For the past several decades, critical illness has been defined by overlapping clinical syndromes such as sepsis, acute respiratory distress syndrome (ARDS), and multiple organ dysfunction syndrome (MODS). Although defining these syndromes has been crucial to improving outcomes and standardizing care in the ICU, there is growing concern that these definitions have run their course (1,2). Within each syndrome lies substantial biologic heterogeneity that has contributed to a multitude of negative clinical trials and frustrated generations of researchers and clinicians. A push to discover more homogenous groups of patients within these syndromes has led to the identification of subtypes that may offer a path forward toward more precise and effective interventions. As more work on subtyping emerges, it appears that many of these subtypes do not exist within one particular clinical syndrome but actually exist across the spectrum of critical illness (2). And even more apparent is that the discovery, identification, and application of these subtypes will rely on new developments in rapid biomarker detection as well as unsupervised and supervised machine learning (Fig. 1).

Figure 1.:

Previously, critical illness consisted of defining patients according to heterogenous clinical syndromes with significant, but ill-defined, overlap. Moving forward, critical care will rely on a combination of biomarkers, machine learning, and clinical data to identify more precise critical illness subtypes with important prognostic and therapeutic implications. Created with BioRender.com. ARDS = acute respiratory distress syndrome, MODS = multiple organ dysfunction syndrome.

In this issue of Pediatric Critical Care Medicine, Sanchez-Pinto et al (3) identified a trajectory-based, sepsis-associated MODS phenotype, termed the persistent hypoxemia, encephalopathy, and shock phenotype using an unsupervised machine learning approach in a large, multicenter cohort of critically ill children. The investigators started with pediatric sequential organ failure assessment (pSOFA) scores recorded for 38,732 children with suspected or confirmed infection over the first 72 hours of ICU admission. Sepsis-associated MODS occurred in 15,246 (39.4%) of these patients. Using derivation and validation cohorts and subgraph-augmented nonnegative matrix factorization, the authors identified four distinct trajectory-based, sepsis-associated MODS phenotypes. Although three of the groups tended toward early resolution of MODS, the group with the persistent hypoxemia, encephalopathy, and shock phenotype had persistent organ dysfunction characterized by greater mechanical ventilation and vasoactive needs as well as evidence of increased encephalopathy, coagulopathy, and systemic inflammation. Sanchez-Pinto et al demonstrated reproducibility of the persistent hypoxemia, encephalopathy, and shock phenotype by applying two additional unsupervised machine learning approaches to the cohort, which identified similar groups with moderate agreement by Fleiss’ kappa statistic.

Identifying subtypes in large databases is easy. Identifying clinically useful subtypes has proven to be a challenge. Sanchez-Pinto et al (3) presented a convincing argument that the persistent hypoxemia, encephalopathy, and shock phenotype could be both prognostically and therapeutically relevant to pediatric intensivists. Despite the persistent hypoxemia, encephalopathy, and shock group representing less than 15% of the total cohort, this group accounted for more than half of the recorded deaths with a four-fold increase in adjusted odds of mortality. Most importantly, patients with this phenotype who received at least 1 mg/kg of intravenous hydrocortisone or 0.5 g/kg of albumin in the first 24 hours of admission had a statistically significant reduction in both persistent MODS on day 7 and in-hospital mortality when compared to propensity-matched untreated controls. Although hydrocortisone and albumin may not have survival benefits in heterogeneous populations of critically ill children, this work raises the possibility that select patients may benefit from early initiation of these therapies.

The major limitation of this work is that the persistent hypoxemia, encephalopathy, and shock phenotype were identified retrospectively using 72 hours of pSOFA scores, whereas the survival benefit required hydrocortisone and albumin therapies to be given within the first 24 hours of admission. However, new developments in rapid biomarker detection and supervised machine learning approaches may offer ways to identify or predict this phenotype much earlier in the course of illness which would facilitate early administration of these therapies. Although research is needed to describe the biologic and clinical characteristics of the persistent hypoxemia, encephalopathy, and shock phenotype to determine its true clinical relevance, its strong association with mortality and the potential for targeted therapies make it an intriguing and vital area for further study.

