Highlight
A large multi-cohort analysis identified four clinically reproducible 14-day oxygenation trajectories in persistent acute hypoxemic respiratory failure: early recovery, stable persistence, biphasic improvement-deterioration, and rapid decline.
These trajectory classes were externally validated across UK and Netherlands cohorts and carried markedly different short-term mortality risks, ranging from 0.3% to 100% at 14 days.
An early prediction model using 12 routinely available variables achieved externally validated discrimination by day 3, suggesting that clinically useful trajectory assignment may be possible before the full course is apparent.
The most severe trajectory was enriched for hyperinflammatory biology, while ARDS was most common in the stable persistence group, supporting the concept that longitudinal clinical evolution captures information not contained in static baseline labels alone.
Background and Clinical Context
Acute hypoxemic respiratory failure is one of the defining syndromes of critical care medicine. It encompasses a broad spectrum of causes, including pneumonia, sepsis, aspiration, trauma, and nonpulmonary inflammatory states, and it frequently overlaps with acute respiratory distress syndrome (ARDS). Yet patients who satisfy similar bedside oxygenation thresholds at a single time point often follow very different clinical courses. Some recover rapidly, others remain stably impaired for days, and some worsen despite initially reassuring trends.
This heterogeneity matters because the ICU still largely categorizes respiratory failure using static definitions, such as a single PaO2/FiO2 ratio, chest imaging, or baseline severity scores. These approaches are practical, but they can mix patients at very different stages of illness. A patient early in recovery and a patient about to deteriorate may share the same oxygenation value on a given day. Such stage-mixing likely weakens prognostic models, complicates trial enrollment, and obscures biological signals.
The study by Marshall and colleagues addresses this gap by asking whether clinically meaningful and reproducible trajectory classes can be derived from the longitudinal course of oxygenation in persistent acute hypoxemic respiratory failure. The central premise is important: disease evolution may be as informative as disease state. If robust trajectory archetypes exist, they could improve prognosis, sharpen enrichment for trials, and provide a more logical framework for linking physiology to biology.
Study Design and Methods
Design and Cohorts
This was a multi-cohort trajectory analysis using one derivation cohort and two external validation cohorts. The derivation dataset was MIMIC-IV, with 3,938 adults. External validation included cohorts from the UK and the Netherlands, bringing the total validation population to 6,480 patients.
Patients were adults with persistent acute hypoxemic respiratory failure defined by PaO2/FiO2 less than 300 mmHg and positive end-expiratory pressure of at least 5 cmH2O for at least 72 hours. This is a clinically relevant design choice because it excludes very transient hypoxemia and focuses on sustained respiratory dysfunction, where trajectory modeling is more likely to capture meaningful disease patterns.
Trajectory Modeling
The investigators modeled daily mean PaO2/FiO2 values through day 14 together with time to ICU discharge or death using a competing-risk latent class mixed model. This is methodologically important. Standard trajectory analyses can be biased when patients leave the ICU early or die, because missing data are not random. By jointly modeling oxygenation with competing outcomes, the study attempts to preserve the relationship between physiologic evolution and meaningful clinical endpoints.
Prediction Modeling
To test whether trajectory class could be recognized early, the authors developed a 12-variable XGBoost model for trajectory prediction. According to the abstract, discrimination in external validation reached mean area under the receiver operating characteristic curve values of at least 0.78 by day 3, with a range of 0.70 to 0.86. While the abstract does not list all 12 variables, the key practical point is that early prediction relied on routine clinical data rather than specialized assays.
Exploratory Biological and Syndrome-Level Analyses
The investigators also examined how trajectory classes related to the prevalence of ARDS and to hyperinflammatory subphenotypes. This is a valuable translational step. The field has increasingly recognized that syndromic definitions such as ARDS capture populations that are clinically useful but biologically diverse. Testing whether longitudinal clinical archetypes align with inflammatory subphenotypes may help bridge bedside physiology and underlying mechanism.
Key Findings
Four Distinct Trajectory Classes
A four-class solution provided the best model fit. The trajectory classes were clinically intuitive and prognostically distinct.
TC1, early recovery, represented patients whose oxygenation improved promptly. This group had the most favorable outcome, with 14-day mortality of 0.3%.
TC2, stable persistence, consisted of patients with sustained but relatively stable hypoxemia over the observation period. Fourteen-day mortality was 8%.
