Reproduce “Deep Representation Learning-Based Dynamic Trajectory Phenotyping for Acute Respiratory Failure in Medical Intensive Care Units” on OpenResearch

About this paper

Alan Wu, Tilendra Choudhary, Pulakesh Upadhyaya, Ayman Ali, Philip Yang, Rishikesan Kamaleswaran

Deep representation learning applied to 24-hour pre-ventilation ICU trajectories identified four clinically interpretable sepsis-associated acute respiratory failure phenotypes: liver dysfunction/heterogeneous, hypercapnia, hypoxemia, and multiple organ dysfunction. The groups differed in 28-day survival, comorbidity patterns, and mortality, with the hypercapnia phenotype showing the lowest mortality (28.01%) and the MODS phenotype the highest (45.60%).

Key insights

• The resulting phenotypes extended beyond respiratory gas exchange: one group was dominated by liver dysfunction and another by broader multiorgan failure.
• The hypercapnia cluster was strongly associated with COPD, elevated PaCO₂, lower pH, and higher bicarbonate, whereas the hypoxemia cluster had high oxygen requirements and reduced oxygenation ratios.
• The MODS group had the greatest comorbidity burden and mortality despite relatively preserved oxygenation measures, indicating that respiratory failure severity alone does not explain risk.

Summary by alphaXiv.