black racism is partly mediated through Associate economic status indicator then there's an arrow that goes You can replace medicalization with like you know, there's an arrow that goes to socio-economic status So then I have to take you back to the slide where I said well that's in the causal pathway That's a that's a that's a mediator So I don't necessarily want to adjust for that if I'm looking for an association between black race and outcome This is not always how it's done. It's not even the only approach to this But I will say that if you if you do end up introducing uh socio-economic status as a variable you're asking a different question Then race and outcome okay, and adjusting for it doesn't necessarily take away the You flatten everybody's income or you flatten everybody's socio-economic status onto whatever scale you're using insurance or or You know social deprivation index. You are doing you are you are meaning everything you're setting everything to the mean version of that And then asking the question but to the extent that some fraction of The association between race and outcome is is mediated through That socio-economic status you've adjusted that away And so it is asking a different question that being said and the few times that reviewers have asked us to do this Okay, race ends up being a more dominant signature than then um Then the socio-economic status And that's been true of every study I've looked out so far The uh, stepsus the stepsus kid mortality study the um the uh Time at bedside study Okay, and like a reviewers really want us to do that one Okay, like actually say like okay, this said is this is socio-economic thing and like the qualitative portion of that kind of Lends itself more to socio-economic status and less toward racism but The context of a young white woman interviewing a parent Okay, it may also contextualize the kind of answers you're going to get from a study like that even on the qualitative portion There Thank you so much. I feel like I'm a little closer to understanding dags finally um I feel like this is obviously very very important research and I was want you alluded to this But I was wondering if you could talk about the ethics of Uh get doing a study on someone who not only doesn't get consent but actively doesn't give consent And then it's been your example of the bedside study or bedside time study Have you had that conversation where someone kind of catches the person doing it? How do you navigate that conversation? None of the nurses in that particular study like we're in the position That's a great question. We actually struggle with this lot the um none of the nurses in that particular study got Called out. Okay, so none of them were actually seen reporting time at bedside Okay, and it was it and we we had gone through a process of training them the In the ethics of this required us to not think that we knew what we were talking about and so So approaching it from a position of humility To go to like the family research council and ask to ask their like this is what we're trying to do this is why we think it's important Okay, this is the concern with doing it through Uh, this consent model All right, we we don't think we'd actually Answer the question we're asking if we went through this consent model How would you help us and then partnering with them and I think that I think that like shifting patients Research participating. This is harder in I think pediatrics. I think adults have this a little bit easier um in my opinion but even critically ill adults like to treat them as research partners Okay, in a way that like that that rather than subjects and so like In academic medical centers like onus to advance knowledge requires advancing like there's no sense in putting out a paper That is asking this question badly Okay, like that that that is that that to me would cause more harm and you come up with like a uh, I Like you'll end up with a study that says like you know parent time at that side is 80 to 100 percent at median 100 with IQR of 90 to 100 as like That's not helpful, dude, and so I wouldn't If you're gonna ask the question if you think this is a fundamentally valid Area study then ask it the right way and then how you engage the family around that is like okay, let's ask Patients who've been in this position like how would if this is what somebody was trying to do Do you think this is valid would you change it up? Would you change what we're trying to do or like would you just make us understand Help us understand how to set this up in such a way. So I think I think like partnering with um equivalent Structures like that or or organization like that is like the first step because the answer may not necessarily be the same in a different institution So I have one quick follow up and I'll go over here you um you use the term racism and I think There's a difference that you're detecting. How do you get to the Conclusion that it's racism that's causing these differences? That's a good question the um I feel I feel like a lot of DEI research like the there's there's middle large push and I support that that like um Racism is not solely interpersonal animus although that is certainly a part of it. Okay, like um a lot of the signal that we're seeing in like the south for example Is not necessarily That physicians there are treating black subjects worse Because of interpersonal racism however if your hospitals are worse if your support structures are worse if you're presenting Because of how your Community has been set up relative to healthcare delivery by design Okay, and so more systemic factors that we colloquially call systemic racism since 2020 I think that that's still importantly labeled racism and so the idea that like That you soften it or use a different term I actually think works against the goal racism is a jarring term But appropriate Like if if we had these kind of disparities in Women dying of A diagnosis that Was because of how Their access to healthcare is limited because of dobs Okay, like that is appropriately called misogyny even if those individual relationships between providers and the victims are not driven by personal misogyny Sort of as a follow-up question. Thank you for your presentation um, so what do you think in terms of what we can do next So in terms of whether it's changing perception changing resources Both locally within hospitals, but also nationally you're identifying a very important things That require a lot of work then to improve upon I think actually you know this this throwaway table Which is a supplement and like this paper is actually I think like like remarkably important I think actually like so like if if my and without I'm hope to God you're not asking how to solve racism and medicine but like yeah, but but My little piece of like of Shandana improves the participation of non-representer grace and is in research and children Um like I actually think the stable actually gives a lot of data um, I think to modifying our approach Is and and being more flexible with our with how we approach subject so like let's see you have a tight 12-hour Algibility window 12 hours is hard man like I mean that that that's just straight up difficult like if you truly truly need a 12-hour blood sample Okay, can you use a residual sample from the CBC that was invariably collected in the ED And is now hanging out in the fifth floor lab Okay, like after initial processing can you actually measure your biomarker in there and then can you then ask the parent like two days later? Is it cool if I use this? Okay, so you've maintained their integrity. They're right over their child and their child samples Okay, you've done your science. You've gotten a clean sample all right But you haven't limited it to like a highly sensitive highly emotionally charged time frame All right, but you've still been able to get in a cute Conval a cute um a cute sample that was necessary to answer your scientific question Does your does your enrollment window need to be in person? Does it like like all you're doing is taking spit like Can I can I really not just call you and ask you about that? Like is this is this not an exception from documentation of informed consent? The informed consent process not me getting a signature the informed consent process is me having a discussion With a parent who says like yes, that's cool That that's and do you understand what we're doing? Yes, I get it. That's cool All right, that's informed consent the documentation of that is necessary. It's legal But it should not hamper Like research at the expense of like creating non-representative cohorts and trials that we make extrapolations from Thank will take there's one more hand I Thank you so much for the talk. I was just curious in your decline study if you had considered also doing interviews of people who Said yes and finding out what their reasoning was afterwards, which might also be interesting Yeah, no, that's a great question. We uh we are going to do that. We want to get the nose first But um we actually think that the reasons why people say yes is it's like very much going to be informative different informative but like very much before we we thought by getting the um the qualitative no was we expected a lot of declines of decline and so We thought that it would actually be harder to enroll so we wanted to get that rolling and actually get our like um Uh if you if you've done qualitative research and like and and i'm there are a lot of people like here have like there's the there's a rhythm that I think that the interviewer has to get into before they get kind of smooth at it and so like they're We want we want to at least one phase to be a little bit smoother before we introduce that variable, but yes, that is something we're hoping to do Nadear, thank you so much for the wonderful talk It's been great to listen to all the work you're doing. Thank you so much. Oh thank you guys