Subtyping in critical illness is gaining traction. One example is the work by Calfee et al (4), identifying biomarker-based inflammatory subphenotypes of ARDS. Interestingly, several subtypes of sepsis, ARDS, and MODS have been identified that bear a striking resemblance to each other, including the persistent hypoxemia, encephalopathy, and shock phenotype. As described by Sanchez-Pinto et al, Knox et al used SOFA scores and an unsupervised machine learning approach to identify a similar sepsis-associated MODS phenotype in adults termed “shock with hypoxemia and altered mental status” (5). Using dozens of clinical variables and unsupervised machine learning, Seymour et al (6) also identified the γ phenotype in a group of adult sepsis patients who similarly had more vasoactive needs, worsened hypoxemia, systemic inflammation, and an increased risk for mortality. These phenotypes also share many characteristics with the ARDS hyperinflammatory subphenotype originally described by Dr. Calfee, which was recently also identified in pediatric ARDS as well as in acute respiratory failure without ARDS (7,8). Excessive inflammation appears to be a principal component of all of these subtypes and Carcillo et al (9) and Horvat et al (10) have described a hyperferritenemic phenotype in sepsis characterized by macrophage activation, increased mortality, and potential response to anti-cytokine therapy.

Whether these critical illness subtypes represent the same, overlapping, or entirely separate conditions remains unclear. However, each was identified using biomarkers and/or machine learning approaches representing the new path forward for critical care. Although biomarkers help peer into the hidden, underlying biology of critical illness, unsupervised machine learning allows us to see hidden patterns and clusters among the immense clinical data available to us. The future of diagnosing and managing critical illness will most likely rely on both. A key step will be incorporating these subtyping strategies into prospective clinical trials.

The true utility of the persistent hypoxemia, encephalopathy, and shock phenotype is yet to be seen, but there is mounting evidence that subtypes of critical illness exist that may respond differently to specific therapies. The traditional approach of treating patients solely according to clinical syndromes such as ARDS, sepsis, and MODS, whereas valuable, lacks precision and has proven inadequate. The development of more effective therapies to improve outcomes in critically ill children will hinge on our ability to describe critical illness subtypes using a combination of biomarkers and machine learning more precisely.

1. Maslove DM, Tang B, Shankar-Hari M, et al.: Redefining critical illness. Nat Med. 2022; 28:1141–1148 2. Reddy K, Calfee CS, McAuley DF: Acute respiratory distress syndrome subphenotypes beyond the syndrome: A step toward treatable traits? Am J Respir Crit Care Med. 2021; 203:1449–1451 3. Sanchez-Pinto LN, Bennett TD, Stroup EK, et al.: Derivation, Validation, and Clinical Relevance of a Pediatric Sepsis Phenotype With Persistent Hypoxemia, Encephalopathy, and Shock. Pediatr Crit Care Med. 2023; 24:795–806 4. Calfee CS, Delucchi K, Parsons PE, et al.; NHLBI ARDS Network: Subphenotypes in acute respiratory distress syndrome: Latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014; 2:611–620 5. Knox DB, Lanspa MJ, Kuttler KG, et al.: Phenotypic clusters within sepsis-associated multiple organ dysfunction syndrome. Intensive Care Med. 2015; 41:814–822 6. Seymour CW, Kennedy JN, Wang S, et al.: Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA. 2019; 321:2003–2017 7. Dahmer MK, Yang G, Zhang M, et al.; RESTORE and BALI study investigators: Identification of phenotypes in paediatric patients with acute respiratory distress syndrome: A latent class analysis. Lancet Respir Med. 2022; 10:289–297 8. Sinha P, Furfaro D, Cummings MJ, et al.: Latent class analysis reveals COVID-19-related acute respiratory distress syndrome subgroups with differential responses to corticosteroids. Am J Respir Crit Care Med. 2021; 204:1274–1285 9. Carcillo JA, Kernan KK, Horvat CM, et al.: Why and how is hyperferritinemic sepsis different from sepsis without hyperferritinemia? Pediatr Crit Care Med. 2020; 21:509–512 10. Horvat CM, Fabio A, Nagin DS, et al.; on behalf of the Eunice Kennedy Shriver National Institute of Child Health and Human Development Collaborative Pediatric Critical Care Research Network: Mortality risk in pediatric sepsis based on C-reactive protein and ferritin levels. Pediatr Crit Care Med. 2022; 23:968–979