Dynamic Phenotyping in ARDS: AI-Driven Insights Reveal Why Corticosteroid Benefits Depend on Inflammatory State

Introduction: Moving Beyond the Syndrome

For decades, Acute Respiratory Distress Syndrome (ARDS) has been treated as a monolithic clinical entity defined by the Berlin criteria. Despite numerous clinical trials, pharmacological interventions—including corticosteroids—have yielded inconsistent results, leading to a long-standing debate in critical care medicine. The emerging paradigm of precision medicine suggests that this inconsistency stems from biological heterogeneity. Recent research has identified two distinct subphenotypes: the hyperinflammatory phenotype, characterized by high levels of pro-inflammatory cytokines and worse outcomes, and the hypoinflammatory phenotype. However, a critical question remained: are these phenotypes stable over time, and can we identify them at the bedside using routine clinical data to guide therapy?

The Quest for a Clinical Surrogate

The study by Pensier et al., published in Intensive Care Medicine, addresses these challenges head-on. By developing an open-source AI Clinical Classifier (AIClarity), the researchers moved away from expensive and slow biomarker assays (like IL-6 or soluble TNF receptor-1) toward routine clinical parameters. This transition is vital for real-time decision-making in the intensive care unit (ICU).

Study Design and Methodology

The researchers employed a robust multi-stage approach to validate their AI tool and assess the temporal stability of ARDS phenotypes.

Development and Validation

The AI Clinical Classifier was developed using individual patient data and biomarkers from six multicenter randomized controlled trials. The development cohort included 1,207 patients, while the validation cohort comprised 2,751 patients. The goal was to ensure the algorithm could accurately distinguish between hyperinflammatory and hypoinflammatory states using only routine clinical data.

The Investigation Cohort

A retrospective investigation cohort of 5,578 patients was then used to observe how these phenotypes evolve over a 30-day period. To handle the complexity of phenotypic shifts, the team utilized a discrete-time Bayesian Markov model, assessing stability at 3-day intervals.

Target Trial Emulation

To determine the effect of corticosteroids, the researchers used target trial emulation and longitudinal logistic regression. This methodology helps mitigate the biases often found in observational data, allowing for a more accurate assessment of how steroid therapy interacts with a patient’s evolving inflammatory profile.

Key Findings: The Dynamic Nature of ARDS

The study’s results provide a transformative look at the pathophysiology of ARDS and the potential for personalized treatment.

Prevalence and Baseline Mortality

The AI Clinical Classifier identified 2,169 (39%) patients as hyperinflammatory and 3,409 (61%) as hypoinflammatory at baseline. The mortality gap was stark: 49% of hyperinflammatory patients died within 30 days, compared to only 24% of hypoinflammatory patients (p < 0.001).

Phenotypic Fluidity

One of the most significant findings was that ARDS phenotypes are not static. Over the 30-day observation period, 49% of patients who were hyperinflammatory at baseline transitioned to the hypoinflammatory phenotype. In contrast, only 7% of hypoinflammatory patients transitioned to the hyperinflammatory state. This suggests that while hyperinflammation is a high-risk state, it is often a transient phase of the illness.

The Corticosteroid Paradox

The response to corticosteroids was diametrically opposed depending on the phenotype:

1. The Hyperinflammatory Benefit

Patients in the hyperinflammatory state at baseline showed a significant reduction in mortality when treated with corticosteroids (IPW-weighted hazard ratio [HR]: 0.81; 95% CI 0.67-0.98, p = 0.033).

2. The Hypoinflammatory Risk

Conversely, patients in the hypoinflammatory state experienced higher mortality when given corticosteroids (IPW-weighted HR: 1.26; 95% CI 1.06-1.50, p = 0.009).

3. The Importance of Persistence

The benefit of steroids was not just about the baseline state. At day 3, a positive response to corticosteroids persisted only among patients who remained in the hyperinflammatory state (adjusted odds ratio = 0.51, 95% CI 0.32-0.80, p = 0.004). If a patient transitioned to a hypoinflammatory state, the benefit vanished.

Expert Commentary: Implications for the Bedside

The findings by Pensier et al. represent a major leap toward precision critical care. The ability to use routine clinical data via the AIClarity tool means that phenotyping is no longer a retrospective research exercise but a potential bedside reality.

Biological Plausibility

The divergent response to steroids makes biological sense. In a hyperinflammatory state, the potent anti-inflammatory effects of corticosteroids likely dampen the cytokine storm that drives lung injury and multi-organ failure. However, in a hypoinflammatory state, the immunosuppressive effects of steroids may interfere with necessary repair mechanisms or increase the risk of secondary infections, thereby worsening outcomes.

The Challenge of Timing

The study highlights that the “therapeutic window” for steroids in ARDS may be narrow and dependent on the patient’s current inflammatory trajectory. This challenges the current practice of starting a fixed course of steroids and suggests that clinicians should perhaps “re-phenotype” patients every few days to decide whether to continue or de-escalate therapy.

Study Limitations

While the study is robust, it is primarily based on retrospective data and trial emulation. The AI Clinical Classifier, while accurate, serves as a surrogate for gold-standard biomarkers. Prospective randomized controlled trials that use real-time phenotyping to assign treatment (predictive enrichment) are the necessary next step to confirm these findings.

Conclusion: A New Standard for ARDS Management?

The characterization of ARDS phenotypes using clinical surrogate data allows physicians to monitor the evolution of the disease with unprecedented clarity. The evidence suggests a clear mandate: corticosteroids should be targeted toward the hyperinflammatory phenotype and used with extreme caution—or avoided entirely—in hypoinflammatory patients. As AI tools like AIClarity become more integrated into electronic health records, the transition from “syndromic” management to “phenotypic” management will likely become the new standard of care, saving lives by matching the right treatment to the right patient at the right time.

References

1. Pensier J, Fosset M, Paschold BS, et al. Temporal stability of phenotypes of acute respiratory distress syndrome: clinical implications for early corticosteroid therapy and mortality. Intensive Care Med. 2025;51(10):1784-1796. doi:10.1007/s00134-025-08089-4.
2. Calfee CS, Delucchi K, Parsons PE, et al. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014;2(8):611-620.
3. Famous KR, Delucchi K, Ware LB, 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.