Clinical phenotyping assigns subgroups using routinely available bedside and electronic health record (EHR) variables including demographic data, comorbidities, vital signs, and standard laboratory tests. Seymour and colleagues derived 4 clinical sepsis phenotypes using early physiologic, laboratory, and organ dysfunction features. In a septic shock trial cohort, Aldewereld and colleagues identified 5 baseline clinical phenotypes using vital signs, routine laboratory tests, and organ dysfunction measures with 14-day all-cause mortality ranging from 7.9% to 36.3% across phenotypes and reported differences in clinical course.

Clinical phenotypes offer the theoretic advantage of a simpler pathway to clinical implementation, albeit their lack of tethering to underlying biology may make identifying targeted therapy more challenging. Prospective implementation of clinical subgroups has nevertheless been limited.

Biomarker-based subtyping has been used to define sepsis subgroups using early circulating markers of the host response that capture variation in systemic inflammation, endothelial injury and coagulopathy. In a representative analysis spanning 2 prospective cohorts, Sinha and colleagues identified a 2-class structure comprising Hyperinflammatory and Hypoinflammatory molecular phenotypes. The Hyperinflammatory phenotype was characterized by higher cytokine concentrations associated with inflammation, coagulopathy, and endothelial dysfunction, greater shock severity, and worse outcomes. Phenotypes were then assigned in PROWESS-SHOCK and VASST using validated classifier models applied to available biomarkers and clinical variables. Notably, treatment effect differed by phenotype for activated protein C in PROWESS-SHOCK, where treatment appeared to harm patients in the Hypoinflammatory phenotype and benefit patients in the Hyperinflammatory phenotype. In a subsequent study of approximately 8000 mechanically ventilated patients with sepsis, the same investigators observed that gram-negative bacteraemia was one of the biggest predictors of the Hyperinflammatory signature, which in turn was contingent on the pathogen’s burden, site of infection, and virulence. In the secondary analysis of a prior trial, they observed differential treatment response in the phenotypes with polymyxin B hemoperfusion.

Transcriptomic profiling supports distinct subgroups that may not be fully captured by routine clinical variables or circulating proteins. In the GAinS consortium, Davenport and colleagues identified 2 sepsis response signatures (SRS1/SRS2), with SRS1 characterised by relative immunosuppression and higher short-term mortality. In the Molecular Diagnosis and Risk Stratification of Sepsis (MARS) consortium, Scicluna and colleagues derived 4 sepsis genomic endotypes (Mars1–4) with separation in 28-day mortality, with Mars1 showing the most consistent excess risk across cohorts and characterized by an immune-suppressed signature. Complementing these cohort-derived signatures, cross-cohort classifier studies have converged on broad immunobiological subgroups with separation in organ dysfunction and mortality risk across independent datasets.

Transcriptomic subgroups also show evidence of treatment-effect heterogeneity. For example, Antcliffe and colleagues applied SRS assignment in septic shock trial participants and demonstrated that hydrocortisone therapy was associated with worse outcomes in SRS2 but not in SRS1. Given these findings were in a subset of patients, they remain hypothesis-generating; nevertheless, they provide a foundation for future transcriptomic guided trial enrichment and stratified treatment evaluation.

More recently, Scicluna and colleagues describe a unified blood transcriptomic framework for sepsis by integrating data from large, well-characterized cohorts (MARS and GAinS) and reconciling previously disparate classification systems. Through cross-cohort harmonization and network-based clustering, the authors identified 3 reproducible consensus transcriptomic subgroups (CTS1–3), each defined by distinct biological programs, including inflammatory/neutrophilic activation, heme and coagulopathic pathways, and interferon-driven adaptive immune responses. Notably, exploratory analyses suggested treatment heterogeneity in one subgroup (CTS2) with corticosteroids, once again in the VANISH trial.

Microbiology-informed subtyping recognizes that sepsis heterogeneity is partly determined by upstream factors related to the infection. Zhao and colleagues compared a host-only model with a host–pathogen model that added 3 microbiology variables: site of infection, pathogen class and drug resistance. Inclusion of microbiology variables altered phenotype assignment and was associated with differences in outcomes and treatment-response signals.

