Site Maintenance October 29 – November 5
SCCM is currently performing major updates to its digital infrastructure to better serve members. Any website that requires an SCCM Customer ID and password to log in is currently unavailable until November 5, 2024. SCCM Customer Service and staff have limited capabilities, and response times will be longer during this time.
Bloodstream infections (BSIs) acquired in the ICU are potentially preventable. Kyle B. Enfield, MD, FSHEA, FCCM, is joined by Sameer S. Kadri-Rodriguez, MD, MS, to discuss the article, “Epidemiology of ICU-Onset Bloodstream Infection: Prevalence, Pathogens, and Risk Factors Among 150,948 ICU Patients at 85 U.S. Hospitals” (Gouel-Cheron A, et al. Crit Care Med. 2022;50:1725-1736). Dr. Kadri-Rodriguez is a critical care and infectious diseases physician at the National Institutes of Health Clinical Center in Bethesda, Maryland. This podcast is sponsored by Sound Physicians.
*If you are unable to play the podcast please click here to download the file.
Transcript:
This podcast is sponsored by Sound Physicians, the employer of choice for critical care physicians, where we seek to transform acute episodes of care. At Sound Physicians, we ensure physicians have the time and resources needed to deliver compassionate care that measurably improves quality and lowers the cost of healthcare for patients in the communities we serve. For more information, please visit careers.soundphysicians.com.
Dr. Enfield: Hello and welcome to the Society of Critical Care Medicine Podcast. I’m your host, Dr. Kyle Enfield. Today, I’m speaking with Dr. Sameer Kadri-Rodriguez, who is a critical care and infectious disease physician at the NIH. Dr. Kadri-Rodriguez has been using large datasets to understand outcomes in the intensive care unit. Today, he is here speaking about the work performed by Dr. Aurelie Gouel-Cheron, titled “Epidemiology of ICU-Onset Bloodstream Infection: Incidence, Pathogens, and Risk Factors Among 150,948 ICU Patients at 85 U.S. Hospitals.” This article can be found online at ccmjournal.org. It provides new information on the impact of bacteremia that develops after a patient is admitted to the ICU and differences in antibiotic-resistant patterns in bacteremia that develops in the ICU as opposed to before ICU admission. Sameer, I’m really excited to sit down and talk with you today. Welcome to the podcast. Before we dig in, do you have any disclosures to report?
Dr. Kadri-Rodriguez: I have no financial disclosures, Kyle. But I will say that I do serve as associate editor for SCCM’s journal Critical Care Medicine.
Dr. Enfield: Thanks for that service as well, because I know that’s a huge job. Sameer, your research has used large datasets to explore important clinical questions. This body of research is often seen as hypothesis-generating. What do you see as the biggest benefit from using large datasets to perform research?
Dr. Kadri-Rodriguez: As all of us know, RCTs, or randomized clinical trials, are obviously the gold standard for traditional assessments of safety, efficacy of interventions, drugs. For example, is drug A better than drug B or placebo? Even the best observational study or emulated trial cannot match up to an RCT. Nonetheless, not everything can be adequately answered using randomized trials. There is a niche for observational studies in critical care. Patients in the ICU might be too sick to randomize or too sick to generalize to results from certain trials. There might not be a financial incentive for a trial. There might be a need to understand real-world care patterns. For example, how are my peers treating disease X, Y, or Z? We might be faced with a new disease and repurposed treatments, just like we saw recently with the pandemic. We might want to look at rare entities. For example, in my research world, that would be studying extensively drug-resistant or pandrug-resistant infections. This 30,000-foot view of diseases sometimes helps put things in perspective. This is the space, I think, where observational studies, and large databases in particular as the data source for these studies, can be invaluable and provide definitive information, not just hypothesis-generating signals.
Dr. Enfield: Along the same lines, what do you see as the challenges or obstacles with this kind of research?
Dr. Kadri-Rodriguez: Far from perfect. First, I think not all observational studies are the same. They span a wide range of quality. When using real-world data for these kind of studies, there are a variety of biases that can invalidate your results and take you to erroneous conclusions. Sometimes the investigator may not even realize that. These can range from confounding by indication in comparative effectiveness studies or ascertainment bias such as, can I really tell if patient A is sicker than patient B with the data tools that I have at my disposal? Billing codes are often used for this type of work. They’re far from perfect. They can often be highly insensitive. Also, two patients coded for, let’s say, the comorbidity diabetes may have very different degrees of comorbidity from their diabetes, and labeling them as having diabetes may put them at the same level of illness through your lens. I think the trick here is that this is all very messy and we need to find a method in this madness. I feel like it’s more of an art than a science. I think we have some experience in this, but we have a lot more to learn.
