Highlight
Continuous inflammatory subphenotype probabilities uncovered marked mortality heterogeneity among patients labeled hypoinflammatory by standard binary classification.
Within hypoinflammatory acute hypoxemic respiratory failure, 90-day mortality rose from 19% to 31% to 40% across probability tertiles, despite all patients remaining below the conventional 0.5 hyperinflammatory threshold.
Longitudinal increases in probability identified especially poor-prognosis patients within the hypoinflammatory group, whereas hyperinflammatory patients had persistently high mortality regardless of trajectory.
These findings suggest that trial enrichment based solely on binary phenotype assignment may exclude biologically intermediate, high-risk patients who could be relevant to targeted therapies.
Background
Acute hypoxemic respiratory failure, including acute respiratory distress syndrome (ARDS), remains a syndrome defined more by physiology than by mechanism. This has limited progress in precision therapeutics. Over the past decade, latent class and biomarker-based analyses have consistently identified two broad inflammatory subphenotypes: hyperinflammatory and hypoinflammatory. Hyperinflammatory patients generally have higher circulating inflammatory mediators, more shock and metabolic derangement, and worse outcomes. These observations have motivated a growing interest in subphenotype-guided clinical trials.
Yet binary phenotyping creates an implicit assumption: that patients on either side of a threshold are biologically and prognostically similar within each category. This assumption is clinically convenient but potentially misleading. A patient just below a probability cutoff may be more similar to a patient just above it than to one with a very low probability. If so, binary assignment could mask clinically important gradients of risk and dilute treatment effects in phenotype-enriched trials.
The study by Venkatesan and colleagues directly addresses this issue. The authors asked whether continuous probabilities from a parsimonious biomarker model offer clinically meaningful information beyond a dichotomous hyperinflammatory versus hypoinflammatory label in adults with acute hypoxemic respiratory failure (AHRF). The answer appears to be yes, and the implications are potentially important for both prognostic enrichment and trial design.
Study Design
Population and cohorts
The primary analysis used 575 critically ill adults with AHRF enrolled in the Pittsburgh Acute Lung Injury Registry. External validation included 1,134 additional patients drawn from the EDEN trial, COVID-19 cohorts, and the RoCI registry. This multi-cohort approach is a notable strength because it tests whether the observed risk structure is reproducible across distinct clinical settings and etiologies of respiratory failure.
Subphenotype model
Continuous subphenotype probabilities were calculated using a parsimonious biomarker model incorporating interleukin-6 (IL-6), soluble tumor necrosis factor receptor-1 (sTNFR-1), and bicarbonate. A probability threshold of 0.5 was used to assign binary phenotype status: hyperinflammatory at or above 0.5, hypoinflammatory below 0.5. The authors also evaluated a procalcitonin-based model and reported similar qualitative findings, supporting robustness across biomarker platforms.
Outcomes and analysis
The primary endpoint was 90-day mortality. The investigators examined mortality across tertiles of continuous probability within each binary phenotype group. They also used restricted cubic spline modeling to assess the shape of the probability-mortality relationship and evaluated longitudinal trajectories in a subset of 330 patients with serial sampling. This is clinically relevant because inflammatory state in critical illness is dynamic, not static.
Key Findings
Binary classification captured one truth but concealed another
Of the 575 patients in the primary cohort, 77 patients, or 13%, were classified as hyperinflammatory and 498 patients, or 87%, as hypoinflammatory. As expected from prior literature, hyperinflammatory patients had a high overall mortality of 55%. However, within this hyperinflammatory category, mortality did not materially differ across probability tertiles, with a non-significant P value of 0.72. In practical terms, once patients crossed the 0.5 threshold, prognosis was already poor and remained relatively homogeneous.
The more revealing result emerged on the hypoinflammatory side. Among patients classified as hypoinflammatory, 90-day mortality increased substantially across probability tertiles: 19%, 31%, and 40%, respectively, with P less than 0.001. These are clinically large absolute differences. A patient categorized as hypoinflammatory but occupying the highest tertile of continuous probability had a mortality risk more than double that of a patient in the lowest tertile. That degree of dispersion is difficult to reconcile with the notion of a homogeneous low-risk phenotype.
The risk curve was non-linear, with the steepest rise below the binary cutoff
Restricted cubic spline modeling demonstrated a strong non-linear association between continuous probability and mortality. Importantly, the steepest increases in risk occurred below the 0.5 threshold. This is arguably the study’s most important conceptual contribution. It implies that the region immediately below the cutoff is not biologically benign. Instead, it contains patients whose risk is rising sharply, yet who remain labeled hypoinflammatory by binary convention.
