Sepsis-associated acute kidney injury: recent advances in enrichment strategies, sub-phenotyping and clinical trials

Identifying SA-AKI sub-phenotypes with a higher risk of poor outcome provides opportunities to target interventions toward the higher risk group and to exclude cohorts that may not benefit or potentially come to harm. Two recent studies by Ozrazgat-Baslanti et al. tracked the clinical trajectories of AKI, one for surgical patients and one for all hospitalized patients with AKI [13, 14]. For surgical patients, the ADQI criteria were used to differentiate between ‘No AKI,’ ‘Rapidly Reversed AKI’ ‘Persistent AKI with Renal Recovery,’ and ‘Persistent AKI without Renal Recovery’ [14]. Surgical patients with sepsis who exhibited ‘Persistent AKI without Renal Recovery’ had the highest hospital mortality (45%), RRT use (40%) and decline of glomerular filtration rate (GFR) in the year following surgery. Among hospitalized patients, those with ‘Persistent AKI without Renal Recovery’ also had the highest hospital mortality (28%), need for RRT (13%) and risk of death within one year of discharge (19%). For clinicians and researchers, the challenge is to identify these at-risk patients early and to investigate potential interventions that may modify any of the outcomes.

Among patients included in the FROG-ICU and ADRENOSS study, the authors showed that an elevated proenkephalin level > 80 pmol/L (found in roughly 6% of the cohorts) at the time of admission to the Intensive Care Unit (ICU) without meeting the serum creatinine (SCr) or urine output criteria of AKI, was associated with an increased risk of mortality [15, 16]. In a planned sub-study of the PROCESS trial (a randomized controlled trial exploring the role of early goal directed therapy), the authors measured urinary cell cycle arrest markers tissue inhibitor of metalloproteinases 2 (TIMP-2) and insulin-like growth factor binding protein 7 (IGFBP7) before and after the 6-hour resuscitation period [17, 18]. They demonstrated that patients who still had an elevated biomarker level ([TIMP-2]·[IGFBP7] > 0.3) after receiving fluid resuscitation were at higher risk for a composite endpoint of progression to severe AKI (Stage 2/3), receipt of dialysis or mortality. The incidence of this composite endpoint was similar in patients with elevated biomarkers post-resuscitation regardless of their pre-resuscitation biomarker status and also similar in those with and without AKI at enrollment based on SCr and urine output criteria.

Accurate sub-phenotyping based on clinical features will be essential to making full use of the large-scale electronic health record (EHR) data to understand and manage AKI. Confusion can result from existing efforts to define sepsis subclasses, where researchers used different approaches for classification (empirical, hypothesis-based or agnostic) and interchangeable terms (such as subgroup, sub-phenotype or endotype) that were not reconciled with terminology applied in previously published studies [19]. Recent studies in unsupervised machine learning (ML) have given promise to the prospect of classifying AKI patients into sub-phenotypes. These ML models characterize sub-phenotypes without adhering to any preconceived hypothesis or guidelines. Three published studies separately developed ML models to sub-phenotype adult ICU patients with AKI [20‐22], adult ICU patients with SA-AKI within 48 h of admission and hospitalized adults with AKI within 48 h of admission. Two studies identified biomarkers that were significantly different between sub-phenotypes, and all studies were able to differentiate sub-phenotypes related to decreased renal function and higher mortality [20, 21, 23]. Finally, AKI trajectories have been proposed to define different phenotypes. The ADQI group published consensus statements regarding the definitions of ‘SA-AKI’ and its timing (‘early SA-AKI,’ within 48 h of diagnosis of sepsis and ‘late SA-AKI,’ defined as AKI between 48 h to day 7 after sepsis diagnosis) and also adopted the previously proposed timelines for AKI (7 days or less), acute kidney disease (1–90 days) and chronic kidney disease (90 + days) [8].