Another major challenge in using large databases for this type of research is that they’re inherently messy. Over the years, they have gotten more granular in terms of not only having billing and coding data and individual charges, but also having data on lab and physiological perturbations from flow sheets, etc., that are really important for critical care research. However, if these datasets are not carefully curated using both scientific and clinical logic, you’ll essentially end up with garbage in, garbage out.
Dr. Enfield: Taking all of that together, you have looked at a fairly large group of patients at 85 different hospitals in the United States. What were the key findings from your study?
Dr. Kadri-Rodriguez: We looked at about 150,000 ICU patients and our first question was, how frequently do these patients develop bloodstream infection? We had two major broad categories of patients. Patients who either came into the ICU with bloodstream infection, they may have been septic or sick with bloodstream infection that brought them to the ICU. Then, the other cohort of patients that the intensivists see, which is bloodstream infection that is acquired while in the ICU.
What we found was that, of the 150,000 ICU patients that we looked at, around 4% were admitted to the ICU with bloodstream infection, and a smaller number, about 1%, developed a bloodstream infection while in the ICU. However, this latter group, this 1% group of patients with ICU-onset bloodstream infections, they represent a unique and concerning group and potentially a target for both quality and modifiable interventions that can reduce the risk of this potentially avoidable entity. We found that this group in particular, the 1% group ICU-onset bloodstream infections, they represented also a group that had a unique pathogen profile. There were more patients with, let’s say, Pseudomonas, Acinetobacter, Enterococcus, Candida, coag-negative Staph in the bloodstream compared to those patients who are coming to the ICU with bloodstream infection.
These patients, the ICU-onset bloodstream infections, the 1% group, they also had a greater burden of antibiotic resistance in the pathogens compared to the bloodstream infections that brought these patients to the ICU. ICU-onset bloodstream infections were also found to have comparatively greater morbidity and mortality burden, which is not surprising because these are patients who are sick to begin with. Then, I think potentially the ICU-onset bloodstream infection makes them sicker and they’re already in the ICU because they’re sick for some other reason. So, I think the two together potentially make them a lot sicker.
I think the bottom line here is that, even though this represents a small group of patients, because of the potential avoidable nature of some of these infections, I feel like it’s our duty as their providers to do everything in our power to minimize these avoidable risks. This led to the second part of our study, which was to identify potential risk factors that might increase their chances of acquiring bloodstream infection in the ICU. This was the hypothesis-generating component of our study. Our risk factor model provides signals, not necessarily definitive, so take it with a grain of salt, but we found that younger patients, male patients, patients who are Black and Hispanic, having greater comorbidity burden, patients who are admitted to the ICU with sepsis, trauma, acute pulmonary, and gastrointestinal presentations, as well as those who had pre-ICU exposure to antibacterials and antifungals, were associated with a greater ICU-onset bloodstream infection risk after adjusted analysis.
Now, because we had a multihospital cohort, 85 hospitals, we were able to also look at some center-level factors that resulted in a higher risk of acquiring ICU-onset bloodstream infection. We found that, compared to medical ICUs and surgical ICUs, the mixed ICUs or the undifferentiated ICUs tended to have greater risk. Urban hospitals tend to have greater risk. Among rural hospitals, the smaller rural hospitals tend to have greater ICU-onset bloodstream infection risk. We also found that the associated risk of acquiring ICU-onset bloodstream infection manifested with any duration of mechanical ventilation.
However, with the other procedures like central line insertions and arterial catheter insertions, that risk only manifested about seven days from insertion. I think the bottom line is that we found many factors, and I think the next step would be to take a deeper dive to understand the independent relationship of these factors in other datasets and other data sources to see if we can validate some of these. If they are validated, then they may become targets for interventions, particularly if they’re avoidable.
Dr. Enfield: That’s fascinating. I know, as an intensivist, this has been information that we’ve all wanted to think about, what antibiotics we should start when the patient is coming straight from the ER or the floor versus the patient that’s been in our ICU for a while. But I wanted to know, where did your interest in this project start at?
