Experts in the field of sepsis advocate for the incorporation of biomarkers and enrichment strategies into clinical trials (1). One form of enrichment involves stratification of study subjects according to baseline mortality risk. This is called prognostic enrichment. With this approach, the number of patients likely to have an outcome can be controlled to maximize the power of the study. For example, patients at low risk of an event might be excluded as being unlikely to benefit from a treatment, while patients at extremely high risk of an event might be excluded if their poor outcome is likely unalterable by the treatment under study. We have developed a stratification tool called Adult Septic Shock Information and Stratification Technology (ASSIST) (2, 3). ASSIST incorporates a panel of biomarkers, age, lactate, and chronic health status to estimate a baseline 28-day mortality probability for adults with septic shock. We envision that ASSIST can serve to enrich the sample enrolled into sepsis trials.
A recent phase IIb trial tested the efficacy of a polyclonal antitumor necrosis factor-α fragment antibody (AZD9773) in patients with severe sepsis or septic shock (4). The trial tested two doses of AZD9773 versus placebo. There was no difference in the primary study endpoint of ventilator-free days between the three study groups, nor was there a difference in 28-day mortality.
In order to test the concept that ASSIST can serve effectively as an enrichment strategy for sepsis trials, we replicated the original analysis of the AZD9773 study but incorporated enrichment by stratifying the cohort using ASSIST. The primary hypothesis is that the potential efficacy of AZD9773 is dependent on baseline mortality risk, as estimated by ASSIST.
Ethics, consent, and permissions
The original phase IIb trial was approved by the Institutional Review Boards of each participating institution, as previously published (4). The consent process allowed for secondary analyses of biological samples and clinical data.
Study protocol and biomarker measurements
Full details of the AZD9773 protocol were previously published (4). Briefly, 300 adults with severe sepsis or septic shock were randomized 1:1:1 to placebo, a low dose of AZD9773, or a high dose of AZD9773. We used plasma samples obtained at study entry to measure the ASSIST biomarkers. These include C-C chemokine ligand 3, C-C chemokine ligand 4, granzyme B (GZMB), heat shock protein 70 kDa 1B, interleukin-8, and interleukin 1α (2). The biomarkers were measured using a multiplex magnetic bead platform (MILLIPLEX MAP) designed for this project by the EMD Millipore Corporation (Billerica, Mass). Biomarker concentrations were measured in a Luminex 100/200 System (Luminex Corporation, Austin, Tex), according the manufacturers’ specifications. Technical assay performance data were previously published (5).
ASSIST was developed by combining biomarker concentrations, age, lactate concentration at study entry, and chronic disease status into a decision tree generated by Classification and Regression Tree methodology (2). The ASSIST biomarkers were originally identified by whole genome expression studies in children with septic shock, designed to identify genes having predictive capacity for mortality (5, 6). The biomarkers were subsequently tested in adult cohorts, leading to the development and testing of ASSIST (2, 7). The decision tree treats chronic health status as a dichotomous variable indicating the presence or absence of any one of the following chronic conditions: New York Heart Association Class IV congestive heart failure (NYHA Class IV CHF), chronic obstructive pulmonary disease, requirement for chronic dialysis, chronic hepatic failure, hematologic or metastatic solid organ malignancy, and requirement for chronic steroids at study entry. Study subjects are allocated to the terminal nodes of the decision tree based on specific rules reflecting the biomarker concentrations, age, lactate, and chronic disease status, with each terminal node assigning a mortality probability. The mortality probabilities of ASSIST range from 0.03 to 0.75. For the current study, study subjects were grouped into one of three mortality risk strata: low risk (mortality probability <10%), intermediate risk (mortality probability 10 to 50%), and high risk (mortality probability >50%).
Our statistical analysis plan was based upon the supposition that the effect of treatment would differ within each ASSIST-based risk stratum. Replicating the original primary analysis (4), we constructed a generalized linear model to estimate the least squares (LS) mean ventilator-free days for each study arm, adjusted for APACHE II and age. Secondarily, we constructed a logistic regression model to estimate the odds of 28-day mortality. To accommodate expected differences in the main effects of treatment within risk stratum, a nested design was used as a more powerful alternative to a stratified analysis. Analyses used SPSS v 23.0 (IBM Corp, Armonk, NY).