For clinicians and trialists, this matters because thresholds simplify at the cost of information loss. If the mortality curve were flat below 0.5 and then rose abruptly above it, binary classification would be defensible. But the opposite pattern was observed: risk gradients were most pronounced among those still considered hypoinflammatory. This suggests that binary phenotyping may be most misleading precisely in the group currently presumed to be lowest priority for inflammatory enrichment.
Clinical severity and immune activation tracked with higher probabilities
The probability gradient was not an abstract statistical finding. Clinical severity scores and biomarkers of immune activation increased progressively across hypoinflammatory tertiles, all with P values below 0.001. This strengthens biological plausibility. Higher continuous probabilities corresponded not only to worse outcomes but also to a more severe clinical and inflammatory profile.
Although the abstract does not enumerate each associated variable, the pattern supports the interpretation that continuous probability captures real pathobiologic transition rather than random measurement noise. In other words, patients in the upper hypoinflammatory range appear to be biologically closer to hyperinflammatory illness than binary labels suggest.
Trajectory analysis added prognostic depth
In 330 patients with longitudinal sampling, rising subphenotype probabilities within the hypoinflammatory group were associated with very poor outcomes, with mortality ranging from 50% to 100%, compared with 16% to 40% in those with stable or declining trajectories, all with P less than 0.001. This is a striking finding because it shows that direction of change matters. A patient who is hypoinflammatory at one time point may still be on a trajectory toward a much higher-risk inflammatory state.
By contrast, hyperinflammatory patients had poor outcomes regardless of trajectory. That observation is consistent with the earlier result showing prognostic homogeneity within the binary hyperinflammatory class. Once in that category, additional granularity may add less discriminative value, at least for mortality.
External validation supports generalizability
The authors validated these patterns in 1,134 patients from EDEN, COVID-19 cohorts, and the RoCI registry. The heterogeneity within the hypoinflammatory group and the non-linear probability-mortality relationship were preserved across cohorts. Similar findings with the procalcitonin-based model further support that the central message is not an artifact of a single biomarker panel.
This cross-cohort replication is especially important because respiratory failure populations are heterogeneous by cause, treatment environment, and era. The inclusion of COVID-19 cohorts suggests that the phenomenon may extend beyond classic pre-pandemic ARDS populations, although disease-specific calibration and treatment interactions still require dedicated study.
Clinical Interpretation
This study challenges the operational simplicity of binary subphenotype assignment. It does not invalidate the hyperinflammatory versus hypoinflammatory framework; rather, it refines it. Hyperinflammatory classification still identifies a small group with consistently high mortality. But the hypoinflammatory category should no longer be considered uniformly lower risk or biologically quiescent.
The most immediate implication is for prognostic enrichment in clinical trials. Trials that enroll only patients crossing a hyperinflammatory threshold may exclude a sizeable subset of patients with high mortality and biologic evidence of immune activation who reside just below the cutoff. If these patients are responsive to immunomodulatory or phenotype-targeted therapies, then strict binary enrollment could reduce both efficiency and external validity.
Continuous probabilities may also improve bedside risk assessment, although the study should not be interpreted as practice-changing for individual patient management yet. Before clinical deployment, several issues remain: assay standardization, real-time turnaround, calibration across populations, and demonstration that acting on these probabilities improves outcomes. Still, the work advances the field from static categorization toward a more graded and dynamic view of critical illness biology.
Relation to Prior Literature
These findings fit well with the broader ARDS subphenotype literature. Calfee and colleagues first demonstrated reproducible hyperinflammatory and hypoinflammatory classes associated with distinct outcomes and treatment responses in secondary analyses of randomized trials. Subsequent work showed that parsimonious biomarker models can approximate latent class assignment using a small number of variables, making clinical translation more feasible. The present study extends that body of work by showing that even parsimonious models contain richer information in their continuous outputs than in their binary labels alone.
Conceptually, this mirrors developments in other areas of medicine, where continuous risk models often outperform threshold-based categories for prognosis. The novelty here is the demonstration that the information loss from dichotomization is not symmetric: it is most consequential within the nominally hypoinflammatory group.