Dr. Kadri-Rodriguez: Sure. For many of us who practice in the ICU, we live by the mantra every day of: do no harm in the ICU. That’s the most important thing above everything else. These unacceptable occurrences like ICU-onset bloodstream infection, I think cleverly coined by somebody as “never events,” should just not occur. We should do everything in our power to avoid them, as I had mentioned earlier.
Let’s say, for example, you did a week in an ICU that had 20 beds and it was full for the entire week. You save some people, some were unsalvageable when they came to you. But occasionally, there’s that patient who picks up a CLABSI or a ventilator-associated pneumonia or a catheter-associated UTI and a bloodstream infection, and they get really sick from it in the ICU, and that’s what tips them over and they die. Now, that one bites a lot. Could we have done something differently to avoid this? This sentiment was the inspiration for me, particularly for this study, to get a step closer to being able to predict and subsequently avoid these untoward consequences of our own care.
Dr. Enfield: That’s, I think, where a lot of us come to from this. I really appreciate you guys taking the time to think about this and find a way to answer some of these question. When I was reading your article, I was reminiscing about an article written by Anthony Harris back in 2015 that really tried to think about using hospital-onset bacteremia as a new quality metric. It makes some fairly compelling arguments about the benefits of using that as a quality metric. I wonder, as you think about your research, how would you respond to that, and what was the most interesting finding to you?
Dr. Kadri-Rodriguez: Sure. We were certainly very motivated by the work of Anthony Harris, Clare Rock, and colleagues. Their study in infection control hospital epidemiology showcasing hospital-onset bloodstream infection as a potential better-quality metric than, for instance, CLABSI, it was foundationally novel and was certainly what kicked off the thoughts that we had to do our study.
In brief, surveillance is carried out for hospital-acquired infections, such as CLABSI. But then, there is an inherent subjectivity—right?—in their adjudication. Nested in that subjectivity is sometimes a suspected tendency to underreport, game the system, wear a blindfold. It’s hard to pinpoint that, but I think there’s a possibility of that occurring and it’s difficult to control that. To overcome this, Dr. Harris and colleagues opted for using hospital-onset BSI instead of central line-associated BSI to take away that subjectivity. We borrowed that concept to the ICU and essentially used that to say, instead of trying to identify predictors of CLABSI, where CLABSI itself might be grossly underreported, why not identify the prevalence, characteristics, and risk factors for the more objective entity, ICU-onset bloodstream infection? Because that’s pretty much black and white. You either have it or you don’t. In that sense, I feel like it might be a more objective quality metric for the ICU down the road.
Dr. Enfield: I would agree that ICU-onset bacteremia is a really nice, fairly black-and-white metric to use and one of the ones that we’ve used at UVA as well to think about the impact of arterial lines and how they might influence bloodstream infections. But all of that research really does rely on using these large datasets. For years, you used the Cerner dataset. I wondered if you might mention to the audience why that dataset was useful to you in doing this research.
Dr. Kadri-Rodriguez: Absolutely. The Cerner dataset comes from Cerner, the company, and essentially, it is a de-identified repository of data from multiple hospitals that have been powered by Cerner for their EHR. It not only includes billing and coding data, but it also includes data on medication orders, laboratory results, microbiology, and antibiotic susceptibility data that were critical for the study that we just did, as well as a subset of hospitals’ physiological data like vitals and vent parameters. It is really very rich and granular, which is critical for doing studies on ICU patients.
In our study, we not only needed information on who had a bloodstream infection, but we also needed the species, the resistance profile, and information such as, was this patient on a ventilator? Did they have a central line? Did they have an arterial line? Did they get a blood transfusion? The linkage of the different domains, such as microbiology, the host factors, and the clinical and counter-level variables, were all really helpful in being able to formulate the design as well as conduct the study using this dataset.
Dr. Enfield: It’s a pretty compelling reason for using the dataset but there are clearly some limitations, as you alluded to early in the conversation. I wondered if you might bring those out for the reader who might not be familiar with the Cerner dataset and how that might impact your findings.
Dr. Kadri-Rodriguez: Sure. I would say that the limitations to the Cerner dataset are not specific to that particular one and probably could be extended to most electronic health record repositories in general. In general, EHR systems are tweaked to fit the needs of specific hospitals and specific care settings. While that’s great for everyday operations, it does tend to introduce noise when you aggregate all these data because then our WBC count, for instance, can be labeled as a white blood cell count somewhere and a white corpuscular count somewhere and a WBC count somewhere. That’s just a simple example, but you can imagine across the range of variables that are of interest to us, if you have 200 different signals or names or labels coming from 200 different hospitals, those data are really not usable in a statistical model.