TC3, biphasic improvement-deterioration, included patients who initially improved but later worsened. This pattern is especially interesting clinically because it resembles a common ICU scenario in which early stabilization is followed by secondary injury, superinfection, fluid overload, ventilator-associated complications, or evolving multiorgan dysfunction. Fourteen-day mortality was 17%.
TC4, rapid decline, was the smallest but most ominous class, characterized by worsening oxygenation and 100% mortality by day 14. Although extreme groups sometimes raise concerns about overfitting, the external reproducibility and biological enrichment reported here support the possibility that this class reflects a genuinely recognizable terminal trajectory.
External Validation and Generalizability
The trajectory archetypes generalized to the external cohorts with high assignment certainty. This is one of the study’s major strengths. Many unsupervised phenotyping studies produce classes that are unstable across datasets because they depend heavily on local practice patterns, case-mix, or measurement timing. Reproducibility across geographically distinct cohorts suggests that these oxygenation trajectories capture clinically durable patterns rather than artifacts of one database.
Associated Biomarker and Clinical Trajectories
The abstract notes that the four classes demonstrated distinct patterns in other clinical biomarker trajectories. Although details are not included in the summary, this finding matters because it implies that oxygenation classes are not isolated respiratory patterns alone. Instead, they likely index broader systemic illness evolution, potentially involving inflammation, organ dysfunction, or treatment response. This strengthens the clinical plausibility of the model.
Relationship to ARDS and Hyperinflammatory Subphenotypes
ARDS was most common in TC2, the stable persistence class, where approximately 50% of patients met ARDS criteria. This is notable because it suggests that conventional ARDS may often represent prolonged but not necessarily rapidly fatal respiratory dysfunction. In contrast, TC4 was enriched for the hyperinflammatory subphenotype, with 41% to 53% prevalence. That observation is consistent with prior literature showing that hyperinflammatory ARDS phenotypes have worse outcomes and distinct responses to supportive strategies.
The result is conceptually important: static ARDS labeling and dynamic clinical trajectories are not interchangeable. The former identifies a syndrome at a point in time; the latter reflects its unfolding course. Both may be useful, but they answer different clinical and biological questions.
Early Prediction Performance
By day 3, early prediction models achieved mean AUCs of at least 0.78 in external validation, ranging from 0.70 to 0.86. For a heterogeneous ICU syndrome, this is encouraging performance. It suggests that trajectory assignment need not wait for the full 14-day course to become visible. If implemented carefully, early trajectory prediction could support risk stratification, trial enrichment, and perhaps more tailored monitoring or rescue decisions.
Clinical Interpretation
This work supports a shift from static phenotyping toward longitudinal phenotyping in critical care. The distinction is more than technical. Clinicians already think in trajectories at the bedside: Is the patient turning the corner, plateauing, or spiraling? This study formalizes that intuition using reproducible statistical methods and external validation.
The immediate clinical value is prognostic. A patient in early recovery and one in rapid decline should not be treated as equivalent simply because both currently meet the same oxygenation threshold. The study suggests that trajectory-aware classification adds prognostic information beyond baseline severity alone.
The broader value is strategic. Trials in acute respiratory failure often fail partly because they enroll biologically and temporally mixed populations. A therapy targeting inflammation, fibroproliferation, or ventilator-induced injury may be diluted if tested across patients who are recovering, stable, and rapidly deteriorating all at once. Trajectory classes could therefore help refine enrollment or define adaptive interventions at more meaningful time windows.
The association between TC4 and hyperinflammatory biology also raises the possibility that simple bedside trajectories can function as pragmatic gateways to deeper biological phenotyping. This may be especially useful in health systems where advanced biomarker assays are unavailable in real time.
Strengths and Limitations
Strengths
The study has several major strengths: a large derivation cohort, robust external validation across different health systems, modeling that accounts for discharge and death as competing outcomes, and practical early prediction using routine variables. The focus on persistent respiratory failure rather than fleeting hypoxemia also improves clinical coherence.
Limitations
Several caveats should be considered. First, trajectory classes are observational constructs and should not be assumed to represent discrete biological entities. Some patients may sit near class boundaries, and classes may partly reflect treatment decisions as well as disease biology.
Second, the primary trajectory variable was PaO2/FiO2, a clinically important but imperfect measure. It is influenced by ventilator settings, fluid balance, hemodynamics, and measurement timing. Although the requirement for PEEP at least 5 cmH2O helps standardize interpretation, residual variability remains.
