SEPSIS SUB-CLASSES: BE CAREFUL OF WHAT YOU WISH FOR

Similar to many conditions encountered in the intensive care unit, sepsis is a heterogeneous syndrome, rather than a uniform disease. It follows that there exist categories of sepsis defined by clinical features, physiology, biology, and/or genetics. Such categories are denoted by various labels, including sub-phenotypes, sub-classes, sub-groups, and endotypes (1). For convenience, we use the term sub-class from this point forward.

The clinical utility of sub-class identification is to inform therapy. Indeed, we already care for patients with sepsis in this manner. The sub-class of patients with sepsis due to gram-positive bacteria receives an antibiotic regimen different to that of the sub-class with gram-negative bacteria. Similarly, those with septic shock primarily characterized by low vascular resistance receive cardiovascular support that is different to those with septic shock primarily due to myocardial dysfunction. These are fundamental and effective sub-classification strategies, but the field can move toward more granular and sophisticated strategies commensurate with the complex biology of sepsis (2).

When evaluating sub-classification strategies, it is important to consider the approach to sub-classification (3). Beyond the initial derivation of the sub-classification strategy, it is imperative to validate the strategy in an independent cohort, wherein the strategy is applied a priori, without modifications, and then one determines if the sub-classification yields similar observations to that seen in the initial derivation phase. Additionally, the derivation and validation phases of the strategy require large numbers of study subjects, preferably from multiple centers, in order to capture patient heterogeneity.

One approach to deriving a sub-classification strategy is knowledge-based. In this approach, the investigator creates sub-classes and membership criteria based on existing knowledge and paradigms, and subsequently tests whether the sub-classes associate with outcomes or treatment responses. The advantages of this approach are that it is based on the traditional scientific method and therefore tends to be hypothesis-driven, and can often take advantage of readily available clinical and biological data. General familiarity with this approach also leads to acceptance by the medical community. The disadvantages of a knowledge-based approach are that it is limited by existing knowledge, it can be based on flawed paradigms, and it is highly susceptible to investigator bias.

In this issue of Pediatric Critical Care Medicine, Dr. Carcillo and colleagues report on inflammation-based sub-classes of sepsis, and their association with multiple organ failure outcome (4). The investigators used a knowledge-based approach. Accordingly, we suggest that the readership consider the data presented in the context described above. That is, consider the important issues of derivation, validation, and study subject number, and the nuances of a knowledge-based approach. We also suggest that the readership consider the vast complexity of inflammation and immune function, and whether this complexity can be reliably captured by the laboratory assays employed in this study. Appropriately, Dr. Carcillo and colleagues stress that their study should not be interpreted as explicitly advocating for the use of Rituximab, Eculizumab, Anakinra, or plasma exchange for pediatric sepsis. We strongly agree with this qualification of the data. With regard to plasma exchange, it is disappointing that a trial was conducted to evaluate this therapy among children with thrombocytopenia associated multiple organ failure, including those with sepsis, but the results have not been published; nor are there any results posted at the ClinicalTrials.gov website despite a trial completion date of February 2012 (NCT00118664).

An alternative approach to sub-class identification is discovery-based. In this data-driven approach, the investigator makes no a priori assumptions regarding sub-class characteristics or membership. The approach typically leverages high throughput technologies such as transcriptomics, proteomics, and metabolomics, or other forms of high dimensional data, and relies on complex bioinformatics and machine-learning tools to identify sub-classes in an unsupervised manner. The disadvantages of a discovery-based approach is that it can be costly, and susceptible to false positive findings because the number of variables considered are typically far greater than the number of study subjects. Additionally, the general medical community is relatively unfamiliar with this type of approach, thus leading to skepticism and a low rate of adoption. However, the costs of these technologies are decreasing in a virtually exponential manner, and the evolution of public repositories for such data provide opportunities to efficiently consider a much larger number of study subjects than is possible by any individual study (5, 6). Further, an increasing number of clinician-investigators are developing expertise with high dimensional data, bioinformatics, and machine learning.

The advantages of a genuine discovery-based approach are that it is unbiased and provides the opportunity to reveal biological mechanisms and associations that would be otherwise unconsidered by approaches limited to existing knowledge and paradigms. It has been proposed that high throughput, discovery-oriented approaches are ideally suited for biologically complex syndromes such as sepsis (7, 8).

Robust examples of discovery-oriented approaches for identifying sub-classes of sepsis and other forms of critical illness are steadily increasing. In all examples provided below, the sub-classes have important differences with regard to outcome. Another critical point is that the sub-classification strategies were initially derived agnostic to outcome. The Knight laboratory in the United Kingdom recently identified and validated two sub-classes of adult patients with sepsis based on large-scale transcriptomic signatures (9). Knox et al. identified four distinct clusters of multiple organ failure among 2,533 patients with sepsis, using self-organizing maps (10). Calfee et al. identified and validated two subclasses of acute respiratory distress syndrome using latent class analysis of a large number of clinical and laboratory variables (11, 12). The sub-classes identified by Calfee et al. appear to have a differential response to PEEP and fluid management strategies. Our own research group identified two subclasses of children with septic shock based on a transcriptomic signature reflecting adaptive immunity and glucocorticoid receptor signaling (13–15). These two groups appear to have vastly different responses to corticosteroids.

We expect this line of investigation, sub-class discovery in sepsis and other forms of critical illness, to expand in the near future. Such studies hold the promise of enabling precision critical care medicine, but much work remains to be done and the validity of any sub-classification strategy needs to be carefully considered. Whether a discovery-based approach is superior to a knowledge-based approach, or vice versa, is irrelevant. More important is for the field to understand the strengths and weaknesses of each approach, and the nuances of sub-class discovery, validation, and clinical relevance.