Strengths and Limitations
Strengths
Important strengths include the use of a biologically grounded biomarker model, an adequately sized primary cohort, external validation across multiple independent datasets, and longitudinal sampling in a substantial subset. The consistency of findings using both the IL-6/sTNFR-1/bicarbonate model and a procalcitonin-based model increases confidence that the observed pattern is robust.
Limitations
Several limitations should be kept in mind. First, this is an observational prognostic analysis; it does not establish that patients with high hypoinflammatory probabilities will benefit from any specific intervention. Second, biomarker availability and turnaround may limit near-term implementation in many ICUs. Third, cohort heterogeneity is both a strength and a challenge: although validation supports generalizability, calibration of a probability model can shift across case-mix and assay platforms. Fourth, the abstract does not provide adjusted effect estimates or detailed discrimination metrics, so the incremental prognostic value over established severity scores cannot be fully assessed from the available data alone. Finally, trajectory analyses, while compelling, are based on a subset and need prospective confirmation.
Practice and Research Implications
For now, the most tangible application is in trial design. Future phenotype-guided studies in AHRF and ARDS should consider continuous probability-based enrichment rather than simple binary inclusion rules. One strategy would be to enroll patients above a lower probability threshold, or to stratify randomization by probability bands rather than phenotype status alone. Another would be adaptive designs that incorporate trajectory over time, recognizing that inflammatory risk may evolve rapidly during critical illness.
For translational science, these findings reinforce the idea that inflammatory phenotypes may represent a continuum rather than two entirely separate states. That framing may improve mechanistic studies linking host-response biology with treatment responsiveness. For bedside clinicians, the message is more cautious but still relevant: a “hypoinflammatory” label should not be equated with low risk, particularly when biomarker-derived probability is near the hyperinflammatory threshold or rising over time.
Funding and ClinicalTrials.gov
The abstract as provided does not report funding details or a ClinicalTrials.gov registration number. The primary analysis was conducted in the Pittsburgh Acute Lung Injury Registry, with validation in EDEN, COVID-19 cohorts, and the RoCI registry. Readers should consult the full article for detailed funding disclosures and cohort-specific regulatory information.
Conclusion
Venkatesan and colleagues provide a persuasive demonstration that binary inflammatory subphenotyping in acute hypoxemic respiratory failure obscures substantial risk heterogeneity, especially within the large hypoinflammatory group. Continuous biomarker-derived probabilities identified clinically meaningful gradients in mortality, illness severity, immune activation, and temporal risk trajectory that were largely invisible to dichotomous classification. Hyperinflammatory patients remained uniformly high risk, but many patients below the conventional 0.5 threshold were far from homogeneous.
The practical implication is clear: precision critical care may need to move beyond yes-or-no phenotype assignment toward continuous and dynamic biologic risk estimation. Whether that shift improves therapeutic targeting remains to be tested prospectively, but this study makes a strong case that the field is currently leaving useful prognostic information on the table.
References
1. Venkatesan N, Shah FA, Bain W, Yang Z, Dela Cruz CS, Baron RM, Zuchelkowski B, Rizzo AN, Joshi S, Arciniegas A, Zambrano K, Aneis H, Mayr FB, Morris A, Talisa VB, McVerry BJ, Nouraie SM, Kitsios GD. Risk heterogeneity within hypoinflammatory acute respiratory failure: continuous probabilities identify high-risk patients masked by binary classification. Intensive Care Medicine. 2026-04-27. PMID: 42043553.
2. Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA; NHLBI ARDS Network. 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: 24815836.
3. Famous KR, Delucchi K, Ware LB, Kangelaris KN, Liu KD, Thompson BT, Calfee CS; ARDS Network. Acute respiratory distress syndrome subphenotypes respond differently to randomized fluid management strategy. Am J Respir Crit Care Med. 2017;195(3):331-338. PMID: 27403916.
4. Sinha P, Delucchi KL, Thompson BT, McAuley DF, Matthay MA, Calfee CS. Latent class analysis of ARDS subphenotypes: a secondary analysis of the STATINS and HARP-2 randomised controlled trials. Lancet Respir Med. 2018;6(8):611-620. PMID: 30017544.
5. Sinha P, Delucchi KL, McAuley DF, O’Kane CM, Matthay MA, Calfee CS. Development and validation of parsimonious algorithms to classify acute respiratory distress syndrome phenotypes: a secondary analysis of randomised controlled trials. Lancet Respir Med. 2020;8(3):247-257. PMID: 31526756.