Temporal stability of phenotypes of acute respiratory distress syndrome: clinical implications for early corticosteroid therapy and mortality.

Summary

Using routine clinical variables, an open-source AI classifier reliably identified hyperinflammatory and hypoinflammatory ARDS phenotypes, which were dynamic over 30 days. Target trial emulation showed corticosteroids reduced 30-day mortality in hyperinflammatory ARDS (HR 0.81) but increased mortality in hypoinflammatory ARDS (HR 1.26), with benefit persisting only in patients who remained hyperinflammatory by day 3.

Key Findings

• Clinical-data AI classifier identified 39% hyperinflammatory and 61% hypoinflammatory ARDS with markedly different 30-day mortality (49% vs 24%).
• Phenotypes were dynamic: 49% of baseline hyperinflammatory cases transitioned to hypoinflammatory by day 30; 7% of hypoinflammatory transitioned to hyperinflammatory.
• Corticosteroids lowered mortality in hyperinflammatory ARDS (IPW-weighted HR 0.81) but increased mortality in hypoinflammatory ARDS (HR 1.26).
• Steroid benefit persisted at day 3 only in patients remaining hyperinflammatory (adjusted OR 0.51).

Clinical Implications

At the bedside, clinicians can use the AI classifier to monitor ARDS inflammatory phenotypes and tailor corticosteroid use—favoring steroids in hyperinflammatory ARDS while avoiding them in hypoinflammatory ARDS, reassessing phenotype within 72 hours.

Limitations

• Treatment effects inferred from observational emulation are subject to residual confounding.
• Phenotyping relied on clinical surrogates rather than contemporaneous biomarker panels in the external cohort.

Future Directions

Prospective, phenotype-stratified RCTs of corticosteroids in ARDS; integration of dynamic phenotyping into decision support with predefined reassessment windows.