The demographic and clinical characteristics of the study cohort were previously published (4). Among the 300 subjects in the original study, there were 286 subjects with enrollment plasma samples available. Eighty-one subjects were classified low risk, 165 subjects were classified as intermediate risk, and the remaining 40 subjects were classified as high risk, based on ASSIST. The 28-day mortality rates for the three groups were 8%, 22%, and 40%, respectively. For estimating the risk of 28-day mortality, ASSIST had an area under the receiver operating characteristic curve of 0.65 (95% CI: 0.58–0.73; P <0.001). This reflects a sensitivity of 90% (95% CI: 79–96) and a specificity of 35% (95% CI: 29–42) for mortality. In comparison, APACHE II had an area under the receiver operating characteristic curve of 0.67 (95% CI: 0.59–0.74, P <0.001) for estimating the risk of 28-day mortality.
Table 1 shows the impact of ASSIST, and low-dose and high-dose AZD9773 on the LS mean ventilator-free days, using the high-dose/high-risk group as the baseline because this group had the lowest number of ventilator-free days. In the model, both APACHE II and the effect of study arm within risk strata were significant (P <0.001 and P = 0.017), whereas age was not significantly associated with outcome (P = 0.680). Table 2 shows the mean effect of low-dose AZD9773 and high-dose AZD9773 on the number of ventilator-free days when compared with placebo within each risk stratum. Within the low-risk group, there was an increase in ventilator-free days for both drug arms. Within the intermediate-risk group, there was an increase in ventilator-free days among those in the low-dose arm, but a decrease in the high-dose arm. Among high-risk patients, there was a decrease of ventilator-free days in both drug arms.
The logistic regression model for 28-day mortality was consistent with the model for ventilator-free days. APACHE II was statistically significant (P = 0.005), age was marginally significant (P = 0.058), and study arm within ASSIST risk strata was marginally significant (P = 0.063). Table 3 shows the odds ratios for 28-day mortality for the low-dose and high-dose arms, relative to placebo within the three ASSIST risk strata. Within the low-risk group, the odds for 28-day mortality were decreased for both drug arms. Within the intermediate risk group, the odds for 28-day mortality were decreased in the low-dose arm, but increased in the high-dose arm. Among high-risk patients, the odds for 28-day mortality were increased for both drug arms.
Taken together, these analyses suggest that the potential effects of AZD9773 on outcomes are dependent on ASSIST-based mortality risk, and the dose effect differs between the three strata.
Enrichment refers to the use of any patient characteristic to select a population in which a drug or treatment effect is more likely than in an unselected population (8). Prognostic enrichment refers to the selection of patients based on the likelihood of a disease-related event. Predictive enrichment refers to the selection of patients who are more likely to respond to an intervention based on an underlying biological mechanism. Prognostic and predictive enrichment are fundamental concepts for embracing precision medicine. In addition, enrichment strategies are particularly effective in trials involving highly heterogeneous syndromes, such as septic shock (9). We propose ASSIST as a candidate prognostic enrichment strategy for septic shock trials and test this concept in the current study.
Unlike the original analysis that adjusted for risk using APACHE II but did not consider enrichment, our models demonstrated statistically significant effects of AZD9773 within ASSIST-based risk strata. We observed an increase in the mean ventilator-free days in the low and intermediate-risk strata treated with a low dose of AZD, but a decrease in the high-risk strata. In addition, we observed a decrease in mean ventilator-free days in subjects treated with a high dose of AZD9773, in both the intermediate and high-risk strata. When considering a mortality outcome, the findings were analogous.
Collectively, the data suggest that a low dose of AZD0773 may be beneficial in adults with severe sepsis or septic shock having a low to intermediate baseline risk of mortality by ASSIST, but potentially harmful in those with a high baseline risk of mortality by ASSIST. In addition, the data suggest that a high dose of AZD9773 may be harmful in subjects with an intermediate to high baseline risk of mortality. This is consistent with a previous trial in patients with septic shock, which used an anti-TNF strategy based on a TNF receptor:Fc fusion protein (10). That study found an association between higher doses of the fusion protein and mortality, leading the investigators to speculate that excessive inhibition of TNF activity might have exacerbated systemic infection in some patients.
These observations support our primary hypothesis that the potential effects of AZD9773 are dependent on baseline mortality risk as assessed by ASSIST. Moreover, they extend the findings of the original study, which showed that in general there was a trend toward worse outcomes for high-dose AZD9773 and better outcomes for low-dose AZD9773 when compared with placebo (4). We posit that had participants for the study been selected based on having a low or intermediate ASSIST mortality, AZD9773 would be more likely to have demonstrated benefit.
