Echocardiography in the Management of Sepsis: Not All Black and White*

In 2018, a study by Feng et al (1) examined the association between the use of transthoracic echocardiography and mortality in patients with sepsis. Using the popular Medical Information Mart for Intensive Care (MIMIC) database, a large repository of electronic medical records from Beth Israel Deaconess Medical Center in Boston, the authors found that having an echocardiogram during an ICU stay for sepsis was associated with a decreased risk of death. One year later, a similar study by Lan et al (2)—also using the MIMIC dataset—found a similar association, albeit using a different statistical approach. In this issue of Critical Care Medicine, the question of whether echocardiography improves outcomes in sepsis is revisited in the study by Blank and Ruth (3). Again the MIMIC database is used (albeit a newer version), with yet another statistical approach being deployed. Why relitigate this issue in the face of two prior studies that had similar results? First, it remains a compelling question. Depending on how it is defined, the prevalence of sepsis-induced cardiomyopathy may be as high as 70%, and its timely diagnosis may have important consequences for treatment (4). Echocardiography might also provide useful insights into volume status and hemodynamic states, and can also be helpful in ruling out alternate causes of shock. Second, to some readers the idea that a one-time diagnostic test could lead to a reduction in mortality lacks face validity (5). This speaks to broader philosophical questions about whether a hemodynamic monitoring tool like echocardiography should be studied as an intervention with the potential to improve patient-centered outcomes. One widely examined case in point is the Pulmonary Artery Catheters in Management of patients in intensive care (PAC-Man) study (6), a large randomized controlled trial that found no difference in hospital mortality between critically ill patients who did or did not receive a pulmonary artery catheter for hemodynamic guidance. There are others as well—devices that use pulse pressure variation, bioreactance, and other technologies—none of which has high-quality evidence to support widespread adoption. Echocardiography may be no exception, with few prospective studies to evaluate its utility (7). Despite this lack of evidence, a closely related modality has been widely deployed in many ICUs. Point of care ultrasound (POCUS), which resembles (but does not replace) formal echocardiography, continues to garner attention, and the results of observational studies are unlikely to change this. Even randomized controlled trials could be met with a shrug; positive results will support its use, whereas negative results may be viewed with skepticism around whether a technique like POCUS, which strictly speaking informs treatment but is not a treatment itself, should be expect to improve complex outcomes like mortality. What did the authors of this latest study find? Their analysis showed that in patients admitted to the ICU with septic shock, there was no association between receiving an echocardiogram and risk of death. This result differs substantially from the prior studies, but may not be entirely contradictory. The first studies were done using the MIMIC III database, which includes patients admitted between 2002 and 2011, whereas the latest result is based on data from MIMIC IV, a more contemporary dataset spanning from 2008 to 2019. It is possible, and indeed likely, that care has evolved during this time to include more POCUS assessments; these may have mitigated the value of a formal echocardiogram, leading to the lack of association reported. The latter dataset also included the approximate year of admission, allowing the authors to control for this potentially important covariable. But another important distinction is the variation in statistical approaches used in these three papers. Ultimately, the aim of any such study is to derive an estimate of the causal effect of echocardiography on mortality. Randomizing patients to receive (or not receive) an echocardiogram would be the most effective way of doing this, because this would, in theory, balance both measured and unmeasured confounders, thereby enabling a more confident attribution of effect. In the absence of randomization, the authors had to deploy methods for causal inference using the observational data on hand (8). The notable difference in the effect estimate seen in this latest study might, therefore, reflect the complexity of causal inference methodology, and the impact of the myriad choices made in developing these models. There is no shortage of nuance and subjectivity here: choosing the parameters to identify the cohort of interest, selecting an appropriate outcome measure, identifying confounders and distinguishing these from mediators, managing bias from missing data (which is seldom missing at random), and dealing with immortal time bias and competing risk. All three studies used 28-day mortality as their primary outcome measure, but in all other aspects, there were important differences in how the models were developed. Furthermore, even optimal control of confounding cannot mitigate risk of bias from unmeasured confounders. Process measures may be useful in this regard. It would be helpful, for example, to know how quickly the echocardiograms were reported, whether the ICU team read these reports, and what actions were taken in response, but much of the process remains opaque in this dataset. Ultimately the measures of effect are estimates, and their interpretation is by no means straightforward. Evidence can and should evolve over time, as both disease patterns and practice patterns change. Perhaps it is this effect that we see reported here; it is possible that echocardiograms provided substantial benefit to patients in the ICU with sepsis in years past, but its impact is abrogated in contemporary practice. But this latest study offers another lesson entirely, one that is by no means new, but bears repeating nonetheless. Causal inference from observational data is not an exact science. Its output must be interpreted just as one might set the gain when doing an ultrasound of the heart—to see and appreciate all the important shades of gray.