Kalil, Andre C. MD, MPH, FCCM; Kellum, John A. MD, MCCM

A protocol for resuscitation known as “early goal-directed therapy” (EGDT) has been promulgated since 2001 when Rivers et al (¹) showed survival benefits in patients with severe sepsis and septic shock. Many observational and quasi-experimental studies have found similar results. Recently, three large randomized trials could not confirm a survival advantage using this approach. A just-published individual patient data meta-analysis performed as a collaboration among the authors of the three randomized trials confirmed that EGDT did not result in better outcomes compared with usual care (²).

In a recent issue of Critical Care Medicine, one of the authors (³) took a combined Bayesian and frequentist methodological approach to evaluate 12 randomized trials (six fully published and six in abstract form) and 31 observational studies that included a total of approximately 20,000 patients with severe sepsis or septic shock. We demonstrated that of all components of EGDT, the only one that explained improved survival was time-to-first antibiotics. This time-to-first antibiotics explained 96–99% of the entire survival difference between EGDT and controls.

During that study, we made a novel observation: the treatment effect of EGDT seemed to depend on the baseline severity of sepsis as measured by Acute Physiology and Chronic Health Evaluation II, by Sequential Organ Failure Assessment, and by the presence of shock. Any benefit of EGDT was significantly reduced when disease was severe by these criteria (³).

In a second study, also recently published in Critical Care Medicine, one of the authors (⁴) evaluated survival outcomes in relation to inflammatory biomarkers measured in patients enrolled in the Protocolized Care for Early Septic Shock (ProCESS) trial. We found that any survival benefit with EGDT was limited to patients with low disease severity (lowest biomarker quartile), whereas EGDT seemed to decrease survival in those with high disease severity (highest biomarker quartile). The most recent individual patient data meta-analysis performed by the Protocolized Resuscitation in Sepsis Meta-Analysis investigators (²) also seemed to suggest a EGDT dependence on baseline disease severity considering the following subanalyses: 1) eligibility criteria met by refractory hypotension and hyperlactatemia; 2) lactate level before randomization; and 3) vasopressor infusion. The dependencies on baseline severity were not statistically significant. However, this lack of power is to be expected owing to the small sample size and lack of power of these subanalyses.

Such a treatment “effect modification” based on varying disease severity is also called a “treatment interaction” or a “heterogeneity of treatment effect.” It means that these patients did not have worse survival only because they were sicker; it means that EGDT itself may have further increased their mortality compared with that of similarly sick controls. The fact that the survival effect with this treatment interaction crossed the risk ratio unity defines a “qualitative interaction,” an interaction that suggests an opposing effect by the same therapy in different clinical situations (i.e., potentially beneficial in low disease severity but harmful in high disease severity). Further, the addition of the variable study design to the regression model in the study by Kalil et al (³) did not change the EGDT interaction by disease severity.

WHAT ARE THE POSSIBLE EXPLANATIONS FOR THIS?

Similar to the U-shaped effect previously observed with other therapies such as tight glucose control in the ICU and low tidal volume ventilation for acute respiratory distress syndrome, it is possible that although EGDT could benefit patients with low disease severity, the fixed bundle protocol approach could harm patients with high disease severity who may require a more individualized and flexible resuscitation approach. Alternatively, EGDT could include beneficial and also harmful effects that become manifest depending on disease severity. For example, suppose increasing fluids is beneficial but targeting central venous oxygen saturation with blood and inotropes is harmful. These components may be applied disproportionally to high- and low-severity patients, and different patients may tolerate the negative effects differently.

Either of these two hypotheses could explain why patients with high disease severity fare worse with EGDT. The indiscriminate administration of a fixed, standard hemodynamic resuscitation to patients who are already in an advanced stage of sepsis may increase susceptibility to complications associated with volume/pressors/inotropes (e.g., fluid overload, cardiac arrhythmias, and transfusion-related acute lung injury). Patients with advanced sepsis may be less likely to benefit against a background of care that is already more robust (e.g., more fluid) due to their increased severity. In other words, they would be more likely to become fluid overloaded and would not be able to manage the fluid overload as well as the low disease severity patients. Consequently, these high-severity patients would be more prone to develop systemic and pulmonary edema, tissue hypoxia, and end-organ damage, all of which would compound to increase mortality. An association of fluid overload with higher morbidity and mortality has been observed in two randomized trials (⁵, ⁶), seven cohort studies (^7–13), and one matched case-control study (¹⁴). Of note, these studies encompassed both adult and pediatric patients from four different continents. Although these studies consistently showed harm with fluid overload, other cointerventions may also have contributed to worse outcomes.

What if a single component or cointervention could be driving this qualitative interaction? Although unmeasured variables or interventions are always a possible source of bias in observational studies, this is infrequent in randomized studies. The study by Kalil et al (³) dissected each bundle by type of intervention and by compliance achievement. Other than the survival benefits from faster and more appropriate antibiotic interventions, the differences in age, country, hospital location, era, systolic pressure, mean arterial pressure, lactate, bundle compliance, and hemodynamic goal achievements were not associated with survival differences between studies.

Neither the immune system nor the nervous system responds linearly to therapies: each is a complex system displaying rich behaviors often disproportionate to a particular perturbation (¹⁵). The exaggerated (in some cases) and suppressed (in other cases) responses to EGDT may also reflect nonlinear responses in cardiovascular, pulmonary, and renal physiologies severely perturbed by sepsis. Indeed, treatments anticipating smooth, linear physiologic responses to therapeutic perturbations rarely seem to work in severely septic patients (¹⁶).

Although chance alone could account for the effects observed, the fact that a statistically significant qualitative interaction was found in two large studies performed by independent research groups (³, ⁴) suggests that the association of sepsis severity and the effect of EGDT may be real. What might be done to move the discussion forward?

Using study-level data, Kalil et al (³) revealed the reasons for the survival discordance between different studies. A patient population trial enrichment can be achieved by the identification of a “sepsis dynamic phenotype”: a combination of real-time physiologic and molecular biomarkers to evaluate the specific moment of the sepsis continuum that would require a precise treatment intervention. This would augment the probability to obtain the most benefits with the least risks for any sepsis therapy, including EGDT (¹⁶). This sepsis dynamic phenotype would also help identify a “tipping point” in which the complex systems associated with the inflammatory response lose their physiologic variability and consequently their opportunity to recover from the sepsis insult. A Bayesian trial enrichment by the performance of an adaptive trial with a stratified randomization based on disease severity might bring the fastest and safest resolution to the EGDT conundrum (¹⁷). The Bayesian approach would further diminish the sample size of the trial, reduce unnecessary patient exposure to either intervention or placebo, speed up the time from trial initiation to completion, accelerate the regulatory approval process, and reduce research costs.

We believe that a solution for the sepsis resuscitation puzzle can be achieved. However, intellectual bias and strong personal beliefs need to be replaced by a more encompassing scientific approach: both population and individual epidemiologic approaches are necessary, both linear and nonlinear systems need to be taken into consideration when evaluating any new experimental therapy for sepsis, both physiologic and molecular biomarker approaches are essential to better define the progression of the sepsis continuum, and both frequentist and Bayesian statistics are relevant and complimentary for the design and analysis of clinical trials. Only when we let science speak will we be able to further improve the outcome of our patients with sepsis.

REFERENCES