Third, the abstract does not provide calibration metrics, decision-curve analyses, or implementation details for the XGBoost model. Discrimination alone does not guarantee clinical utility. Whether the model improves decisions compared with standard severity assessment remains to be shown prospectively.
Fourth, mortality differences across classes are striking, especially the 100% mortality in TC4. Such extreme separation is clinically compelling but also warrants careful assessment in the full paper to understand class prevalence, confidence around estimates, and the role of treatment limitations or withdrawal of life-sustaining therapy.
Finally, the study appears to focus on ICU patients with persistent oxygenation failure. Generalizability to less severe ward-based hypoxemia, noninvasive respiratory support populations, or settings with different ventilation practices may be limited.
Expert Commentary and Relation to Prior Literature
This study fits within a broader movement in critical care toward phenotype-based and treatable-trait approaches. Prior work in ARDS has shown that latent classes, particularly hyperinflammatory and hypoinflammatory subphenotypes, are reproducible and prognostically important. Studies by Calfee and colleagues demonstrated that inflammatory phenotypes differ in outcomes and possibly treatment response. Marshall and colleagues add a complementary axis: time-dependent respiratory course.
Importantly, trajectory phenotyping may be easier to operationalize than biomarker-driven endotyping because oxygenation data are routinely available. That said, trajectory classes should probably not be viewed as replacements for biological phenotypes. The most promising future direction may be integration: combining trajectory, physiology, imaging, and molecular signals into multimodal risk frameworks.
Guidelines for ARDS and acute respiratory failure continue to emphasize lung-protective ventilation, conservative fluid strategies where appropriate, prone positioning in selected severe cases, and careful management of the underlying cause. This study does not alter those standards directly. Instead, it offers a new framework for understanding why similarly treated patients diverge so markedly over time.
Implications for Practice and Research
In clinical practice, trajectory-aware thinking could improve communication, prognosis, and timing of reassessment. By day 3, clinicians may be able to identify whether a patient is on a recovering, persistent, biphasic, or rapidly declining path. That may help prioritize diagnostic reevaluation, escalation to adjunctive therapies, family discussions, or transfer decisions.
For research, the implications are substantial. Future interventional studies could stratify or enrich by predicted trajectory class, reducing heterogeneity and potentially increasing power to detect treatment effects. Mechanistic studies could also focus sampling around trajectory transitions, especially the biphasic class, which may represent a window where preventable secondary injury occurs.
A key next step will be prospective validation embedded in care pathways. It will be important to test whether early trajectory prediction adds actionable information beyond established severity scores, whether it remains stable under varying treatment strategies, and whether it can be integrated into electronic health record workflows without creating excessive complexity.
Conclusion
Marshall and colleagues provide compelling evidence that persistent acute hypoxemic respiratory failure can be organized into four reproducible oxygenation trajectory archetypes with distinct outcomes and biological associations. The work moves the field beyond static snapshots of respiratory failure toward a more clinically intuitive and scientifically credible model of illness evolution. If prospectively validated, these trajectories could support earlier prognostic enrichment, more precise trial design, and tighter linkage between bedside physiology and critical illness biology.
For intensivists and respiratory clinicians, the key message is simple: in acute respiratory failure, where the patient is going may be as important as where the patient is.
Funding and Registration
The abstract and citation provided do not report a ClinicalTrials.gov registration number. Specific funding details should be confirmed from the full-text article.
References
1. Marshall DC, Green AD, Komorowski M, Patel BV, Parbhoo S, Antcliffe DB. Reproducible clinical archetypes in acute respiratory failure: a multi-cohort trajectory analysis. Intensive Care Medicine. 2026 May 17. PMID: 42149243.
2. Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014;2(8):611-620. PMID: 24853585.
3. Famous KR, Delucchi K, Ware LB, Kangelaris KN, Liu KD, Thompson BT, et al. Acute respiratory distress syndrome subphenotypes respond differently to randomized fluid management strategy. Am J Respir Crit Care Med. 2017;195(3):331-338. PMID: 27513822.
4. Matthay MA, Arabi YM, Siegel ER, Ware LB, Bos LDJ, Sinha P, et al. Phenotypes and personalized medicine in the acute respiratory distress syndrome. Intensive Care Med. 2020;46(12):2136-2152. PMID: 32965465.
5. Ranieri VM, Rubenfeld GD, Thompson BT, Ferguson ND, Caldwell E, Fan E, et al. Acute respiratory distress syndrome: the Berlin Definition. JAMA. 2012;307(23):2526-2533. PMID: 22797452.