Subphenotyping prone position responders with machine learning.

Summary

Unsupervised machine learning on 353 mechanically ventilated ARDS patients identified three subphenotypes with different mortality, but could not predict which patients would benefit from prone positioning. The study highlights ARDS heterogeneity and the limitations of current clinical variables for precision stratification.

Key Findings

• Three unsupervised machine learning-derived ARDS subphenotypes were identified among 353 proned patients.
• Subphenotypes exhibited different 28-day mortality rates, indicating prognostic heterogeneity.
• Available supine respiratory mechanics and oxygenation variables could not predict benefit from prone positioning.

Clinical Implications

Clinicians should continue evidence-based criteria for proning while recognizing heterogeneity; current routine variables may be insufficient to select responders, supporting collection of richer physiologic and multimodal data.

Limitations

• Retrospective design with potential selection and information biases.
• Lack of multimodal data (e.g., imaging, biomarkers) may limit predictive performance.

Future Directions

Integrate multimodal data (omics, imaging, ventilator waveform analytics) and prospective validation to enable actionable phenotypes for personalized prone positioning.