A major challenge is the fact that current methods for data-driven phenotyping are heterogeneous, there are no reproducible approaches across differing methodologies and datasets to identify sub-phenotypes and endotypes, and resources to aggregate the existing strategies for clinical impact are lacking [24]. These models were developed in specific cohorts and settings and their generalizability remains uncertain across different health systems and critical care units. Confirmatory studies in diverse settings will be needed. Any proposed sub-phenotypes should be assessed by testing for (1) consistency and reproducibility in other datasets, (2) biological plausibility and (3) clinical utility (i.e., ability to identify patients at high risk for a specific outcome or to predict treatment response). Finally, the ideal phenotyping algorithm should impact clinical decision making in real time and provide increased value over current severity scores.

Because of the heterogeneous nature of critical illness, many clinical studies have not been able to identify treatment benefits. The goal of precision medicine is to match the best available treatment option with the specific patient populations. Developing a classification system for biomarker-driven AKI endotypes will improve the understanding of AKI and enable researchers to engage in more specific therapeutic trials, thus getting closer to the goal of providing precision medicine to patients with AKI [24, 25] (Fig. 2).

In the setting of SA-AKI, an appealing approach is using big data to identify cohorts with shared biology (i.e., comorbidities, laboratory results, clinical variables, biomarkers) [16, 18]. Several groups have used artificial intelligence and advanced ML methods to find a variety of signals in the wealth of available data.

In the FINN-AKI cohort, investigators applied latent class analysis (LCA) and differentiated between two endotypes of AKI in patients with sepsis [21]. Patients with endotype-2, defined by higher plasma levels of inflammatory and endothelial injury markers, had higher 90-day mortality compared to endotype-1 (41% vs 29%) and also a lower probability of short-term renal recovery.

In a retrospective analysis of two ICU cohorts, Bhatraju and colleagues used LCA to characterize two distinct SA-AKI endotypes [26]. All patients had plasma collected within 48 h of ICU admission; 29 different clinical and laboratory values and seven vascular host and inflammatory biomarkers were included in the analysis. Different levels of specific biomarkers (including angiopoietin 1 and 2) and significant differences in the incidence of certain single-nucleotide polymorphism (SNPs) within angiopoietin-2 were identified across the 2 endotypes. A two-group model best separated the data with approximately 60% of patients in endotype-1 and 40% in endotype-2 category. These findings were replicated in a validation cohort where endotype-2 was associated with a 2 or more greater risk of renal non-recovery and 28-day mortality compared to endotype-1, even after adjusting for severity of illness.

The authors then developed a parsimonious prediction model that included the ratio of angiopoietin-2/1 and soluble tumor necrosis factor receptor-1 (sTNFR-1). The model had fairly good c-statistic to predict AKI sub-phenotypes; patients with lower biomarkers of endothelial dysfunction and inflammation were characterized as endotype-1. When applying the model to the VASST database (vasopressin versus norepinephrine in patients with septic shock), the authors observed that these two AKI endotypes identified prior to randomization showed heterogeneity of treatment effect for the early addition of vasopressin to norepinephrine [26]. Specifically, patients classified as endotype-1 had a lower 90-day mortality with the early addition of vasopressin, while mortality was not significantly different in the endotype-2 group if randomized to vasopressin. Further research is necessary to test these endotypes in prospective trials in SA-AKI. Of note, these sub-phenotypes share characteristics with other sub-phenotypes reported in other conditions such as acute respiratory distress syndrome (ARDS) [27]. This observation may not be surprising given that (i) the mechanisms of sepsis-induced organ failure are likely (at least partially) shared between organs, and (ii) similar biomarkers were used to identify sub-phenotypes. In this line, sepsis is the leading cause of ARDS and sub-phenotypes were replicated between sepsis and ARDS [28].

Seymour and colleagues harnessed data from four distinct sepsis cohorts (over 40,000 patients) to develop and validate four distinct phenotypes of sepsis [29]. They demonstrated differences in the incidence of organ injury (AKI, liver failure), as well as differences in biomarkers (inflammatory, coagulopathy, etc.) between sub-phenotypes with cross-variation in sub-phenotypes within trials and the differential treatment effects. These studies suggest that heterogeneity of treatment effect exists across sub-phenotypes and that so-called 'negative' therapies may be reconsidered in enriched SA-AKI populations and among specific sub-phenotypes.