Candidate microbiology inputs for host–pathogen subtyping include infection site, organism class and measures of burden where available; however, culture-derived features are often delayed and show context-dependent associations with outcome, supporting their use as incremental evidence rather than stand-alone real-time stratifiers. ^,

ARDS subgroups are well described, with consistent identification of Hyperinflammatory and Hypoinflammatory subphenotypes in multiple studies with biological differences similar to those described above in sepsis. ^, These biological differences translate into clinically meaningful outcomes. Hyperinflammatory ARDS is associated with fewer ventilator-free days, fewer organ-failure–free days and higher mortality compared with the Hypoinflammatory subphenotype. ARDS subtyping has also been used to explore heterogeneity of treatment effect. In secondary analyses of randomized trials, simvastatin therapy is associated with improved survival in the Hyperinflammatory subphenotype but not in the Hypoinflammatory subphenotype. Differential treatment responses were also observed in these molecular phenotypes with positive end-expiratory pressure (PEEP) strategy in secondary analysis of a randomized trial and an observational study. While such analyses should be interpreted cautiously and require prospective confirmation, they have motivated phenotype-informed trial designs.

To make this phenotyping schema clinically actionable, parsimonious classifiers have been developed that combine small biomarker sets with routine physiology to approximate latent class membership. Machine learning–based variable selection identified 6 key classifier variables (IL-8, IL-6, protein C, soluble TNF receptor-1, bicarbonate and vasopressor use) from which 3-variable and 4-variable logistic models were derived. The best-performing models reproduced latent class assignment with high accuracy in an independent validation cohort and preserved separation in clinical outcomes and treatment-response signals.

AKI provides further proof-of-concept for biologically informed subtyping. In critically ill adults with AKI, latent class analysis identified two reproducible molecular subphenotypes distinguished primarily by endothelial dysfunction and inflammation, with the higher-risk subphenotype showing a greater risk of 7-day renal nonrecovery, defined as death, incident dialysis, or lack of resolution of AKI by day 7. Parsimonious classifiers enabled practical assignment of subphenotype membership. In VASST, these subphenotypes showed differential response to vasopressin, with lower 90-day mortality in the lower-risk subphenotype.

In pediatric septic shock, the PERSEVERE biomarker panel has been used to identify patients at risk of severe sepsis-associated AKI and to distinguish those more likely to recover renal function. Although this approach is better viewed as prognostic enrichment than subgroup discovery, it further supports the principle that early host-response profiling can refine renal risk stratification beyond conventional clinical variables alone.

Extending a parsimonious biomarker model derived from prior AKI subphenotyping work, Kiernan and colleagues applied it in a secondary analysis of CLOVERS and identified a sepsis subphenotype with greater endotheliopathy and inflammation that had higher 28-day mortality and fewer organ-support–free days with a liberal versus restrictive fluid strategy.

Arguably, among the multivariate subtyping schemes described above, the Hyperinflammatory and Hypoinflammatory molecular phenotypes are closest to being operationalized in the clinical space. There are 2 specific challenges that needed overcoming for the clinical implementation of these phenotypes. First, the phenotypes were originally identified using latent class analysis models that comprised 20 to 35 variables that were standardized to the population mean. ^, Second, these models also comprise 6 to 8 protein biomarkers with almost all lacking clinical-grade assays.

To address this problem, investigators used unbiased variable selection approaches to identify several 3- to 4-variable parsimonious classifier models (PCMs) to help identify the phenotypes at the bedside. The PCMs showed high performance metrics in validation sets comprising adult and pediatric ARDS ^,, and in adult critically ill patients with sepsis.

The reproducibility of these PCMs across multiple cohorts has motivated prospective evaluation of near-real-time implementation of molecular phenotypes. In the PHenotypes IN the Acute Respiratory Distress Syndrome trial (PHIND), plasma interleukin-6 (IL-6) and soluble tumor necrosis factor receptor-1 (sTNFR-1) were measured at enrollment using the Evidence MultiSTAT (Randox) near-patient immunoassay and combined with arterial bicarbonate in a validated parsimonious model to assign Hyperinflammatory and Hypoinflammatory ARDS subphenotypes in real-time. Prospectively assigned Hyperinflammatory patients had substantially higher 60-day mortality than Hypoinflammatory patients, supporting the clinical relevance of bedside classification. Similar feasibility work is underway in the Sub-Phenotypes of ARDS using Rapid Prospective Classification study, embedded within the United States-based I-SPY network, which is evaluating real-time prospective biologic subphenotype assignment in severe acute respiratory failure using rapid biomarker testing on the Evidence MultiSTAT (Randox) and Ella (Bio-Techne) platforms (NCT04488081). Building on this feasibility work, PANTHER is a Bayesian adaptive platform trial in acute hypoxaemic respiratory failure that incorporates real-time biologic classification at enrollment to enable phenotype-stratified evaluation of candidate therapies (ISRCTN81435672).