In the original study, adjustments for APACHE II were used to consider risk (4). ASSIST includes biomarker information that reflects the underlying biology while APACHE II reflects symptoms and clinical risk factors (2, 3, 5, 11, 12). As such, enrichment using ASSIST may better stratify patients according to pathophysiology. We note that the ASSIST biomarkers reflect inflammation and immune function (3, 6). It is possible that an antitumor necrosis factor strategy could modify the biology reflected by these biomarkers, and hence modify outcome.
Our results must be interpreted with caution because they represent a post hoc analysis. We cannot reliably conclude that AZD9773 per se is beneficial in certain mortality risk subgroups of patients with septic shock. Rather, our data support the concept that ASSIST can potentially serve as an enrichment strategy for sepsis clinical trials. This concept requires prospective testing, which in turn requires the development of a rapid assay platform that can measure the ASSIST biomarkers in a timeframe allowing for prompt decision making in critically ill patients. We are currently working toward the development of such an assay platform.
1. Cohen J, Vincent JL, Adhikari NK, Machado FR, Angus DC, Calandra T, Jaton K, Giulieri S, Delaloye J, Opal S, et al. Sepsis
: a roadmap for future research. Lancet Infect Dis
2015; 15 5:581–614.
2. Wong HR, Lindsell CJ, Pettila V, Meyer NJ, Thair SA, Karlsson S, Russell JA, Fjell CD, Boyd JH, Ruokonen E, et al. A multibiomarker-based outcome risk stratification
model for adult septic shock. Crit Care Med
2014; 42 4:781–789.
3. Alder MN, Lindsell CJ, Wong HR. The pediatric sepsis
biomarker risk model: potential implications for sepsis
therapy and biology. Expert Rev Anti Infect Ther
4. Bernard GR, Francois B, Mira JP, Vincent JL, Dellinger RP, Russell JA, Larosa SP, Laterre PF, Levy MM, Dankner W, et al. Evaluating the efficacy and safety of two doses of the polyclonal anti-tumor necrosis factor-alpha fragment antibody AZD9773 in adult patients with severe sepsis
and/or septic shock: randomized, double-blind, placebo-controlled phase IIb study. Crit Care Med
2014; 42 3:504–511.
5. Wong HR, Salisbury S, Xiao Q, Cvijanovich NZ, Hall M, Allen GL, Thomas NJ, Freishtat RJ, Anas N, Meyer K, et al. The pediatric sepsis
biomarker risk model. Crit Care
2012; 16 5:R174.
6. Kaplan JM, Wong HR. Biomarker discovery and development in pediatric critical care medicine. Pediatr Crit Care Med
2011; 12 2:165–173.
7. Wong HR, Walley KR, Pettila V, Meyer NJ, Russell JA, Karlsson S, Shashaty MG, Lindsell CJ. Comparing the prognostic performance of ASSIST
to interleukin-6 and procalcitonin in patients with severe sepsis
or septic shock. Biomarkers
2015; 20 2:132–135.
8. Simon R. Clinical trial
designs for evaluating the medical utility of prognostic and predictive biomarkers
in oncology. Per Med
2010; 7 1:33–47.
9. Wong HR, Atkinson SJ, Cvijanovich NZ, Anas N, Allen GL, Thomas NJ, Bigham MT, Weiss SL, Fitzgerald JC, Checchia PA, et al. Combining prognostic and predictive enrichment
strategies to identify children with septic shock responsive to corticosteroids. Crit Care Med
2016; 44 10:e1000–e1003.
10. Fisher CJ Jr, Agosti JM, Opal SM, Lowry SF, Balk RA, Sadoff JC, Abraham E, Schein RM, Benjamin E. Treatment of septic shock with the tumor necrosis factor receptor:Fc fusion protein. The Soluble TNF
Study Group. N Engl J Med
1996; 334 26:1697–1702.
11. Wong HR, Weiss SL, Giuliano JS Jr, Wainwright MS, Cvijanovich NZ, Thomas NJ, Allen GL, Anas N, Bigham MT, Hall M, et al. Testing the prognostic accuracy of the updated pediatric sepsis
biomarker risk model. PLoS One
2014; 9 1:e86242.
12. Wong HR, Weiss SL, Giuliano JS Jr, Wainwright MS, Cvijanovich NZ, Thomas NJ, Allen GL, Anas N, Bigham MT, Hall M, et al. The temporal version of the pediatric sepsis
biomarker risk model. PLoS One
2014; 9 3:e92121.
Keywords:© 2016 by the Shock Society
Biomarkers; clinical trial; enrichment; sepsis; stratification; TNF; APACHE; Acute Physiology and Chronic Health Evaluation; ASSIST; Adult Septic Shock Information and Stratification Technology; LS; least squares