Most AKI sub-phenotypes were discovered or validated retrospectively using existing databases. In only a few prospective studies and clinical trials, biomarker-based sub-phenotyping was implemented for predictive enrichment. The EUPHRATES trial was an interventional RCT where a biomarker (endotoxin activity) was used to risk stratify patients for enrollment, regardless of AKI status. Patients with an endotoxin activity of 0.6 or higher were randomized to receive two hemoadsorption treatments using polymyxin B and usual care versus sham hemoperfusion and usual care [30]. There was no difference in 28-day mortality between both groups. However, a subgroup analysis identified a ‘sweet spot’ of endotoxin activity (between 0.6 and 0.9) where patients may derive a benefit from hemoadsorption [31]. The trial did not include an assessment of the response to therapy and the protocol had no customization to patients’ response. It could be argued that some patients may have benefitted from greater customization of therapy.

A new trial, Tigris (NCT03901807), is currently underway to test this hypothesis prospectively. Importantly, the statistical analysis plan describes combining Tigris and EUPHRATES data using Bayesian statistics [32]. This approach is similar in concept to adaptive clinical trials which drop certain arms or groups as the trial progresses or even add or eliminate interventions. These adaptations incur less penalty when planned prospectively and may be important tools for future studies in SA-AKI.

In the ATHOS-3 trial [33], patients with vasodilatory shock on high-dose vasopressors were randomized to receive synthetic Angiotensin II or placebo with continuation of other vasopressors. The trial demonstrated that about 70% of patients in the intervention arm met the primary endpoint of a mean arterial pressure (MAP) > 75 mmHg or increase by 10 mmHg or more from baseline within 3 h of drug initiation. While the phase III trial did not show a mortality difference, subsequent work in specific subgroups suggested a survival benefit, including those with septic AKI and patients in whom Angiotensin II was introduced at lower doses of vasopressors [34, 35].

Healthy volunteers have low levels of angiotensin I and angiotensin II and a low angiotensin I/II ratio. In contrast, some patients with vasodilatory shock have very high levels of angiotensin I with a significantly elevated angiotensin I/II ratio, suggesting reduced conversion of angiotensin I to angiotensin II, resulting in angiotensin II deficiency. Patients with a substantially elevated Ang I/Ang II ratio, who were randomized to receiving exogenous Angiotensin II, had a survival benefit compared to the group that received placebo [36]. In a post hoc analysis of the ATHOS-3 trial, the authors measured angiotensin I and renin levels over time (at study initiation and 3 h later) and demonstrated that the angiotensin I and renin concentrations did not decrease in patients who received placebo while the levels fell in those who were randomized to angiotensin II therapy [37]. Based on this finding, the authors hypothesized that the activity of the angiotensin converting enzyme (ACE) was reduced in patients with septic shock and/or endothelial dysfunction [38]. Giving angiotensin II to a subgroup of septic patients with hyperreninemia enrolled in the ATHOS-3 trial was associated with lower mortality. ACE is largely a pulmonary capillary endothelial enzyme, the activity of which decreases with increase in severity of lung injury. It is similarly proposed to decrease in those with other reasons associated with altered pulmonary blood flow, such as cardiac surgical patients or patients on extracorporeal membrane oxygenation (ECMO) [39]. ACE activity is difficult to measure. However, reduced ACE activity usually results in increased renin release and potential diversion of the RAAS pathway through ACE 2, leading to more Ang 1–7 than Ang II [40, 41]. Thus, renin can be considered a surrogate marker of ACE activity. However, a true point of care bedside renin assay to target therapy is lacking and further work to define the role of biomarkers of the RAAS for predictive enrichment to guide Angiotensin II administration is needed. Finally, these post hoc analyses highlight the overlap between sub-phenotypes, have a relative small sample size with a risk of type 1 error and should be considered exploratory (Table 2).