While these innovations have enabled advancing molecular phenotype-informed trials, the tests are relatively expensive and may not be ubiquitously available. Novel approaches to molecular diagnostics and protein quantification, such as plasmonic fluor lateral flow assays, electrochemical biosensors, and microfluidics are promising bioengineering innovations that may provide rapid and relatively inexpensive protein quantification ( Table 2 ).

A second challenge to the translation of these models is inconsistent availability of technology that quantifies protein biomarkers at the bedside. To circumvent the need for rapid quantification of research protein biomarkers, Sinha and colleagues used only readily available clinical data to classify molecular phenotypes using a machine learning algorithm called XGBoost. This model was trained to predict the Hyperinflammatory phenotype and showed good performance metrics in validation cohorts, including differential treatment responses in the 2 phenotypes for multiple therapies that included PEEP, fluid management, activated protein C, and polymyxin B.

As with molecular phenotypes, transcriptomic subtyping studies have sought to compress high-dimensional clustering solutions into parsimonious gene sets that enable more rapid clinical classification.

Hector Wong arguably pioneered this field through a series of studies in pediatric sepsis that delineated a translational pathway from transcriptomic endotyping discovery to clinical applicability. Initial work identified 2 reproducible pediatric septic shock endotypes (A and B) based on coordinated gene-expression programs, with endotype A characterized by repression of adaptive immunity and glucocorticoid signaling and associated with higher mortality. Subsequent studies demonstrated that these endotypes are biologically conserved and clinically relevant, extending across age groups and aligning with independently derived adult transcriptomic signatures such as the sepsis response signature (SRS). ^,

Sweeney and colleagues used large-scale transcriptomic integration to develop parsimonious classifiers of sepsis heterogeneity. In a multicohort meta-analysis across infectious and sterile inflammatory states, an 11-gene host-response signature was derived that distinguished infection from noninfectious inflammation across platforms and populations. They subsequently used unsupervised transcriptomic analysis to identify 3 reproducible sepsis subgroups (Inflammopathic, Adaptive, and Coagulopathic) that were conserved across independent cohorts and associated with distinct immune and coagulation pathways, clinical severity and outcomes. Together, these studies establish a coherent strategy in which cross-cohort harmonization can be used to identify sepsis signatures that are in turn identifiable via parsimonious gene-sets, bridging high-dimensional discovery with scalable clinical translation.

Likewise, the above-mentioned SRS, MARS, and CTS signatures of sepsis have all described parsimonious gene-sets that would theoretically allow rapid classification of the subgroups in critically ill patients. The residual challenge for the field for translating transcriptomics signatures to the bedside in sepsis involves using high-throughput molecular analysis that is rapid, reproducible, and relatively inexpensive that would allow prospective evaluation of therapies in transcriptomic subgroups in clinical trials. Antcliffe and colleagues have shown feasibility of prospective deployment of a parsimonious gene-set derived classifier across 17 intensive care units (ICUs) in Europe in 357 patients using an 11-gene panel measured on the FilmArray system at multiple timepoints after ICU admission, which they used to stratify patients into high-risk and low-risk groups.

Notably, there are several commercial point-of-care platforms that appear to have the technical capabilities using reverse transcription with either loop-mediated isothermal amplification or polymerase chain reaction to quantify gene-expression rapidly at the bedside for the purpose of sepsis recognition and management, but not real-time phenotyping. Specifically, the Inflammatix TriVerity system, an FDA-cleared platform uses a 29-gene mRNA panel and machine learning algorithm to distinguish bacterial from viral infections. Similarly, the SeptiCyte RAPID utilizes a focused 4-gene signature to provide a SeptiScore , helping clinicians differentiate sepsis from noninfectious systemic inflammation. Other emerging technologies, such as the Sepset lab-on-a-chip microfluidic device, aim to automate the process further, offering a potentially scalable way to identify high-risk endotypes directly in the emergency department or ICU. Together, these tools demonstrate the potential of moving sepsis management toward a model of real-time precision medicine. All these tools, however, require prospective evaluation of validity and clinical utility.

Routinely captured vital signs can support dynamic subtyping in early sepsis. Bhavani and colleagues derived 4 reproducible subphenotypes using vital signs recorded over the first 8 hours of hospital presentation and demonstrated meaningful separation in vasopressor use and 30-day mortality. In secondary analyses using SMART trial data, subphenotype membership modified the association between balanced crystalloids and saline with Group D patients (hypotensive without accompanying elevations in heart rate, temperature, and respiratory rate) demonstrating lower 30-day mortality with balanced crystalloids than saline. Based on these findings, there is now a prospective trial underway that is evaluating the efficacy of crystalloids versus saline among Group D patients in a randomized controlled trial.