The first step is very thorough curation of the data. I think that Cerner and many other datasets like this, the main limitation is they’re not ready-to-go datasets. You cannot just take it and put the data in your model. We spend a lot more time cleaning the data than actually performing the analysis, just to give you some reference. I think that is one big limitation. The other limitation is that it’s an all-payer type of dataset. The data does not come from one particular payer or one particular HMO. I think one of the disadvantages there is that you can determine if the patient died in the hospital or was discharged to hospice, for instance, but beyond that, it’s really difficult to tell more long-term signals.
For instance, it’s difficult to assess chronic critical illness or long-term survival analyses in these type of datasets because, while you might be able to pick up encounters downstream among survivors of the particular index critical care encounter, what if they went to a hospital that’s out of the system and they were very sick and they died there? We would not have that information. So, there is the possibility of a type 2 error or an error where there’s a false negative where, if you don’t have information on them, does not mean that they survived or they did well.
Also, the mortality outcomes are not linked to datasets like the Social Security Death Index, which would’ve been great because it could have taken our outcomes beyond the hospital. Also, vitals and flow sheet data tend to have a lot of missingness because, after all, it’s not a trial. It’s not a prospectively collected dataset. It is real-world data. Think about the way in which your flow sheets get data in them; it’s far from systematic. That sort of nonsystematic generation of data where data is just available if it’s put in can create some problems when it comes to trying to use that information in your statistical models.
Dr. Enfield: One quick follow-up question to that. I think that some listeners will hear the phrase “cleaning the data” and worry about the introduction of bias. I know, from my experience that, when you say cleaning the data, you mean a very specific thing of what that entails. I wonder if you might expound a little bit more about what that means.
Dr. Kadri-Rodriguez: Yeah, that’s a good point. By cleaning the data, I mean homogenizing and curating the data so that it is in a format that makes it analyzable in a statistical model. It does not necessarily mean excluding large chunks of data or changing results. All it means is, if there is a certain variable that is heterogeneously expressed in the large repository, we try to reduce those complex dimensions into simple homogeneous dimensions that the computer can recognize as one. I think that, for the most part, that is a major part of the cleaning that’s done.
The other aspect of quote-unquote cleaning is identifying implausible signals. So, at the end of the day, these datasets are, as I said, real-world datasets. The data is generated at hospitals and then goes to a central repository and then is generated and then transmitted to us. Along the way, it is possible that there could be errors in transmission, errors in the way the data is relayed. I think that it’s really important to do some quality control checks and plausibility checks. That is nested within what we call cleaning.
Then there’s the aspect of missingness. We assess the degree of missingness and sometimes missingness is so large that you cannot really use the variable. When missingness is small and manageable, you still ask the question, is there a possibility that there is a differential bias in this missingness, like, is it missing for a reason? Even if you use that 50% data, you want to be confident that the missingness is not due to a specific reason or a specific tendency to treat or not to treat or to test or not to test, because that can really introduce a significant confounding. So, we use techniques like multiple imputations with chained equations, and we only use them when we are fairly convinced that the missingness is nondifferential. So, in summary, these are some of the elements in what we call cleaning.
Dr. Enfield: That’s great. I think that’ll be really helpful for our audience who may be less familiar with this kind of research. When I hear you talk, and I’m sure that there are some critical care fellows out there, the idea of doing research with large datasets could be really intriguing and I can see why they might be interested in doing that. What advice would you have for a fellow or a young faculty member who wants to pursue their own questions using these methods going forward?
Dr. Kadri-Rodriguez: I’d say go for it. But a couple words of caution. I say, respect the process of data cleaning and curation. It could make or break your study. Second, I would say, have a data analysis plan a priori. Even better, publish it upfront online. This creates a sense of transparency for yourself, within your group, and for the people reading and absorbing your work. And it leaves a time-stamped paper trail and minimizes data dredging and enables others, as I said, to trust your work more. In the era of data transparency and sharing, this is becoming almost an expectation. I suspect, over the next couple years, we will start to see more and more people doing observational research, providing a clear-cut a priori design, as well as statistical analysis brand upfront.
I think, when we end up analyzing and looking at the data, often we may not find what we expect in terms of how messy it might be or certain things you expected were there and they’re not there. So, you always have an opportunity to amend some of these a priori published protocols just like you would in a clinical trial. I think that just adds to the transparency while still allowing you to reasonably deviate from a plan you had thought of based on obvious needs.
Dr. Enfield: As we think about that transparency, Sameer, where is your research heading?
Dr. Kadri-Rodriguez: My team at NIH and I intend to continue to do these sort of observational studies. We’re constantly on the lookout for bigger, better data and questions of immediate and national and global importance. Hopefully, there’ll be more from us down the road. We are working currently on creating an antibiotic escalation algorithm. One of our fellows, Morgan Walker, in our program is leading that effort, along with Christina Yek. The idea is to create this antibiotic escalation algorithm for intensivists that is informed by current patterns of co-resistance that are observed among the pathogens seen in patients in U.S. hospitals who need ICU care.
Essentially, we’re trying to look at a very large dataset of bugs and the co-resistant patterns, so that if somebody is already on a certain type of antibiotic therapy and they’re decompensating and you want to escalate their therapy in a more informed fashion rather than slap on aminoglycosides to everybody, we feel like this is a good opportunity to leverage large databases to provide that informed decision that can not only inform these algorithms but potentially even inform sepsis guidelines down the road.
The other study that we have just completed, that hopefully we’ll submit to a journal soon, is looking at the clinical utility of the biomarker procalcitonin when it’s tested on admission. There is a lot of admission procalcitonin testing that’s happening these days in U.S. hospitals. I think it’s still a bit unclear to critical care docs in terms of what to do with the information, or can they hang their hat on it. This study is hopefully going to throw more light on that controversy. So, stay tuned for that one.
Dr. Enfield: That is all really exciting, and I know our testing stewardship team here will be looking forward to that paper as it is an often brought-up conversation. For me today, the takeaways from what I heard today are that BSI rates in the ICU are maybe even a little bit more common than we think, with 4% of our admitted patients and 1% of our patients developing BSI after they come into the ICU. Those that do get that BSI after they come in really do have a different antibiotic pattern than those who come in with their bloodstream infection ongoing. So, that is really going to be great information for us to take back to our patients today. The other thing that I really heard you say is the real power of doing this kind research, but also that it is a rigorous method that requires knowledge and skill set to really be able to trust what we get out after we do the projects. I really appreciate you sharing that with everyone today. Anything else that the listener should take home from our conversation?
Dr. Kadri-Rodriguez: I think you summarized that really well, Kyle. Thank you. I think the only thing I would add to this is, for the people who are interested in this kind of research, don’t be discouraged by the people that don’t believe in this work. Keep doing it. It is really a work in progress for us as a community. I think we just have to work together to strive to leverage better data, do even better at homogenizing the data, and doing even better at analyzing the data so that we can trust these more and utilize that information in our everyday work with more confidence.
Dr. Enfield: Sameer, I really appreciate that and the encouraging words for us to continue to do this kind of research. I want to thank you for everything you’ve done for the critical care community and this article, which I’m hoping everyone will download today and read. Again, you can find Sameer’s article at ccmjournal.org. For the Society of Critical Care Medicine Podcast, I’m your host, Dr. Kyle Enfield. This will conclude this episode. Thank you all for joining in.
This podcast is sponsored by Sound Physicians, the employer of choice for critical care physicians, where we seek to transform acute episodes of care. At Sound Physicians, we ensure physicians have the time and resources needed to deliver compassionate care that measurably improves quality and lowers the cost of healthcare for patients in the communities we serve. For more information, please visit careers.soundphysicians.com.
Kyle B. Enfield, MD, is an associate professor of medicine in the Division of Pulmonary and Critical Care at the University of Virginia. He received his undergraduate degree from the University of Oklahoma.
Join or renew your membership with SCCM, the only multiprofessional society dedicated exclusively to the advancement of critical care. Contact a customer service representative at +1 847 827-6888 or visit sccm.org/membership for more information.
The SCCM Podcast is the copyrighted material of the Society of Critical Care Medicine. All rights are reserved. Find more episodes at sccm.org/podcast.
This podcast is for educational purposes only. The material presented is intended to represent an approach, view, statement, or opinion of the presenter that may be helpful to others. The views and opinions expressed herein are those of the presenters and do not necessarily reflect the opinions or views of SCCM. SCCM does not recommend or endorse any specific test, physician, product, procedure, opinion, or other information that may be mentioned.
Some episodes of the SCCM Podcast include a transcript of the episode’s audio. Although the transcription is largely accurate, in some cases it is incomplete or inaccurate due to inaudible passages or transcription errors and should not be treated as an authoritative record.
